Traitement en cours

Veuillez attendre...

Paramétrages

Paramétrages

Aller à Demande

1. WO1999001985 - PROCEDE ET APPAREIL DESTINES A L'INSPECTION DE PLAQUETTES A SEMICONDUCTEURS ET DE DISPOSITIFS AFFICHEURS A CRISTAUX LIQUIDES ET FAISANT APPEL A LA DECOMPOSITION ET A LA SYNTHESE D'IMAGES MULTIDIMENSIONNELLES

Note: Texte fondé sur des processus automatiques de reconnaissance optique de caractères. Seule la version PDF a une valeur juridique

[ EN ]

METHOD AND APPARATUS FOR SEMICONDUCTOR WAFER AND LCD INSPECTION USING MULTIDIMENSIONAL IMAGE

DECOMPOSITION AND SYNTHESIS
The invention relates to a method and apparatus for image decomposition and synthesis and more particularly to a method and apparatus for the inspection of semiconductors and liquid crystal displays using multidimensional image decomposition and synthesis.
BACKGROUND OF THE INVENTION
During the manufacture of a successful semiconductor substrate or liquid crystal display, LCD, the substrate receives a number of inspections and material characterizations. These inspections and characterizations attempt to detect a variety of defects including process defects, scratches and contamination defects, etc. These inspections and material characterizations also attempt to detect significant process variations. Defects and the effects of significant process variations occur on a variety of scales including both the macro level scale and micro level scale. Defects also occur in different materials and on different layers on the wafer. These defects and significant process variations, of different scales, whether macro or micro, of different materials and layers, require accurate and repeatable detection and classification. In some instances this detection may require three dimensional multi-modality imaging.
The drive by semiconductor manufacturers and LCD manufacturers to smaller feature sizes, larger substrate sizes, and/or improved yield places greater challenges on identifying and resolving defects. Defect detection, classification, and the effective use of defect data play an essential role in yield learning during ramp-up and for process control during production. The technical challenges of improving and maintaining yields continues to grow as device sizes shrink, the number of die on each wafer increase, and the structure of circuit designs become more complex on more layers. Conventional methods of wafer inspection and material characterization are becoming unsuitable for these new challenges. A need has arisen to keep up with defect detection application demands .
Prior art solutions often overlook the way data is represented which in many cases is as critical, in computer image analysis systems for semiconductor wafer and LCD inspection, as the algorithms applied to the data. Simple algorithms applied to proper representations have the potential to be more powerful than sophisticated algorithms applied to improper representations. Encoding a digital wafer image as an array of data points reflects the measurements of the imaging sensors . But this raw data format is not suitable for most computer image analysis tasks. Alternatively, an image may be represented by its Fourier transform, with
operations applied to the transform coefficients rather than to the original data. This may be appropriate for some simple wafer material
characterization tasks, but inappropriate for others .
In addition, the existence of too much
information may create a fundamental problem in image data representation for wafer inspection.
Useful and irrelevant information are often mixed together in an image. In many cases the huge volume of irrelevant or redundant data is difficult or impossible for a computer to remove while retaining enough useful defect information.
SUMMARY OF THE INVENTION
The invention provides a semiconductor wafer or LCD inspection apparatus comprising a means for imaging the semiconductor wafer or LCD active matrix panel; an automated stage positioned to move the wafer under the means for imaging the semiconductor substrate; a substrate storage, loading and
unloading means for transporting the semiconductor wafer or LCD to and from the automated stage; and a means for image processing to scan the semiconductor wafer or LCD and to perform an image decomposition, processing and synthesis for automatic inspection of the semiconductor wafer or LCD.
In one embodiment of the invention, the means for image processing further comprises a means for decomposing the raw image data into a set of partial information channels, each channel reflecting predetermined aspects of the image.
In one embodiment of the invention, the means for image processing further comprises a means for synthesizing the image to recover the original image .
In one embodiment of the invention, the means for image processing further comprises a means for constructing an improved image.
In one embodiment of the invention, the means for imaging the semiconductor wafer or LCD further comprises an automated microscope.
In one embodiment of the invention, the means for imaging the semiconductor wafer or LCD further comprises a light source and a sensing device to image the area of the wafer under inspection.
In one embodiment of the invention, the means for image processing further comprises a host computer.
In one embodiment of the invention, the means for image processing further comprises multiple controllers.
In one embodiment of the invention, the means for image processing further comprises a high speed image processing unit .
The invention further provides a method for semiconductor wafer inspection comprising the steps of : obtaining an image of the semiconductor
substrate; performing an image decomposition to generate a decomposition data set of spatial
frequency and orientation bandpass component images that represent the decomposition of the image;
performing a multiple channel feature detection on the decomposition data set; performing an image synthesis on the decomposition data set to generate a synthesized image; and detecting defects in the synthesized image.
In one embodiment of the invention, the step of performing an image decomposition to generate a decomposition data set of spatial frequency and orientation bandpass component images further comprises the steps of: performing a linear lowpass filtering of the image to generate a lowpass image; performing a sub-sampling on the lowpass image to generate a coarser image; and repeating the above steps a predetermined number of times to generate a predetermined number of coarser images.
In one embodiment of the invention, the
invention further provides the steps of : expanding the coarser image to form an interpolated image; and adding the interpolated image to the next finer resolution image to generate a bandpass image.

In one embodiment of the invention, the step of expanding the coarser image further comprises the step of generating a point replicated image.
In one embodiment of the invention, the step of performing a linear lowpass filtering of the image to generate a lowpass image further comprises the step of performing an isotropic lowpass
decomposition.
In one embodiment of the invention, the step of performing a linear lowpass filtering of the image further comprises forming lowpass filters at
different directions.
In one embodiment of the invention, the step of performing an image synthesis on the decomposition data set to generate a synthesized image further comprises the steps of : expanding the coarser image to form an interpolated image; and adding the interpolated image to a bandpass image to generate a next finer resolution image.
The invention further provides the step of performing a multiband contrast enhancement step on the lowpass and bandpass images.
The invention further provides the step of performing a multiband noise coring to selectively remove high frequency components from the bandpass image .
The invention further provides the step of performing an edge preserving averaging on the lowpass image.
The invention further provides a method of image decomposition to generate a decomposition data set of spatial frequency and orientation bandpass component images further comprising the steps of : performing a linear lowpass filtering of an image to generate a lowpass image; performing a sub-sampling on the lowpass image to generate a coarser image; and repeating the above steps a predetermined number of times to generate a predetermined number of coarser images.
The invention further comprises the steps of: expanding the coarser image to form an interpolated image; and adding the interpolated image to the next finer resolution image to generate a bandpass image.

In one embodiment of the invention, the step of expanding the coarser image further comprises the step of generating a point replicated image.
In one embodiment of the invention, the step of performing a linear lowpass filtering of the image to generate a lowpass image further comprises the step of performing an isotropic lowpass
decomposition.
In one embodiment of the invention, the step of performing a linear lowpass filtering of the image further comprises forming lowpass filters at
different directions.
In one embodiment of the invention, the step of performing an image synthesis on the decomposition data set to generate a synthesized image.
In one embodiment of the invention, the step of performing an image synthesis on the decomposition data set to generate a synthesized image further comprises the steps of : expanding the coarser image to form an interpolated image; and adding the interpolated image to the bandpass image to generate a next finer resolution image.
Other objects, features and advantages of the present invention will become apparent to those skilled in the art through the description of the preferred embodiment, claims and drawings herein wherein like numerals refer to like elements.

BRIEF DESCRIPTION OF THE DRAWINGS
To illustrate this invention, a preferred embodiment will be described herein with reference to the accompanying drawings .
Figures 1A and IB show the method and apparatus of the invention for semiconductor wafer and LCD inspection using multidimensional image
decomposition and synthesis .
Figure 2 shows the method of the invention to generate the lowpass and bandpass decomposition of raw image data .
Figure 3 shows the spatial
frequency/morphological pattern spectra plot of the decompositions in three-dimensions.
Figure 4 shows the lowpass filters at different directions .
Figure 5 shows the directional decompositions derived from the lowpass decomposition.
Figure 6 shows the coarse to fine synthesis process of the invention.
Figure 7 shows the method of the invention to perform multiband local contrast enhancement.
Figure 8 which shows the method of the
invention to perform multiband noise coring.
Figure 9 shows a representation of a
semiconductor wafer substrate having replicated die patterns containing detailed image structure in the presence of a defect having scale and directional characteristics caused by a process defect.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT
Refer now to Figures 1A and IB which show the method and apparatus of the invention for
semiconductor wafer or LCD inspection using
multidimensional image decomposition and synthesis. The automatic wafer inspection system 10 comprises a wafer storage and loading module 18 and a wafer storage and unloading module 20, where both can be combined into one module, to transport a wafer 15 to and from an automated stage 16. The automated stage 16 moves the wafer 15 under a microscope 17 for scanning. The invention further provides an
illumination source 14 and a sensing device 12 to image the area of the wafer 15 under inspection. The illumination source 14 may further comprise a visible light source, an ultra-violet light source, an infrared light source or an electron beam source. The sensing device 12 is appropriately configured to receive and image from the illumination source 14. A processing unit 22 comprises a host computer 21, multiple controllers 23, a high speed image
processing unit 25 and a power supply 27. The sensing device sends a raw image 24 to processing unit 22 through connection 19. The processing unit 22 controls the scanning of the wafer 15 and
performs the image decomposition and synthesis processing for automatic inspection. The host computer 21 operates to facilitate the information extraction process by implementing the decomposition and synthesis method in software. Those skilled in the art will recognize that other methods of
implementing the decomposition and synthesis method may be used, including a hardware implementation. The invention decomposes the raw image data 24 into a set of partial information channels, each channel reflects certain aspects, a modality, of the
information. The partial information can thus be processed independently or cooperatively and
synthesized to recover the original image as
recovered image 24A or to construct an improved image 33. In this manner, useful information can be retained and irrelevant information can be rejected effectively.
Refer now to Figure IB which shows the method of the invention to perform automatic wafer defect classification. The method starts with the
acquisition of a wafer image 24 to provide the raw image data 24 of the wafer 15. The process then performs an image decomposition in step 26 on the raw image data 24 to generate a decomposition data set of spatial frequency and orientation bandpass component images that represent the decomposition of the raw image data 24. The method of the invention then performs multiple channel feature detection and enhancement on the decomposition data set in step 28. The invention then performs image synthesis in step 30 on information from the multiple channel feature detection and enhancement of the
decomposition data set 35. In step 32, the
invention performs final defect detection. Image defects 37 are provided in step 34 from the final defect detection step 32.
The representation used by the invention retains spatial localization while maintaining the localization in the spatial-frequency spectrum pattern and spatial orientation domains. The invention decomposes the raw image data 24 into a set of spatial frequency and orientation bandpass component images, represented by decomposition data set 35. Individual samples of a component image represent image pattern information that is
appropriately localized, while the bandpassed image as a whole represents information about a particular fineness of detail or scale. There is evidence that the early processing in the human visual system uses a similar representation. By selectively decomposing the data and enhancing certain
information, details which are not obvious can be highlighted and subsequently extracted. The result of the decomposition is a complete representation of the data, namely, the original raw image data 24 can be exactly recovered from this representation as recovered image 24A.
Particular semiconductor fabrication processes have characteristic defects. For example, spin depositions in semiconductor processes may have streaking. Or processes that are sensitive to contamination by particulates may contain small scale imperfections in the presence of larger scale structure and small features within that structure. By organizing image information, the structure of the characteristic difficulties of the process can be made more apparent, and easier to access. Images can be represented efficiently to allow access to the structural characteristics of the defects which are characteristic to the image processing problems and their associated processing structures. In particular, image processors can have feature calculations that have higher signal to noise and require less processing time when the image data is decomposed from a conventional representation and recomposed into structures that are appropriate to the defects which need to be detected.
Refer now to Figure 9 which shows a
representation of a semiconductor wafer substrate having replicated die patterns containing detailed image structure in the presence of a defect having scale and directional characteristics caused by a process defect. In the example of Figure 9, the image of the wafer is decomposed directionally and spectrally into a family of macro images.

Referring to the macro image 900, the defect 912 is a streak viewed in the context of circuit detail and large scale wafer structure. In a conventional image, a great deal of morphological processing and feature extraction effort must be expended to characterize the entire image. Within this
characterization lies not only the defect, but a number of artifact representations whose
characteristics are difficult to separate from the whole image characteristics because the whole image is viewed in its entirety. Alternatively, images 901 through 904 show directional filters that if applied to the image, produce filtered images 905 through 908, respectively. In this family of representations of the input macro wafer image, the defect properties are easily separable using
conventional image processing techniques.
The same input macro image is alternatively spectrally decomposed into low pass image 909, bandpass image 910, and high pass image 911. The large nature of the defect in contrast to the small scale features of the wafer design are apparent in a large object shown in image 909. Note that the spatial context of the defect has been maintained and can be studied in conjunction with the
directionally filtered images 905, 906, 907 and 908. The signal strength represents the strength of the features of the defect . Defect feature values come through intact for the most part after the
orientation filtering process. Features calculated for the remaining portions of the image are reduced in value. This increases the ability of the
invention to detect defects in the presence of other intended structures. Certain characteristics of the basic wafer image are also apparent. The basic image information has reduced contrast in the residual image compared to the residual contrast of the defect shown in filtered images 909, 905 or alternatively the defect may be rejected by the filtering the result of which is shown in image 910, 911, 900, 907 and 908. In both cases a separation has occurred. Thus, the invention separates defects from other image information. This separation reduces the difficulty of image processing to detect a defect by selection of the decomposition,
filtering, and synthesis operations on the input image information. The combination of the features of these images further strengthens feature strength for defect detection.
The importance of analyzing images at many scales arises from the nature of images, since the wafer 15 contains macro and micro defects and images of the wafer 15 are themselves variable, and
processes have different characteristics when viewed at different scales, and the designed circuit patterns contain features of many sizes. Types of macro and micro defects include: Gray Spot, Gray Streak, Gray Spot and Gray Streak, Particles, Multi-Layer Structure, Line Break, Subsurface Line,
Scratch, Hillocks, Grass, Worm-hole, Starburst, Speedboat, Orange Peel, Resist Gel Defect,
Controlled Collapse Chip Connection (C4) ,
Microbridge, Submicron, Micron, Micron Sphere, U. Pattern, Contamination, Protrusion, Break,
Intrusion, Nuisance,
Mask-Related (Shorts), Haze, Micro-contamination, Crystalline (Stacking Fault) , Spots, Break,
Recticle, Hard-Defects (Pinholes, Pindots,
Extrusions) , Semi-Transparent (Resist Residues, Thin Chrome) , Registration (Oversized, Undersized, Mislocated), Corner, Extra Metal, Metal Missing and Opens (Pattern Missing) . As a result, an analysis procedure that is applied only at a single scale may miss information at other scales. The invention provides a solution to this problem by isolating features at different scales and carrying out analysis at all scales simultaneously.
The invention provides for the organization of image information to facilitate rapid and enhanced image processing algorithms. Following the
arrangement of image data into a set of partial information channels, each channel of which
represents predetermined aspects of the raw image data, the image processing apparatus acts to detect and classify defects on the imaged substrate.
Defects are detected by conventional application of morphological operations to segment objects, and to characterize those objects by calculation of
features selected to differentiate between objects and artifacts. The calculated feature values for each object are the input to conventional
classifiers such as Fisher Tree classifiers or neural network classifiers . The output of the classifier may serve to simply detect a defect or alternatively to classify that defect within a set of possibilities. The construction and operation of these classifiers may be influenced by the features available from decomposed images which would
otherwise not be available from the original image, or be so difficult and time consuming to compute as to be impractical to obtain from the original image. The invention views objects of different types at different scales but at the same time preserves the spatial relations between objects so that object characteristics can be related easily between decomposed images . Other characteristics of these objects might be similarly derived and related such as direction, color, structure (e.g. periodicity), geometry, symmetry and so forth. Each
characteristic would be most simply seen in a decomposed image or set of images which individually were processed using conventional image processing techniques as described above.
An object represented by a combination of multiscale data can be analyzed by selectively correlating these data. The multiscale
representation provides a unique way to associate local information with global context by correlation between data of different scales. Thus, local features can be enhanced differently depending upon their global context encoded in the coarse
representations .
Based on the sampling theorem, different sampling is required to sufficiently store
information at different scales. Objects at fine scale require high image resolutions and objects at coarse scale require only modest image resolutions for storage. Therefore, the data in the multiscale decompositions can be stored at different
resolutions. Only a small amount of storage is required to store the global image information, and the representation at the finest scale is the only data requiring the full image resolution. Thus, the image decomposition can be efficiently stored in a multiresolution format. This lends itself to a coarse-to-fine processing strategy. Processing can be conducted at the coarse resolution to detect the objects of interest, and the follow-up processing can be applied only to those objects of interest. The dynamic processing strategy of the invention brings significant increases in efficiency and performance . The method and apparatus of the invention is thus able to handle multidimensional images or high resolution wafer images with
advantage since this type of problem demands a huge volume of storage and intensive computation.
The decomposition and synthesis image
representation and analysis method of the invention are related to the mask free primitive
characterization modules of the US patent
application entitled "Method and Apparatus for Mask free Semiconductor Wafer Inspection" by Lee, et al . which is incorporated in its entirety by reference thereto .
While the invention is described with image data in three dimensions, those skilled in the art will recognize that the methods of the invention can be equally applicable to higher dimensional image data or two dimensional data as well.
Refer now to Figure 2 which shows the method of the invention to the image decomposition step 26 described in Figure IB. In step 26 the invention generates lowpass and bandpass decomposition of image 24. The image decomposition data set 35 are data structures designed to isolate wafer image 24 features at different scales and orientations and to support efficient scaled neighborhood operations through reduced image representation. The image decomposition data set 35 comprises a lowpass decomposition stage, a bandpass decomposition stage, and a directional decomposition stage. The lowpass decomposition consists of a sequence of copies of the original image in which both sample density and resolution are decreased in regular or irregular steps. These reduced resolution levels of the decomposition are themselves obtained through a highly efficient iterative algorithm. The bottom, or zero-th level of the lowpass decomposition, L0, decomposition data image 36 is equal to the original image, the raw image data 24. This decomposition data image 36 is linearly or morphologically lowpass filtered 53 and down-sampled 55, in one example embodiment by a factor of two in each dimension, to obtain the next decomposition level, Lx,
decomposition data image 38. L is then filtered 57 in the same way and down-sampled 59 to obtain L2, decomposition data image 40. Further repetitions of the filter 61 and down-sample 63 steps generate the remaining lowpass decomposition levels 42. Thus,
L0 = I;
L± = D(F(Li.1) ) , i > 1,
where F(.) is a multidimensional lowpass filter operation and D(.) is a multidimensional down-sample operation. In one example embodiment, the lowpass filter can be a linear convolution filter or
nonlinear morphological filters such as the well known dilation, erosion, opening, closing, and other operations. The lowpass decomposition is equivalent to filtering the original image with a set of equivalent linear or nonlinear neighborhood
functions. The equivalent functions increase in width in proportion to down-sample factor with each level. In the case that the F(.) is a linear
Gaussian convolution with a five point kernel and a down-sample ratio of two in each dimension, the equivalent filters act as lowpass filters with the band limit reduced correspondingly by one octave with each level. The linear lowpass decomposition using Gaussian convolution is equivalent to the well known Gaussian pyramid data structure.

The bandpass decomposition, decomposition data images 44, 46 and 48, can be generated by
subtracting each lowpass decomposition level from the next lower level in the decomposition using subtractors 37, 39 and 41. Because these levels differ in their sample density, the invention interpolates, using expansions 45, 47 and 49, new sample values between those in a given level before that level is subtracted from the next lower level. In one embodiment, the interpolation can be achieved by point replication followed by a linear lowpass filtering. The levels of the bandpass
decomposition, B, decomposition data images 44, 46 and 48 can thus be specified in terms of the lowpass decomposition levels as follows :
B1 = A - L(EXP(L1+1) ) ,
where L(.) is a multidimensional linear lowpass filter and EXP ( . ) is a multidimensional data
replication.
Refer now to Figure 3 which shows the spatial frequency/morphological pattern spectra plot 59 of the decompositions in three-dimensions. Given a multidimensional image, the bandpass decomposition step 26 decomposes the image into different bands in spatial frequency or morphological pattern scale. The linear bandpass images, decomposition data images 44, 46, 48, 50, as with the Fourier
transform, represent pattern components that are restricted in the spatial-frequency domain. But unlike the Fourier transform, the images are also restricted to local volumes in the spatial domain. Spatial as well as spatial- frequency localization can be critical in the analysis of wafer images that contain multiple circuit pattern features and defects at multiple scales so that the characteristics of a single defect or pattern feature are extracted. Without this separate ability, the image confounds the characteristics of many defects or pattern features. In this way, individual samples of a component image represent image pattern information that is appropriately localized, while the bandpassed image as a whole represents information about a particular fineness of detail or scale.
Refer now to Figure 4 which shows the lowpass filters at different directions. Decompositions provide insight into the data in different
perspectives. As the bandpass decompositions reveal critical components of the wafer image patterns at different scales for analysis, these components can be further decomposed into different orientations. In this way, the information of different directions can be isolated at multiple scales. Different directional operators are applied to perform the decomposition. This is done by decomposing a linear lowpass function into the normalized additive combinations of different directional operators . As shown in Figure 4, the filter F is the normalized additive combinations of F1, F2, ... F8. Thus,

1 8
8 h
where the Fp's are lowpass filters 54, 56, 58, 60,

62, 64, 66, 68 and 70, at different directions. At a decomposition level, different directional
components can be generated along with an isotropic lowpass image which is an average of all the
directional components.
A lowpass decomposition in direction p at level p
i, L{ is generated by


and the isotropic lowpass decomposition at level i,
p
Ll 7 can be derived from Lf 's:



which is an average of all the directional lowpass
p
decompositions Lt 's. For example, consider the filter F shown in Figure 4.
1 M
F = ∑ Fp
8 p=i

The filter F is the normalized additive combinations of F1, F2, ... F8 representing lowpass filters 54, 56, 58, 60, 62, 64, 66, 68 and 70, at different directions. The isotropic lowpass image can then be used to generate the next level of decompositions including different directional components at that level. The directional bandpass images can be generated at each level in the same way the isotropic bandpass images are generated.
p
Thus, Bl is generated by subtracting L(EXP(LtP+1))

from Li :

B P = Li - L(EXP(LlP+x)) .

Those skilled in the art will recognize that
several image transformation or wavelet basis functions in image coding applications can be used as the generating kernels for the directional components of the invention. These image coding applications include the Harr transform, the
Quadrature mirror filters and others.
Now refer to Figure 5 which shows the
directional decompositions derived from the lowpass decomposition ^ . The lowpass decomposition image 110 is low pass filtered by filters 72, 74, 76, 78, 80, 82, 84 and 86 to generate the directional decompositions 90, 92, 94, 96, 98, 102, 104 and 106. The combination of the directional decompositions generate the bandpass decomposition 108 and the low pass decomposition 88.
Refer now to Figure 6 which shows the coarse to fine synthesis process of the invention. The bandpass and directional decomposition is a complete image representation; the step used to construct the decomposition may be reversed to synthesize the original image exactly. The bandpass decomposition can be recovered from different directional
components:

Bt = L - L(EXP(Lι+1))



1 M M
- ∑ ( , - UE P(L^ - - ∑

Thus, B1 can be recovered fully by averaging
p
over the directional decomposition 5, 's
decomposition data images 120, 122 and 124.
To recover hx decomposition data images 112, 114, 116 and 118, the lowpass decomposition level

L1+1, data images 114, 116 and 118, are interpolated, using expansions 130, 132 and 138, and added to B1 using summations 126, 125, and 136. This procedure can be repeated to recover L^, L1_2, and so on until the original image is recovered. Thus,
Lx = B^^ + L(EXP(L1+1)) for all i e {l,...,N-l}, and


An image can be sufficiently represented by its bandpass decomposition B0 through BN-1 and the top lowpass image LN. As an exact representation of the original image, bandpass decomposition may place the data in a more compact form so that the data can be stored, processed and transmitted more efficiently. Although the bandpass decomposition has a little more sample elements,
1
(2Λ-1)

in N dimensional and the regular down-sample by 2 case, than the original image, the values of these samples tend to be near zero, and therefore can be stored with a small number of bits. Further data compression can be obtained through quantitization: the number of distinct values taken by samples is reduced by binning the existing values. This results in some degradation when the image is recovered, but if quantitization bins are carefully chosen, the degradation can be reduced to an
acceptable level.
Another nice feature of the bandpass
decomposition is that the coarse resolution images, image data 118 and 116 for examples, are recovered first in the synthesis process. This invention provides a controlled synthesis scheme. Volumes of interest can be identified first at coarse
resolution and are synthesized subsequently at finer resolution as required in an analysis process. The data outside the volumes of interest are
sufficiently represented by its coarse resolution data and need not be synthesized further. Thus, only the data in the volumes of interest need to be progressively transmitted up to the finest
resolution. In this way, a controlled resolution strategy is set up which uses the lowest resolution sufficient for the application and restricts
processing of highest resolution image to volumes of interest. This methodology is in line with how human perception works. Humans view the world through sequences of fixations. They gather
information selectively and the vast majority of scene information is ignored.
The method and apparatus of the invention performs feature detection in step 28. Bandpass and directional decompositions based on different filtering schemes can isolate critical components of the wafer circuit patterns at different scales and orientations so that they are more accessible for analysis. When a linear Gaussian filter is used, the bandpass decomposition components include a highpass image and several bandpass images derived from taking the difference of images filtered by different sized equivalent Gaussian filters. These bands highlight image edge information at different scales. In fact, differences of Gaussian filters are approximations to the well known Laplacian of Gaussian filters. Hence, the zero-crossing points of the Gaussian based bandpass decomposition
correspond to the edges of different scales detected by the Laplacian of Gaussian edge detectors.
When a temporal bandpass filter is applied prior to an image decomposition, a Gaussian
decomposition can be used for motion detection. In motion detection, each decomposition band represents motions of different velocities from objects of different sizes.
The decomposition and synthesis image
representation and analysis disclosed in this invention can support the mask free primitive characterization modules disclosed in copending United States patent application entitled "Method and Apparatus for Mask free Semiconductor Wafer Inspection" by Lee, et al . which is incorporated by reference hereto. In addition to the linear
filtering, nonlinear filters such as morphological dilation, erosion, opening, and closing, etc. can be used for image decompositions. The bandpass
decompositions based on different morphologic filters can detect wafer defect patterns of
different kinds. Dilation decompositions detect dark edges of different scales; erosion decomposition detects bright edges of multiple scales; opening decomposition detects bright corners of multiple scales; and closing decomposition detects dark corners of multiple scales. The
directional decompositions further isolate the specific image features at different spatial
directions .
As the original wafer image can be exactly recovered by the image synthesis step 30 from the bandpass and directional decompositions, a modified image can be constructed by the synthesis of the modified bandpass and directional components. The modification can be performed to enhance features at desirable bands and directions and to reject
undesirable artifacts or to exclude certain
undesirable information. As an example, highpass filtered images of different scales can be generated by selectively excluding low frequency components of the images. This can be accomplished by a selective synthesis strategy. Highpass images of the smallest scale can be generated from B0 :
H = B0
which is equivalent to
L0 - L(EXP(LX) ) ,
and since L0 = I , the original image, and Li =
D(F(I) ) , thus,
H = I - (EXP(D(F(I) ) ) ) ,
which is the difference between the original image and its equivalent lowpassed image through the
L(EXP(D(F(.) ))) operation. The L (EXP (D (F ( . ) ) ) ) operation involves one lowpass filtering, a down-sampling operation, a replication and a linear filtering, This sequence defines one iteration of the equivalent lowpass filter.
The invention synthesizes a highpass filtered image of larger scale by
H' = B0 + L(EXP(BX) ) ,
which can be considered as an image recovery process starting from Bi rather than hx . Since


and
L2 = D(F(D(F(I) ))) ,
we have B-. = L-. - L (EXP (D (F (D (F (I) ) ) ) ) ) and


L(EXP(L(EXP(D(F(D(F(I) )))))) )
= I - L(EXP(L(EXP(D(F(D(F(I) ))))))) .
This can be recognized as the difference between the original image 24 and a lowpassed image by two iterations of the equivalent lowpass filters. The invention synthesizes other highpass images of larger scales by the selective synthesis strategy in a similar manner. The highpass images of small scale can enhance local image features such as peaks, pits, ridges, valleys, and the boundaries of objects. These are important features for wafer defect detection. As the scale of the highpass filtering increases, the objects themselves are enhanced.
One other useful selective synthesis strategy is the selective exclusions of certain frequency bands. A lowpassed image can be synthesized by excluding B0 band. Thus,
K = L (EXP (Li) ) = L0 - B0 = I - B0.
Bi can be excluded by
K' = B0 + L(EXP(P) ) ,
where P = L(EXP(L2)) . Since P = Lx - B1;
K' = B0 + L (EXP (Li - Bi) ) = B0 + L (EXP (Li) ) -

= L0 - L(EXP(B1)) = I - L(EXP(Bi)).
Similarly, other B±'s can be excluded by the selective synthesis strategies. Images excluding different bandpass channels can be used to exclude certain known circuit patterns in a wafer image.
In general, any combinations of LN and Bi's can be used in the selective synthesis process. The selections can be down to different directional components to synthesize only the features at the desirable orientations. This provides a general platform to derive and present information of different emphases from a given wafer image with certain circuit pattern features .
In addition to the selective synthesis strategy based on the initial bandpass decompositions, bandpass channels can be modified prior to the synthesis. This provides a richer and more powerful platform for wafer defect feature detection, enhancement, and regular circuit pattern and
variation artifact rejection.
The method and apparatus of the invention performs an independent enhancement then synthesis. The modification can be done independently at each band before synthesis is applied. In this way, information can be modified differently at different scales . There is evidence that the human visual system processes spatial-frequency information independently of the information in adjacent bands. Alternatively the independent enhancement then synthesis steps can be performed by multiband local contrast enhancement, multiband noise coring, or edge preserved averaging.
Now refer to Figure 7 which shows a method of the invention to perform multiband local contrast enhancement . Contrast enhancement can be performed independently at each scale by correlating between the lowpass image, lowpass decomposition image data 142, 144 and 146, and the bandpass image, bandpass decomposition image data 148, 150 and 152, at each decomposition level. Thus, the enhancement in the bandpass features is adaptively adjusted based on the lowpass images. The adjustment weights 154, 156 and 158 can be encoded as lookup tables with
bandpass and lowpass images as the input and the enhanced image as the output, contrast enhanced bandpass decomposition image data 160, 162 and 164. After the enhancement is performed at each band independently, the contrast enhanced bandpass decomposition can be used to synthesize a
reconstructed contrast enhanced image 168.
Refer now to Figure 8 which shows the method of the invention to perform multiband noise coring. The noise coring operation of the invention
selectively removes high frequency components from an image. Noise is suppressed by setting the bandpass or highpass image data value to zero when the noise value is below a predefined threshold. The predefined threshold is adjusted based on the decomposed image band. The noise suppressed
bandpass images 188, 190 and 192 are then added with the lowpass image 170 as described in Figure 6 to construct a noise reduced image 194. When
performing at a single scale, too much low-amplitude edge information may be removed along with the true noise, which causes edge transitions to look blurred and ragged. Alternatively, the invention decomposes the images and performs noise coring at each band independently and then synthesizes the image
following the method shown in Figure 8. The noise reduced image 194 is analogous to the reconstructed original image 112. The invention removes less image information while still reducing the equivalent amount of noise.
The invention also performs edge preserved averaging. A multiresolution representation of the original image can be rather useful in several applications. Although the lowpass decomposition can be used for this purpose, the edges are blurred at lower resolution because of the lowpass filtering effect. The invention constructs a multiresolution representation of the original image 24 based on an edge preserved blurring method. To accomplish this, the invention uses a directional decomposition method. In the directional decomposition, a
p
directional lowpass decomposition set Ll at level
p i and a directional bandpass decomposition set Bl_1

at level i-1 are constructed. The edge preserved lowpass image at level i, L ' x can then be generated by the following rule:

!,'(*) = LP(k)

when

DiB'tKk) < D(B^)(k)

for all q < p, where k is the index for a
multidimensional point and D(.) is the down-sampling operation. In this way, only the directional averaging yielding the lowest Dψfλ value will be

used for L'x. The method of the invention tends to preserve edges when averaging is performed.
The invention also performs between band correlation then synthesis . As the bandpass decomposition provides both spatial and spatial frequency localization, spatial neighborhood operators can be applied to different spatial frequency bands for detection and enhancement. An even more useful method is to perform operations between bandpass channels of the same image to detect and strengthen feature characteristics of a process .
Alternately, different embodiments of the invention provide a number of between band
correlations then synthesis methods. The
correlation strategies include a weakening strategy and an enhancement strategy. Information can be removed from a wafer image by a correlation between different bandpass decomposition bands. Given a frequency band, a weakening strategy constrains the data intensity by taking the minimum with its adjacent bands. The adjacent bands can be
reprocessed by neighborhood operations such as morphological dilations before correlation. In the case of using B1+1 to correlate B1, the method of the invention expands Bx+1 and dilates the result by an appropriate structuring element before performing the minimum operation with Bx :
B'x = MIN(BX, dilate (EXP (B1+1) ) ) .
In this way, the value in Bx is constrained to the dilated value of the expanded Bx+1.
The correlation of B1 can also be performed using the adjacent higher frequency bands B^ as follows:
B'x = MIN(BX/ dilate (D(BX.X))).
The weakening operations can be applied iteratively from the coarse (fine) resolutions to the fine (coarse) resolutions.
Image features can be enhanced by band correlation in a similar fashion. Given a frequency band, an enhancement strategy increases the data intensities by a constrained dilation operation. The constraint is imposed by correlating its dilated image with the data from its adjacent bands by the following rule:
B'x = MAX(BX, MIN (dilate (Bx) , EXP (B1+1) ) ) .
An alternative method is to add to the image by a fraction of the constrained dilation image:
B'x = Bx + *MAX(Bx, MIN (dilate (Bx) , EXP (Bx+1) ) ) .

The enhancement strategy picks up the
components in band Bx+1 having intensity values higher than the intensity value in band Bx . To further constrain the data for enhancement, only the data within the spatial dilation of band Bx is enhanced. The enhancement can be done by a maximum operation or by adding a fraction of the enhancement component to band Bx . In addition, the enhancement correlation can be performed using the adjacent higher frequency band Bi_1 as well. As with the weakening operations, the enhancement operations can be applied iteratively from the coarse (fine) resolutions to the fine (coarse) resolutions.
In wafer inspection applications, weakening and enhancement operations can be applied iteratively before the image synthesis step 30 is performed to selectively detect and enhance desirable features and patterns in the image and suppress the
undesirable information. Those skilled in the art will recognize that other methods of selective image enhancement may be used without deviating from the scope of the invention.
The following patents and patent applications are incorporated by reference hereto:
U.S. Patent Application Serial No. 08/571,686, filed December 13, 1995, for which the issue fee has been paid, entitled "Method For Identifying Normal Biomedical Specimens" to Nelson et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No. 07/838,064, filed February 18, 1992.
U.S. Patent No. 5,528,703, issued June 18, 1996, entitled "Method For Identifying Objects Using Data Processing Techniques" to Lee, which is a file wrapper continuation of abandoned U.S. Patent
Application Serial No. 07/838,395, filed February 18, 1992.
U.S. Patent No. 5,315,700, issued May 24, 1994, entitled "Method And Apparatus For Rapidly
Processing Data Sequences" to Johnston et al .
U.S. Patent No. 5,361,140, issued November 1, 1994, entitled "Method and Apparatus for Dynamic Correction of Microscopic Image Signals" to Hayenga et al .
U.S. Patent Application Serial No. 08/302,355, filed September 7, 1994, for which the issue fee has been paid, entitled "Method and Apparatus for Rapid Capture of Focused Microscopic Images" to Hayenga et al . , which is a continuation-in-part of abandoned U.S. Patent Application Serial No. 07/838,063 filed on February 18, 1992.
U.S. Patent No. 5,757,954, issued May 26, 1998, entitled, "Field Prioritization Apparatus and
Method" to Kuan et al .
Pending U.S. Patent Application Serial No.
08/927,379, filed September 12, 1997, entitled
"Apparatus for Automated Identification of Cell Groupings on a Biological Specimen" to Wilhelm et al . , which is a file wrapper continuation of
abandoned U.S. Patent Application Serial No.

08/309,061, filed September 20, 1994.
Pending U.S. Patent Application Serial No.
08/969,970, filed November 13, 1997, entitled
"Apparatus for Automated Identification of Thick Cell Groupings on a Biological Specimen" to Meyer et al . , which is a file wrapper continuation of
abandoned U.S. Patent Application Serial No.
08/309,116, filed September 20, 1994.
U.S. Patent Application Serial No. 08/667,292, filed June 20, 1996, for which the issue fee has been paid, entitled "Biological Analysis System Self Calibration Apparatus" to Lee et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No. 08/309,115, filed September 20, 1994.
Allowed U.S. Patent Application Serial No.
08/678,124, filed July 11, 1996, entitled "Apparatus for Identification and Integration of Multiple Cell Patterns" to Lee et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No. 08/308,992, filed September 20, 1994.
U.S. Patent No. 5,627,908, issued May 6, 1997, entitled "Method for Cytological System Dynamic Normalization" to Lee et al .
U.S. Patent No. 5,638,459, issued June 10,

1997, entitled "Method and Apparatus for Detecting a Microscope Slide Coverslip" to Rosenlof et al .
U.S. Patent No. 5,566,249, issued October 15, 1996, entitled "Apparatus for Detecting Bubbles in Coverslip Adhesive" to Rosenlof et al .
Pending U.S. Patent Application Serial No.
08/867,017, filed June 3, 1997, entitled
"Cytological Slide Scoring Apparatus" to Lee et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No. 08/309,931, filed September 20, 1994.
Allowed U.S. Patent Application Serial No.
08/309,250, filed September 20, 1994, entitled
"Apparatus for the Identification of Free-Lying Cells" to Lee et al .
U.S. Patent No. 5,740,269, issued April 14,

1998, entitled "A Method and Apparatus for Robust

Biological Specimen Classification" to Oh et al .
U.S. Patent No. 5,715,327, issued February 3, 1998, entitled "Method and Apparatus for Detection of Unsuitable Conditions for Automated Cytology

Scoring" to Wilhelm et al .
U.S. Patent Application Serial No. 08/526,138, filed September 5, 1995, for which the issue fee has been paid, entitled "Method for Testing Proficiency in Screening Images of Biological Slides" to Nelson et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No.
08/153,293, filed November 16, 1993.
U.S. Patent No. 5,699,794, issued December 23,

1997, entitled "Apparatus for Automated Urine
Sediment Sample Handling" to Fleck.
Pending U.S. Patent Application Serial No.
08/485,182, filed June 7, 1995, entitled
"Interactive Method and Apparatus for Sorting
Biological Specimens" to Lee et al .
U.S. Patent No. 5,647,025, issued July 8, 1997, entitled "Automatic Focusing of Biomedical Specimens

Apparatus" to Frost et al .
U.S. Patent No. 5,677,762, issued October 14,

1997, entitled "Apparatus for Illumination
Stabilization and Homogenization" to Ortyn et al . , which is a file wrapper continuation of abandoned

U.S. Patent Application Serial No. 08/309,064, filed September 20, 1994.

Pending U.S. Patent Application Serial No.
08/697,480, filed August 26, 1996, entitled
"Biological Specimen Analysis System Processing Integrity Checking Apparatus" to Ortyn et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No. 08/309,249, filed September 20, 1994.
U.S. Patent No. 5,715,326, issued February 3, 1998, entitled "Cytological System Illumination Integrity Checking Apparatus and Method" to Ortyn et al.
U.S. Patent No. 5,581,631, issued December 3, 1996, entitled "Cytological System Image Collection Integrity Checking Apparatus" to Ortyn et al .
U.S. Patent No. 5,557,097, issued September 17,

1996, entitled "Cytological System Autofocus
Integrity Checking Apparatus" to Ortyn et al .
U.S. Patent No. 5,499,097, issued 3/12/96, entitled "Method and Apparatus for Checking
Automated Optical System Performance Repeatability" to Ortyn et al .
U.S. Patent No. 5,692,066, issued November 25,

1997, entitled "Method and Apparatus for Image Plane Modulation Pattern Recognition" to Lee et al .
Pending U.S. Patent Application Serial No.
08/897,392, filed 7/21/97, entitled "Method and Apparatus for Highly Efficient Computer Aided
Screening" to Lee et al . , which is a file wrapper continuation of abandoned U.S. Patent Application Serial No. 08/315,719, filed September 30, 1994.
U.S. Patent Application Serial No. 08/472,389, filed June 7, 1995, for which the issue fee has been paid, entitled "Image Enhancement Method and
Apparatus" to Oh et al .
U.S. Patent No. 5,642,441, issued June 24, 1997, entitled "Apparatus and Method for Measuring Focal Plane Separation" to Riley et al .
U.S. Patent No. 5,625,706, issued April 29,

1997, entitled "Method and Apparatus for
Continuously Monitoring and Forecasting Slide and Specimen Preparation for a Biological Specimen
Population" to Lee et al .
U.S. Patent No. 5,745,601, issued April 28,

1998, entitled "Robustness of Classification
Measurement Apparatus and Method" to Lee et al .
U.S. Patent No. 5,671,288, issued September 23, 1997 entitled "Method and Apparatus for Assessing Slide and Specimen Preparation Quality" to Wilhelm et al .
U.S. Patent No. 5,621,519, issued April 15,

1997, entitled "Imaging System Transfer Function Control Method and Apparatus" to Frost et al .
U.S. Patent No. 5,619,428, issued April 8, 1997 entitled "Method and Apparatus for Integrating An Automated System to a Laboratory" to Lee et al .
U.S. Patent No. 5,781,667, issued July 14, 1998 entitled "Apparatus for High Speed Morphological Processing" to Schmidt et al .
U.S. Patent No. 5,642,433, issued June 24, 1997, entitled "Method and Apparatus for Image
Contrast Quality Evaluation" to Lee et al .
U.S. Patent No. 5,710,842, issued January 20,

1998, entitled "Method for Identifying Objects Using Data Processing Techniques" to Lee, which is a divisional of U.S. Patent No. 5,528,703, issued June 18, 1996.
Pending U.S. Patent Application Serial No.
08/788,239, filed January 25, 1997, entitled "Method and Apparatus for Alias Free Measurement of Optical Transfer Function" to Oh et al .

U.S. Patent No. 5,654,535, issued August 5,

1997, entitled "Cytological System Autofocus
Integrity Checking Apparatus" to Ortyn et al . , which is a divisional of U.S. Patent No. 5,557,097, issued September 17, 1996.
Allowed U.S. Patent Application Serial No.
08/784,316, filed January 16, 1997, entitled "Method and Apparatus for Detecting a Microscope Slide
Coverslip" to Rosenlof et al . , which is a divisional of U.S. Patent No. 5,638,459, issued June 10, 1997. Pending U.S. Patent Application Serial No.
08/767,457, filed December 16, 1996, entitled
"Method and Apparatus for Efficacy Improvement in Management of Cases With Equivocal Screening
Results" to Lee et al .
Pending U.S. Patent Application Serial No.
08/821,699, filed March 20, 1997, entitled
"Cytological System Autofocus Integrity Checking Apparatus" to Ortyn et al . , which is a divisional of U.S. Patent No. 5,654,535, issued August 5, 1997.
U.S. Patent No. 5,760,387, issued June 2, 1998, entitled "Cytological System Autofocus Integrity Checking Apparatus" to Ortyn et al . , which is a divisional of U.S. Patent No. 5,654,535, issued August 5, 1997.
U.S. Patent Application Serial No. 08/823,793, filed March 20, 1997, for which the issue fee has been paid, entitled "Cytological System Autofocus Integrity Checking Apparatus" to Ortyn et al . , which is a divisional of U.S. Patent No. 5,654,535, issued August 5, 1997.
U.S. Patent No. 5,763,871, issued June 9,

1998, entitled "Cytological System Autofocus
Integrity Checking Apparatus" to Ortyn et al . , which is a divisional of U.S. Patent No. 5,654,535, issued August 5, 1997.
Pending U.S. Patent Application Serial No.
08/821,694, filed March 20, 1997, entitled
"Cytological System Autofocus Integrity Checking Apparatus" to Ortyn et al . , which is a divisional of U.S. Patent No. 5,654,535, issued August 5, 1997.
Allowed U.S. Patent Application Serial No.
08/816,837, filed March 13, 1997, entitled
"Astigmatism Measurement Apparatus and Method Based on a Focal Plane Separation Measurement" to Riley et al . , which is a divisional of U.S. Patent No.
5,642,441, issued June 24, 1997, entitled "Apparatus and Method for Measuring Focal Plane Separation."
Pending U.S. Patent Application Serial No.
08/888,115, filed July 3, 1997, entitled "Method and Apparatus for Maskless Semiconductor and Liquid Crystal Display Inspection" to Lee et al .
Pending U.S. Patent Application Serial No.
08/888,120, filed July 3, 1997, entitled "Method and Apparatus for A Reduced Instruction Set Architecture for Multidimensional Image Processing" to Lee et al .

Pending U.S. Patent Application Serial No.
08/888,119, filed July 3, 1997 entitled "Method and Apparatus for Incremental Concurrent Learning in Automatic Semiconductor Wafer and Liquid Crystal Display Defect Classification" to Lee et al .
Pending U.S. Patent Application Serial No.
08/888,116, filed July 3, 1997, entitled "Method and Apparatus for Semiconductor Wafer and LCD Inspection Using Multidimensional Image Decomposition and
Synthesis" to Lee et al .
All of the above patent applications and patents are incorporated herein, in their entirety, by the foregoing references thereto.
The invention has been described herein in considerable detail in order to comply with the Patent Statutes and to provide those skilled in the art with the information needed to apply the novel principles and to construct and use such specialized components as are required. However, it is to be understood that the invention can be carried out by specifically different equipment and devices, and that various modifications, both as to the equipment details and operating procedures, can be
accomplished without departing from the scope of the invention itself.
What is claimed is :