LT5766B - Inspectation method - Google Patents

Abstract

An inspection method for fiat objects with repeating structures, especially wafers, comprising the steps of: Establishing a reference image; Taking an image of the object surface; and Detecting Defects on the object surface by comparison of the image of the object surface to the reference image; is characterized in that the establishing of a reference image is effected by: Taking images (38) of the surface of different sections (20) of a sample object, the sections being selected such that a plurality images of corresponding positions of the repeating structure are generated; Calculating a reference image (46) for all corresponding positions by computing the average, median or another representative value of the individual values of the images; and Replacing individual values of the reference image which are outside a tolerance range about the respective reference value of the reference image by the reference value.

Description

The invention relates to screening methods for screening planar objects with repetitive structures, in particular plates, comprising the steps of:

(a) creating a template image;

(b) selecting a surface image of the object; and (c) detecting defects on the object surface by comparing the object surface image with a template image.

In various industries, flat products are inspected for defects by optical imaging. As such, one could be wafer in the semiconductor and solar cell industry. The plates are discs of semiconductor, glass, tape or ceramic material. In certain applications, the plates are usually tested completely or at least over most of the surface. This check is called macro check. The lateral resolution needed to recognize popular defects increases with the advancement of mainstream technology. Typically, a resolution of 5 pm is necessary for macro-verification in subsequent technologies. At the same time, high throughput devices are preferred for wafer inspection.

Currently known macro-verification systems meet either the high throughput requirements or the desired resolution, but unfortunately not both at the same time. Therefore, there is a need for a faster macro-verification system with simultaneous improved resolution.

Similar requirements exist in other industries. Defective plates in the industry must be inspected during production. Imaging techniques that often display the entire image there are used to detect defects. In the electrical industry, mounting plates are inspected for defects in a series of objects subject to optical testing, especially mounting plates, for defects.

The need for rapid testing for a large number of commonly similar test objects is a common feature of all these applications. Such objects include mounting plates, semiconductor plates, solar cells, images, and the like. The use of sensors to generate large images of test objects is also a common feature of applications. Depending on the type of defect present, images must be generated by an optical image-capture system, including point-to-point sensors. Optical image capture systems are, for example, arrays or linear cameras. Thus, performance sensors are, for example, detectors for reflection of optical rays, microwaves or sound waves. Magnetic sensors can also be used.

State of the art

Typically, many plates or other objects of the same type are tested. In this case, known techniques use particularly good quality plates as a template that may not be defective. This is a disadvantage of such techniques because such defective plates are not always appropriate. Thus, automation of defect inspection and detection is not possible.

WO 00/04488 (by Rudolph) discloses a method for generating a template image using optical monitoring of a plurality of known good quality disks. Unknown plates are checked using a model that generates the use of such a template image.

US 4,644,172 (by Sandland) discloses a verification method in which a template image is manually selected through production experience and stored. Validation is performed on a pre-selected geometry and compared to the stored template image. Mean values and standard deviation are calculated.

The essence of the invention

It is an object of the present invention to provide a reliable automated method for detecting defects in which perfectly prepared plates and a plurality of good quality plates are not used to generate template images. According to the present invention, this object is achieved by creating a template image according to step (a) by:

(d) selecting images (38) of surfaces of different sections (20) of the sample object, the sections being selected to generate most of the repeating structure images at respective positions;

(e) calculating a template image (46) for all relevant positions, processing the data to obtain an average, midpoint, or other representative value of the individual values of the images; and (f) replacing the individual values of the template image outside the tolerance with the corresponding template value of the template image in accordance with the template value.

By axis, it can be seen that plates and many other objects have repetitive structures that can be used. Such repetitive structures are, for example, forms. Proper selection of sections for inspection ensures that each section corresponds to a plurality of corresponding sections at different positions on the same plate. In the best, non-defective object, all sections could produce identical images. Therefore, the relevant sections can serve as a template between them. Typically, defects occur on the object to be evaluated when the template image is defined by the corresponding section images. According to the present invention, this is done by replacing the individual values of the template image which are outside the tolerance with the corresponding template value of the template image according to the template value. In other words, if the value deviates significantly from the mean value, it is replaced by the mean value. The margin of tolerance is selected according to the circumstances. This can be customized or the value shifted dynamically. The process of the present invention does not require expensive, apparently defective plaques and does not require many good plaques.

The method is particularly suitable if the replaceable structures are on a stamped plate.

In a particularly valuable modification of the invention, the size of the sections in each direction is selected such that the repeatability of the sections corresponds to that of the repeating structures in each direction. The simplest example is an example where the section size is equal to the shape size. Because many different formats are available, the size of the shapes is essentially different to the size of the image. The size of the sections is therefore different than the size of the repeating structures. For example, the size of sections nx m (n, m integers) may be of the form size i x j (i, j integers). Then there are n x m different sections where the template image is defined for each section.

In a particularly convenient modification of the invention, the tolerance limits are multiple, in particular standard deviations of individual images of a stencil image, 2 to 4 times, and especially 3 times. The tolerance limits may also be expressed in the form of a threshold image as described below.

Rather, template image generation occurs at regular time intervals for each object type. So a long set time works and the values are taken into account. You can use the test image itself to generate a template image, where the template image is unnecessary and recognizable, notes a fixed trend over time, and can be automatically compensated for local defects at the same time.

In other modifications of the invention, the sections overlap. Thus, data processing errors at edges and edges can be reduced.

In a particularly valuable modification of the invention, the exemplary object is one of the objects being inspected at the same time. No separate template plate is required.

Preferably, the sections are selected such that multiple images are generated for individual positions in the repeating structures.

In a particularly preferred modification of the invention, the method further comprises the steps described below, wherein (a) the sample object images are further selected for each section that is offset relative to the original image of the section at a resolution below;

(b) calculating a threshold image for all images selected in step (a) by computing a standard deviation or other representative deviation of the individual values of the images, wherein the individual values for the deviation within the threshold image are replaced by the corresponding threshold value of the threshold image. and (c) detecting defects at a position where the difference between the selected image and the template image is greater than the corresponding threshold image threshold.

In this modification of the invention, an individual variable threshold is assigned to each point. The threshold at the edges and edges is slightly higher due to the use of moving images if the threshold is below. In this way, it is possible to operate at high sensitivity, corresponding to a low threshold, when no shape of artifact defects are found at the edges. The threshold is reached by selecting images that are shifted within the sub-pixel. The standard deviation of such an image is defined and a threshold is set for each of its points.

In another modification of the invention, the color variant images are selected for each section. Color variants images are generated at auto-masking surfaces that respond differently to process changes. They are also applicable for correction or for processing data with various parameters and / or algorithms. A plate with a thin oxide layer, for example, reflects in different colors with varying oxide layer thicknesses through interference, although this effect does not affect metal surfaces. Such strong effects of color variations can be detected automatically, for example, by defining a surface with a higher standard deviation or other appropriate statistical value for signal evaluation.

In a particularly good modification of the invention, the selected images are processed by image processing techniques and a certain level or bevel gain before calculating the mean or standard deviation or other typical value.

Other modifications of the invention are the subject of the dependent claims. The embodiment is described in detail below with reference to the accompanying drawings.

Brief description of the drawings

Fig. 1 is a schematic view of a portion of the observed surface with shapes where sections are marked.

Fig. 2 is a flow diagram illustrating a way of generating a template image (gold image).

Figure 3 a-d is a flow chart illustrating in detail the processing steps for generating a template image.

Fig. 4 is a flow diagram illustrating a method for processing image data to detect defects.

Description of Implementation of the Invention

Embodiments of the invention are described with reference to plates having repetitive shape structures. Such plates with shape structures are generally known. Figure 1 schematically illustrates a portion of a plate surface, generally designated by the number 1. The portion comprises a continuous thin line 12 formed by the shape structures 12. The shape structure in this portion represents - by way of example - 63 shapes obtained from columns 14 and 9 of columns 18.

To detect defects, the plate is examined by a suitable camera, such as a linear detector or CCD-area detector, and an appropriate optical kit known in the art. The image is generated by the camera section by section or divided into sections. Each section in this particular embodiment corresponds to 11 million pixels. As an example, the 12 sections are shown in Figure 1 using dashed lines 26, which are divided into 3 lines 22 and 4 columns 24.

The size of the sections 20 is chosen such that the width of the four sections is exactly the width of six shapes and the height of the three sections is exactly the height of the five shapes. In this way, the image is selected on the entire plate. It will be appreciated that the shape structure selected from section 28 corresponds to the section structure of section 30, which may be selected three sections below. There are a number of sections corresponding to the number of duplicate sections in the 3 * 4 pattern. Each of these matching sections can be used as a template for the other sections, provided the plate is accurate.

Figure 2 shows a basic method for generating a template image (gold image). First, any of the repeating plates is placed on the plate 32. The plate is adjusted on the plate 34. The plate is then moved to an agreed position 36. In this position, the surface image of the plate is selected 38. The selected image is corrected by common and known image processing techniques. due to imaging errors. Such errors include, for example, shadow correction or white balance. Shadow Correction is an optical fiber edge shadow correction. The white balance is an adjustment for the camera's dependence on wavelength errors. The plate is moved to another position when an image of section 36 is selected and an image of another section is selected 38. This is repeated for all sections 48.

It is then calculated for each section of the images if the section is "masked" 42, that is, if the defined values are used for each section, or if it is replaced by the central or mean corresponding other images. The center value or average is used for sections where the deviation of values is greater than the reported threshold from the center size or mean 44. Such values are then collected in the template image 46.

Figure 3 shows a more detailed calculation of the template image. After inserting the card, the row view camera image 50 is selected first. After Bayer-interpolation 52, a red image 54, a green image 56, and a blue image 58 are selected. Alternatively, the tri-color channel images can be directly generated by a 3-CCD camera. Further calculation is done separately for those three color channels. Those color images 52, 54, and 56 are copied to the respective memory 60, 62, and 64. In other words, the mean value is calculated from the values of the respective sections 66, 68, and 70 and spaced between 74, 76, and 78. Steps 36 and 50 through 70 are repeated for all sections 72. As a consequence, averages are obtained for each color and individual values for each section.

The stored individual images are then normalized to all averages for each image. For this purpose, each image is multiplied by the ratio of the average of all pixels of all images in the same color channel to the actual average pixel of the individual color image. This is necessary to eliminate large color trends that are inappropriate for defect detection. Such color trends can be generated, for example, by making small changes in the thickness of the top layers of the object being inspected. Notwithstanding this step, there would be differences between the individual relevant images and the distortion of the template position. The formula is shown in Figure 3b.

The next step is to add custom images. This step is called "add to buffer". The average image is then calculated by dividing the sum by the number of images. This step is called "1 / silver image CountMax".

For adjustable tolerance, the threshold is calculated for each individual image. This can be achieved by using only the images used to generate the golden images as shown in Figure 3. It can also be achieved by selecting additional images with a sub-pixel scroll that has been deliberately generated. Such sub-pixel scroll images are generated with a random positioning scroll inside a well defined area. They can be selected along the same path of image selection. For example: 1 * Positioning in exact position; 1 x shift positioning. They can be chosen by a different path afterwards. Moreover, it may be observed that the true shift between the respective images occurs within the limited positioning accuracy of each mechanical plate system. This impulse generates the same principled result as a consciously generated impulse. If this inevitable positioning inaccuracy is too small for the application images to be selected with conscious shifting, it can still be used. The purpose of this discussion is to raise the threshold to the edges and edges to the extent that small-edge products are interpreted as defects while the threshold value between them remains low to guarantee high sensitivity.

For threshold image calculations, the standard deviation is generated for each individual image using different rectangular minus mean values of the corresponding other images and normalized to the formula in Figure 3c. This Sdev - image (Sdev standard deviation) is obtained by multiplying by the "factor" parameter and using the "threshold" parameter for each individual image.

The difference between masking to generate each individual image is the difference between the individual image (Silver Image Buffer) and the average size image (Avg Buffer) compared to the corresponding threshold image. All difference values in the +/- threshold about the size of the mean are masked with "1" and contribute to the template image at a later stage. All other values of camouflage images are "0".

The "gold image" template image is formed for all individual images "silver images" by using all products "silver image" * "masking value" for each pixel value and normalizing to the number of images actively masked for that pixel (see figure 3d). In this way, only the active image set "1" is added to the stencil image and at the same time the normalization is maintained at the same level.

The color channel template images thus obtained can then be used to define the images to be inspected, either individually or by assembling the images to form a color image.

Figure 4 shows a more detailed detection of defects using a template image obtained in this way.

After loading and adjusting the plate, the plate is pulled so that the pattern of images to be inspected covering the entire surface is selected from the area to be inspected, which must be all or part of the plate. Image Selection Position is selected to customize positions for selecting template images. Test images are corrected for image errors such as shadow correction and whiteness with the same image processing methods and settings, and are used for template images.

The average or center of the image is calculated to recognize and compensate for major color trends and is normalized to the template image by using the stored average of the corresponding template image, that is, the template image that was generated for the actual section.

The difference between the check image and the stencil image is calculated and filtered with image processing techniques such as masking areas that are of no interest and doubled compared to the corresponding threshold image. In a doubled image generated only in this way, the pixel values where the difference between the inspected image and the template image is larger than the corresponding threshold at the position of that pixel are marked "1". All other values of the image are "0". All areas marked in this way are called "blobs". They are potential defects and may be further processed in other ways by a list of defects, visual features, and so on. for generating. In particular, morphological image processing steps are used to aggregate ("group") individual blobs, which obviously belong to common defects. Scratches, for example, in defect images are not shown with solid, clearly drawn lines, but have spaces. This grouping serves to recognize all the scratches from the linear parts and to sum them up. Many ways to accomplish such tasks have been researched and published.

Evaluation of the images being tested can be done with color images or with black and white images. Color images can be evaluated in color channels differently according to the color space system used for pre-RGB, HSV, and so on. The difference calculation must also be done with a suitable decimal system (distance definition), for example in 3-dimensional RGB space.

The above description was made through an example with exact values in terms of pixel count, section size and shape, and averaging and threshold calculations. It is to be understood, however, that such meanings describe only one of the many embodiments of the present invention. The invention has to be adapted to various ratios of section sizes to shape sizes. Also, the type of camera, tracked subject, resolution, averaging methods used, as well as center-to-average calculation methods, HSV or other channel separation instead of RGB color separation, and model size may change without departing from the main invention. Also provided herein are modifications suitable for special applications such as operation without sub-pixel shifting for primary threshold images or general image processing techniques such as morphological adjustment, filtering, or smoothing for further use without departing from the invention.

The same way can be used in another embodiment for black and white images. Alternatively, images selected with a color detector may be processed as a grayscale image instead of a computer processed color image according to DE-Bayer, where adjacent pixels in a grayscale image include the same signal values for different color areas.

The following positions of the drawings and a more precise process will help to better understand the invention.

Figure 2 illustrates the process of generating a gold image. 32 - Placing the plate on the platform, 34 - Plating the plate, 36 - Pushing the platform to the fold position, 38 - Finding the image, 40 - Shading the image, 42 - Calculating the mask for each image, 44 - Masking each image Calculation, 48 - n-fold repetition.

Figure 3a depicts Albatra image processing-gold image generation. Initial prerequisite: silver image CountMax = 20 and silver image Count = 0. Elsewhere in the text are positions: 36 - motion stage, 50 - row camera image 8U1, 52 deBayering, 54 - red image after deBayering 8U1, 56 - deBayer green image 8U1 , 58 - blue image after deBayering 8U1, 60 copy to red buffer, 62 copy to green buffer, 64 - blue buffer copy, 66 - calculator, 68 - calculator, 70 - calculator, 74 - red silver tool Vai Float (Count Max of silver image), 76 - Vai Float green tool (Count Max of silver image), 78 - Vai Float blue tool of count silver (Count Max of silver image),

- red silver image buffer 8U1 (silver image count max), 82 - green silver image buffer 8U1 (silver image count max), 84 - blue silver image buffer 8U1 (silver image count max).

FIG. 3b depicts Albatra Image Processing - Gold Image Generation (cont'd), condition i = 0. Elements not elsewhere specified in the text: 90 - Red silver image buffer (s) 8U1, 91 - Green silver image buffer (i) 8U1, 92 - Blue silver image buffer image buffer (i) 8U1, 93 - Normalizer All / measure (i), 94 - Normalizer All / measure (i), 95 - Normalizer All / measure (i), 96 - Add to buffer, 97 - Add to buffer, 98 - add to buffer, 99 - red silver amount - buffer image 32F1, 100 - green silver amount - buffer image 32F1, 101 - blue silver amount - buffer image 32F1, 102 1 / silver image Count Max, 103 - 1 / silver image Count Max, 104 1 / silver image Count Max, 105 - Avg bumper red 32F, 106 - Avg bumper green 32F, 107 - Avg bumper blue 32F.

Figure 3c shows a continuation of the process for all RGB channels (red, green, blue). Positions: 108 - Silver Image Buffer 8Ul (Silver Image Count Max), 109 - Avg Buffer 32F1, 110 - StdDev Calculation for each Image, 111 StdDev Image Buffer 32F1 (Silver Image Count Max), hereafter - masking generation where: 112 - Silver Image buffer 8U1 (silver image count max), 113 - Avg buffer 32F1, 114 - stdDev image buffer 32F1, (silver image count max), 115 - calculation of camouflage for each image, 116 - factor, threshold, 117 - camouflage buffer 8U1 (silver image) Count Max).

Figure 3d shows a continuation of the process. Same procedure for all RGB channels (red, green, blue). 118- silver image buffer, 119- camouflage buffer, 8U1 (silver image Count Max), 120 - Avg Calculation, 121 - gold image 8U1.

FIG. Figure 4 shows an overview of Albatra image processing. 122- gold image 8U3, 123 - image scan 8U3, selected according to scan plan, 124 - no difference, 125 - difference image 8U3, 126 - filter, 127 - threshold image8U3, 128 - filtered difference image 8U3, 129 - doubling, 131 - blob finding, 132- defect sheet (-PosX, -POsY, -space), 133 - included defects, 134 - defect result image 8U3, 135 - clustering, 136 cluster sheet (-posX, - Pos Y, - space, -defects) sheet), 137 - cluster engagement, 138 cluster result image 8U3.

Claims (12)

Definition of the Invention 1. A method of testing planar objects with repetitive structures, in particular plates, comprising the following steps: (a) generating a template image; (b) selecting an object surface image; and (c) detecting defects on the object surface by comparing the object surface image with the template image; characterized in that the step of generating the template image in step (a) is as follows: (d) selecting images (38) of surfaces of different sections (20) of the sample object, and selecting the sections such that most of the images are generated at the respective positions of the repeating structure; (e) calculating a template image (46) for all relevant positions by processing the data to obtain an average, midpoint, or other representative value of the individual values of the images; and (f) replacing the individual values of the template image which are outside the tolerance at the corresponding template value of the template image with the template value. 2. A method according to claim 1, characterized in that the repeating structures on the plate are shaped according to the shapes (14). 3. A method as claimed in claim 1 or 2, wherein the section size is selected in each direction such that the multiple of section (20) corresponds to the size of the repeating structure (14) in each direction. 4. A method as claimed in any preceding claim, wherein the tolerance limit is multiple, in particular 2 to 4 times, and in particular 3 times the standard deviation of the individual images of the template image. 5. The method of any preceding claim, wherein the generating a template image for each type of object is performed at regular intervals. 6. A method according to claim 3, wherein the size of the sections is not the size of the repeating structures. 7. A method according to any one of the preceding claims, characterized in that the sections overlap. 8. A method according to any one of the preceding claims, characterized in that the sample object is one of the objects being checked at the same time. 9. The method of any preceding claim, wherein the sections are selected such that multiple images are generated for individual positions of repeating structures. 10. The method of any preceding claim, further comprising: (a) further selecting images of the sample object for each section displaced relative to the original image of that section at a resolution below the resolution; (b) calculating a threshold image for all images selected by computing a standard deviation or other representative deviation of the individual values of the images for each section in step (a), wherein the individual values for the threshold values outside the tolerance at the corresponding threshold value of the threshold image are replaced. and (c) detecting defects at a position where the difference between the selected image and the stencil image is greater than the corresponding threshold image threshold. 11. A method as claimed in any preceding claim, wherein selecting a color image variation for each section. 12. A method as claimed in any preceding claim, wherein the selected images are processed by image processing techniques and, in particular, smoothes or bends the edge enhancement prior to calculating the mean or standard deviation or other typical value.