AOI detection method based on semiconductor chip consistency defect

Abstract

The invention discloses an AOI detection method based on semiconductor chip consistency defects, and the method comprises the following steps: 1, image training: obtaining N qualified chip pictures as training samples, carrying out the training of the image features, carrying out the statistics of the image features, and finally obtaining a mean image and a variance image for comparison, the step 1 comprises the following steps of: 1.1, carrying out image affine and alignment; step 1.2, image enhancement and blurring, and gray scale adjustment; 1.3, calculating a mean value variance; step 2, defect detection: comparing a chip image needing to be detected with a mean image and a variance image used for comparison, detecting an abnormal point diagram of the chip image needing to be detected and the mean image and the variance image, and if the abnormal point diagram exceeds a set threshold value, judging that the chip image needing to be detected is unqualified, and the step 2 comprises the following steps: step 2.1, performing image affine and alignment; 2.2, image enhancement and blurring, and gray scale adjustment are carried out; and step 2.3, defect connection and treatment.

Description

An AOI detection method based on semiconductor chip consistency defects Technical Field

[0001] This invention relates to the field of chip inspection technology, and specifically to an AOI inspection method based on semiconductor chip consistency defects. Background Technology

[0002] For semiconductor chip products, AOI (Automated Optical Inspection) is a common method for detecting chip defects. Simply put, AOI is a machine vision system that uses optical principles and image processing technology to automatically detect various defects on the surface of products (such as chips).

[0003] The main workflow of common AOI inspection includes the following steps: 1. Image acquisition: The high-resolution camera on the AOI device scans the chip under test under the illumination of a specific light source (such as LED lights) to obtain a high-definition image of its surface.

[0004] 2. Image Processing: The system preprocesses the acquired images, such as enhancing contrast and reducing noise, in order to more clearly identify features.

[0005] 3. Feature analysis and comparison: The system will compare the acquired images with standard images.

[0006] 4. Defect detection: By comparison, the system will identify any discrepancies with the standard image. For example, a component may be missing, solder joints may be misaligned, etc.

[0007] 5. Result Output and Processing: After the inspection is completed, the system will mark the location of all defects and display it on the monitor. At the same time, it can control the actuator to remove defective products from the production line.

[0008] AOI (Automated Optical Inspection) has many advantages such as high efficiency, high precision, and good stability, and is widely used in the semiconductor chip industry. For example, Chinese invention application with patent document number CN119000694 A discloses a "semiconductor chip manufacturing intelligent inspection and analysis system and method based on machine vision"; Chinese invention patent with patent document number CN115035101B discloses a "defect verification method, detection method and AOI equipment for semiconductor chip carrier".

[0009] Generally speaking, semiconductor chips, due to their tight manufacturing process, typically exhibit excellent consistency. However, during image capture, variations in light sources and camera noise can cause slight differences in the images. Image processing is relatively simple for chips without texture. For images with backgrounds, or even complex backgrounds, conventional image processing algorithms struggle. Therefore, a statistical processing method is needed that can effectively overcome errors caused by lighting conditions and accurately identify defects within the material.

[0010] In response to this, the inventor proposes the following technical solution. Summary of the Invention

[0011] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and propose an AOI detection method based on semiconductor chip consistency defects.

[0012] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: an AOI detection method based on semiconductor chip consistency defects, characterized in that the method includes the following steps: Step 1: Image training, acquiring N qualified chip images as training samples, training their image features, and statistically analyzing the image features to obtain the final mean map and variance map for comparison. Step 1 includes: Step 1.1: Image affine transformation and alignment; Step 1.2: Image enhancement and blurring, grayscale adjustment; Step 1.3: Mean and variance calculation; Step 2: Defect detection, comparing the chip image to be detected with the mean map and variance map used for comparison, detecting the abnormal point maps of the two, if the abnormal point maps exceed a set threshold, then judging that the chip image to be detected is unqualified. Step 2 includes: Step 2.1: Image affine transformation and alignment; Step 2.2: Image enhancement and blurring, grayscale adjustment; Step 2.3: Defect connection and processing.

[0013] Furthermore, in the above technical solution, step 1.1 includes: Step 1.1a: Obtain N qualified chip images as training samples; Step 1.1b: Select one image from the N training samples as a reference, select the region region0 with unique features in the image, perform Sobel algorithm gradient in the XY direction, and extract the Sobel image; Step 1.1c: Perform non-maximum suppression on the extracted Sobel image, that is, perform maximum convolution on the Sobel image to obtain the maxConv image, compare the maxConv image with the current image pixel by pixel, and set the positions of pixels smaller than the maxConv image to 0. The image after setting is the model image; Step 1.1d: Use the obtained model image as the template image, and record the center C0 of the template image region as the template center, and the angle angle0 = 0; Step 1.1e: Select any chip image from the remaining training samples, repeat the above steps, and finally obtain an image centered on C(x,y) as the template. Find the position closest to C(x,y), and the new contour point set is Cn(x,y). Record the matching score and position at this time; Step 1.1f: Adjust different angles, from the starting angle to the ending angle. Repeat the above steps 3.1b to 3.1e for each unit angle adjustment, and record the matching score, position, and angle; Step 1.1g: Take the position and angle corresponding to the highest matching score as the matched position; Step 1.1h: Subpixel interpolation, that is, perform bilinear subpixel interpolation on the image for the position with the highest matching score to obtain a weighted image.

[0014] Furthermore, in the above technical solution, step 1.2 includes: step 1.2a: performing two blurring processes on the weighted image using two different blurring parameters to obtain two blurred images; step 1.2b: subtracting the two blurred images to obtain the image edge features of key information and filtering out image noise; step 1.2c: adding the weakly blurred image to the subtracted image to obtain the enhanced image; step 1.2d: calculating the average value of the enhanced image and adjusting the brightness of the image to make the average brightness of each image consistent.

[0015] Furthermore, in the above technical solution, step 1.3 includes: Step 1.3a: Take one image from the N images processed in step 1.2 and calculate the median gray value of a pixel in the four neighborhoods of that pixel; Step 1.3b: Calculate the median gray value of the corresponding pixels in the remaining images in the same manner as in step 1.3a; Step 1.3c: Take the N median gray values ​​of the corresponding pixels as an array and calculate the mean and standard deviation of the array; Step 1.3d: Place the calculated mean value in the position of the corresponding pixel in the mean map and place the variance value in the position of the corresponding pixel in the variance map; Following steps 1.3da to 1.3d, traverse all pixels in all N images to obtain the complete mean map and variance map.

[0016] Furthermore, in the above technical solution, step 2.1 is as follows: align the chip image to be detected with the template image according to the method in step 1.1, and then subtract the mean image from the chip image in step 1.3 to obtain the difference image.

[0017] Furthermore, in the above technical solution, step 2.2 is: to perform image enhancement and brightness adjustment on the difference image obtained in step 2.1 according to the method in step 1.2.

[0018] Furthermore, in the above technical solution, step 2.3 is as follows: subtract the variance map from the difference map obtained in step 1.3 to obtain the difference map, and obtain the anomaly map. The area of ​​the anomaly map that is greater than the threshold is the abnormal area. Finally, the length, area and shape characteristics of the abnormal area are statistically analyzed, and the chip image to be detected is determined as unqualified according to the set threshold.

[0019] By adopting the above technical solution, this invention has the following beneficial effects compared with the prior art: This invention is highly applicable to the field of semiconductor chips where material consistency is good, solves problems that are difficult to handle with ordinary image processing algorithms, and can replace relatively complex deep learning training models, solving their slow optimization iteration problem. This invention only requires selecting a small number of samples to quickly statistically analyze features and achieve stable detection results, making it a very practical and reliable consistency detection method in the semiconductor field. Detailed Implementation

[0020] The present invention will be further described below with reference to specific embodiments.

[0021] In the chip industry, chip defects are intolerable. Failure to detect defective chips can lead to intractable problems. While chips generally exhibit high consistency due to their inherent precision characteristics, existing AOI (Automated Optical Inspection) methods are susceptible to image discrepancies caused by lighting variations. These discrepancies can result in significant errors in the final chip defect detection. This invention addresses this issue by proposing an AOI detection method based on semiconductor chip consistency defects. This method primarily utilizes image optimization processing to effectively overcome errors caused by lighting during AOI inspection, ultimately achieving accurate identification of defects within the chip and ensuring the quality of the supplied chips.

[0022] This invention mainly includes two steps: Step 1: Image training, acquiring N qualified chip images as training samples, training their image features, and statistically analyzing the image features to finally obtain the mean map and variance map for comparison.

[0023] Step 2: Defect detection. Compare the chip image to be inspected with the mean map and variance map used for comparison, and detect the outlier points in both maps. If the outlier points exceed the set threshold, the chip image to be inspected is determined to be unqualified.

[0024] In step 1, a small number of training samples are selected to quickly statistically analyze image features, obtaining mean and variance maps for comparison. In step 2, the same image processing method as in step 1 is used to process the actual chip image to be detected. The processed image is then compared with the mean and variance maps used for comparison, which can accurately identify defects in the chip.

[0025] The detection method of the present invention is described in detail below. Step 1 includes: Step 1.1: Image affine transformation and alignment. Specifically, Step 1.1 includes: Step 1.1a: Obtaining N qualified chip images as training samples. The shooting positions and angles of these N chip images are slightly different. Under normal circumstances, the present invention does not require a large number of training samples. For ease of explanation, N=30 is used as an example below.

[0026] Step 1.1b: Take the first image from the 30 training samples as the baseline and select the region region0 within that image that can serve as a unique feature. This unique feature refers to a feature that can be identified in the image. If it is not a unique feature, subsequent judgment will be difficult, such as a bump feature in a chip image. Then, perform the Sobel algorithm gradient in the XY directions on region0 to extract the Sobel image. The Sobel algorithm is a method that approximates the image gradient through simple convolution operations, thereby detecting the edges of objects in an image. By calculating the spatial gradient of image brightness, it highlights areas in the image where gray values ​​change sharply (i.e., the edges of the corresponding chip).

[0027] Step 1.1c: Perform non-maximum suppression on the extracted Sobel image, i.e., perform maximum convolution on the Sobel image with convolution coefficients of 1 and kernel size w*h to obtain the maxConv image. Compare the maxConv image with the current image pixel by pixel, and set the positions of pixels smaller than the maxConv image to 0. The image after setting is the model image.

[0028] Step 1.1d: Use the obtained model image as the template image, and record the center C0 of the template image region as the template center, with angle angle0 = 0.

[0029] Step 1.1e: Select any chip image from the remaining 29 training samples and repeat the above steps to obtain an image centered at C(x,y). Find the position closest to C(x,y), and the new contour point set is Cn(x,y). Record the matching score and position at this point. The matching score is calculated using NCC matching, and the specific formula is as follows:

[0030] C(i,j): Set of gradient points for template contour; Cn(x+i,y+j): Set of gradient points for the current search region contour; , This represents the mean of the template and the search area.

[0031] Step 1.1f: Adjust different angles, from the starting angle to the ending angle. Repeat steps 3.1b to 3.1e for each unit angle adjustment (e.g., every 0.01° interval), and record the matching score, position, and angle. Step 1.1g: Take the position and angle corresponding to the highest matching score as the matched position. Step 1.1h: Subpixel interpolation, that is, perform bilinear subpixel interpolation on the image for the position with the highest matching score. Bilinear subpixel interpolation is a mathematical method to estimate the pixel value of the image at non-integer coordinate positions by calculating the weighted average of the four pixels around the target point, and finally obtains a weighted image.

[0032] Step 1.2 involves enhancing, blurring, and adjusting the grayscale of the weighted image obtained in Step 1.1 to ensure accurate mean and variance calculations in subsequent steps. This includes: Step 1.2a: Blurring the weighted image twice using two different blur parameters to obtain two blurred images; Step 1.2b: Subtracting the two blurred images to obtain the key image edge features and filter out image noise; Step 1.2c: Adding the weakly blurred image from the two blurred images to the subtracted image to obtain the enhanced image; Step 1.2d: Calculating the average brightness of the enhanced image and adjusting its brightness to ensure consistent average brightness across all images.

[0033] For example: First, calculate the average brightness of the enhanced image obtained in step 1.2c; then, determine a target brightness, such as setting the target brightness value to 200.

[0034] Next, calculate the required brightness adjustment value grayValue = 200 – average brightness value; finally, add grayValue to the image to obtain the brightness-adjusted image, so that the average brightness of each image is consistent.

[0035] This image processing method can eliminate lighting differences and effectively reduce brightness differences caused by exposure, light source intensity, angle, etc.

[0036] Step 1.3 is the mean and variance calculation, which includes: Step 1.3a: Take one image from the 30 images processed in step 1.2, and calculate the median gray value of a certain pixel in the image, assuming it is the gray value of the four neighborhoods of pixel (0,0); Step 1.3b: Calculate the median gray value of the corresponding pixel (0,0) in the remaining images in the same way as in step 1.3a; Step 1.3c: Take the 30 median gray values ​​of the corresponding pixel (0,0) as an array, and calculate the mean and standard deviation of the array; Step 1.3d: Place the calculated mean value at the position of the corresponding pixel (0,0) in the mean map, and place the variance value at the position of the corresponding pixel (0,0) in the variance map; Following steps 1.3da to 1.3d, traverse all pixels in all 30 images to obtain the complete mean map and variance map.

[0037] At this point, the mean and variance plots for comparison are obtained in step 1. Next, we proceed to step 2. Step 2 involves using an AOI (Automated Optical Inspection) system to photograph the chip to be inspected, obtaining an image of the chip. This image is then compared with the mean and variance plots used for comparison. By determining whether the chip image is acceptable, we can ultimately determine whether the chip has defects. Step 2 includes: Step 2.1: Image Affine Mapping and Alignment; Aligning the chip image to be inspected with the template image according to the method in step 1.1, and then subtracting the mean plot from the chip image in step 1.3 to obtain the difference plot.

[0038] Step 2.2: Image enhancement and blurring, grayscale adjustment; perform image enhancement and brightness adjustment on the difference image obtained in Step 2.1 according to the method in Step 1.2.

[0039] Step 2.3: Defect Connection and Handling. Subtract the variance map from the difference map obtained in Step 1.3 from the difference map obtained in Step 2.2 to obtain the outlier map. Areas in the outlier map that are larger than the threshold are considered outlier areas.

[0040] Finally, the length, area, and shape of the abnormal regions are statistically analyzed, and a threshold is set to determine whether the chip image to be inspected is unqualified. If the chip image is unqualified, the chip to be inspected is determined to have a defect.

[0041] This invention is highly applicable to the field of semiconductor chips where material consistency is good. It solves problems that are difficult to handle with ordinary image processing algorithms, and can replace relatively complex deep learning training models, thus solving the problem of slow optimization iteration.

[0042] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

Claims

1. An AOI detection method based on semiconductor chip consistency defects, characterized in that, The method includes the following steps: Step 1: Image training. N qualified chip images are acquired as training samples, and their image features are trained. The image features are statistically analyzed to obtain the mean map and variance map for comparison. Step 1 includes: Step 1.1: Image affine transformation and alignment; Step 1.2: Image enhancement and blurring, grayscale adjustment; Step 1.3: Mean and variance calculation; Step 2: Defect detection. The chip image to be detected is compared with the mean map and variance map used for comparison to detect the abnormal point maps. If the abnormal point map exceeds the set threshold, the chip image to be detected is determined to be unqualified. Step 2 includes: Step 2.1: Image affine transformation and alignment; Step 2.2: Image enhancement and blurring, grayscale adjustment; Step 2.3: Defect connection and processing.

2. The AOI detection method based on semiconductor chip consistency defects according to claim 1, characterized in that: Step 1.1 includes: Step 1.1a: Obtain N qualified chip images as training samples; Step 1.1b: Select one image from the N training samples as a reference, select the region region0 with unique features in the image, perform Sobel gradient algorithm on the XY direction, and extract the Sobel image; Step 1.1c: Perform non-maximum suppression on the extracted Sobel image, that is, convolve the Sobel image to obtain the maxConv image, compare the maxConv image with the current image pixel by pixel, and set the positions of pixels smaller than the maxConv image to 0. The image after setting is the model image; Step 1.1d: Use the obtained model image as the template image, and record the center C0 of the template image region as the template center, and the angle angle0 = 0; Step 1.1e: Select any chip image from the remaining training samples, repeat the above steps, and finally obtain an image centered on C(x,y) as the template. Find the position closest to C(x,y), and the new contour point set is Cn(x,y). Record the matching score and position at this time; Step 1.1f: Adjust different angles, from the starting angle to the ending angle. Repeat the above steps 3.1b to 3.1e for each unit angle adjustment, and record the matching score, position, and angle; Step 1.1g: Take the position and angle corresponding to the highest matching score as the matched position; Step 1.1h: Subpixel interpolation, that is, perform bilinear subpixel interpolation on the image for the position with the highest matching score to obtain a weighted image.

3. The AOI detection method based on semiconductor chip consistency defects according to claim 2, characterized in that: The matching score is calculated using NCC matching, and the specific formula is as follows: C(i,j): Set of gradient points for template contour; Cn(x+i,y+j): Set of gradient points for the current search region contour; , This represents the mean of the template and the search area.

4. The AOI detection method based on semiconductor chip consistency defects according to claim 2, characterized in that: Step 1.2 includes: Step 1.2a: performing two blurring processes on the weighted image using two different blurring parameters to obtain two blurred images; Step 1.2b: subtracting the two blurred images to obtain the image edge features of key information and filtering out image noise; Step 1.2c: adding the weakly blurred image to the subtracted image to obtain the enhanced image; Step 1.2d: calculating the average value of the enhanced image and adjusting the brightness of the image to make the average brightness of each image consistent.

5. The AOI detection method based on semiconductor chip consistency defects according to claim 4, characterized in that: Step 1.3 includes: Step 1.3a: Take one image from the N images processed by step 1.2, and calculate the median gray value of a pixel in the four neighborhoods of the image; Step 1.3b: Calculate the median gray value of the corresponding pixels in the remaining image according to the method in Step 1.3a; Step 1.3c: Take the calculated median gray values ​​of the corresponding pixels as an array, and calculate the mean and standard deviation of the array; Step 1.3d: Place the calculated mean value in the position of the corresponding pixel in the mean map, and place the variance value in the position of the corresponding pixel in the variance map; Following steps 1.3da to 1.3d, traverse all pixels in all N images to obtain the complete mean map and variance map.

6. The AOI detection method based on semiconductor chip consistency defects according to claim 5, characterized in that: Step 2.1 is as follows: Align the chip image to be detected with the template image according to the method in step 1.1, and then subtract the mean image from the chip image in step 1.3 to obtain the difference image.

7. The AOI detection method based on semiconductor chip consistency defects according to claim 6, characterized in that: Step 2.2 involves performing image enhancement and brightness adjustment on the difference image obtained in step 2.1 according to the method in step 1.

2.

8. The AOI detection method based on semiconductor chip consistency defects according to claim 7, characterized in that: Step 2.3 is as follows: subtract the variance map from the difference map obtained in step 1.3 to obtain the difference map, and obtain the outlier map. The area of ​​the outlier map that is greater than the threshold is the outlier area. Finally, the length, area and shape characteristics of the outlier area are statistically analyzed, and the chip image to be detected is determined as unqualified according to the set threshold.

Images