Semiconductor chip defect detection method, device, equipment and storage medium

Abstract

The application provides a semiconductor chip defect detection method, device, equipment and storage medium, the method comprises: carrying out channel extraction and combination of multiple color spaces to the semiconductor chip image to be detected, carrying out segmentation processing on the chip image based on the combined color channel, obtaining multiple initial regions; statistical features are extracted for each initial region respectively, obtaining the feature set corresponding to each initial region, the statistical features include color features and spatial features; according to the feature set of each initial region, the multiple initial regions are merged, and multiple functional regions of the chip are obtained; for different functional regions, the corresponding feature subset is selected, the abnormality of the functional region is recognized, and the defect detection result of the chip image is obtained. The application avoids the problem that the unified detection strategy cannot consider the characteristics of different regions by recognizing different functional regions of the chip and selecting detection features accordingly, and improves the accuracy and reliability of defect detection.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, device, and storage medium for detecting defects in semiconductor chips. Background Technology

[0002] During semiconductor manufacturing, various defects may appear on the chip surface, such as scratches, contamination, oxidation, and particle residue. These defects can affect the chip's performance and reliability. To ensure product quality, defect detection of chips is necessary. Traditional manual inspection methods are inefficient and easily affected by subjective factors; therefore, automated visual inspection technology is gradually becoming the mainstream.

[0003] Most existing semiconductor chip defect detection methods employ a uniform feature extraction and detection strategy, using the same detection algorithm for the entire chip image. However, the surface of a semiconductor chip typically contains multiple functional regions with significant differences in material, structure, and texture characteristics. Using a uniform detection strategy makes it difficult to take into account the characteristics of each region, resulting in low detection accuracy and a high likelihood of missed or false detections. Summary of the Invention

[0004] The main objective of this invention is to solve the technical problem that the existing semiconductor chip defect detection methods have low detection accuracy due to the use of a uniform detection strategy. This invention provides a method for detecting defects in semiconductor chips, the method comprising: The semiconductor chip image to be detected is subjected to multi-color space channel extraction and combination, and the semiconductor chip image is segmented based on the combined color channels to obtain multiple initial regions; Statistical features are extracted from each initial region to obtain the feature set corresponding to each initial region; Based on the feature sets of each initial region, multiple initial regions are merged to obtain multiple functional regions of the semiconductor chip; By selecting appropriate feature subsets for different functional regions, anomaly identification is performed on the functional regions to obtain the defect detection results of the semiconductor chip image.

[0005] The present invention also provides a semiconductor chip defect detection device, the semiconductor chip defect detection device comprising: The image segmentation module is used to extract and combine channels in multiple color spaces from the semiconductor chip image to be detected, and to segment the semiconductor chip image based on the combined color channels to obtain multiple initial regions. The feature extraction module is used to extract statistical features from each initial region to obtain the feature set corresponding to each initial region. The region merging module is used to merge multiple initial regions according to the feature set of each initial region to obtain multiple functional regions of the semiconductor chip. The defect detection module is used to select corresponding feature subsets for different functional areas, perform anomaly identification on the functional areas, and obtain the defect detection results of the semiconductor chip image.

[0006] The present invention also provides a semiconductor chip defect detection device, comprising: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a circuit; the at least one processor invokes the instructions in the memory to cause the semiconductor chip defect detection device to perform the steps of the above-described semiconductor chip defect detection method.

[0007] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the steps of the above-described semiconductor chip defect detection method.

[0008] The aforementioned semiconductor chip defect detection method, apparatus, device, and storage medium extract and combine channels in multiple color spaces from an image of the semiconductor chip to be detected. Based on the combined color channels, the chip image is segmented to obtain multiple initial regions. Statistical features are extracted from each initial region to obtain a feature set corresponding to each initial region. These statistical features include color features and spatial features. The multiple initial regions are merged according to their feature sets to obtain multiple functional regions of the chip. For different functional regions, corresponding feature subsets are selected to identify anomalies and obtain the defect detection result of the chip image. This invention, by identifying different functional regions of the chip and selectively choosing detection features, avoids the problem of a unified detection strategy failing to adequately consider the characteristics of different regions, thus improving the accuracy and reliability of defect detection.

[0009] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0010] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the first embodiment of the semiconductor chip defect detection method in this invention; Figure 2This is a schematic diagram of a second embodiment of the semiconductor chip defect detection method in this invention; Figure 3 This is a schematic diagram of one embodiment of the semiconductor chip defect detection device according to the present invention; Figure 4 This is a schematic diagram of one embodiment of the semiconductor chip defect detection device in this invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0014] To facilitate understanding of this embodiment, a semiconductor chip defect detection method disclosed in this invention will first be described in detail. For example... Figure 1 As shown, this method includes the following steps: 101. Extract and combine channels in multiple color spaces from the semiconductor chip image to be detected, and segment the semiconductor chip image based on the combined color channels to obtain multiple initial regions; In this embodiment, the process of extracting and combining channels in multiple color spaces from the semiconductor chip image to be detected, and then segmenting the semiconductor chip image based on the combined color channels to obtain multiple initial regions includes: extracting different channels from the semiconductor chip image in multiple color spaces and combining them across spaces to obtain a reconstructed three-channel color representation; setting multiple initial cluster centers on the image plane for the three-channel color representation, calculating the comprehensive distance between the initial cluster centers and pixels within a preset range; and assigning cluster affiliations to the pixels based on the comprehensive distance, with pixels belonging to the same cluster center forming an initial region, thus obtaining the multiple initial regions.

[0015] Specifically, this embodiment uses an LED chip as an example. LED chips exhibit different visual characteristics in different color spaces. The electrode area of ​​the chip typically appears golden or silver-white, the light-emitting area appears grayish-brown, and the epitaxial area is similar in color to the light-emitting area but located at the chip edge. A single color space cannot simultaneously highlight the boundaries and defect features of all functional areas. The RGB color space directly corresponds to the physical signal captured by the camera, and its red channel is particularly sensitive to the contrast between the chip's electrode area and the light-emitting area. The HSV color space is closer to how the human eye perceives color; its saturation channel quantifies the purity of the color, while its brightness channel reflects changes in lightness and darkness.

[0016] The raw chip image is input into the color space conversion module, where RGB and HSV conversions are performed separately. The RGB conversion process directly reads the raw data from the camera sensor, separating the red, green, and blue channels. The HSV conversion requires calculating the maximum and minimum color component values ​​for each pixel, and then using the difference and ratio between these values ​​to calculate hue, saturation, and brightness. The red channel extracted from the RGB space is retained, while the saturation and brightness channels extracted from the HSV space are also retained. This selective extraction avoids interference from redundant information.

[0017] After extraction, the three channels are reassembled into a new three-dimensional data structure. The red channel occupies the first dimension, the saturation channel occupies the second dimension, and the brightness channel occupies the third dimension. This reconstructed three-channel representation retains the red channel's strong ability to distinguish the boundary between the electrode and the luminous area, the saturation channel's ability to capture high-saturation defects such as pinholes, and the brightness channel's ability to provide grayscale contrast for all boundaries and defects.

[0018] Next, region segmentation is performed on the reconstructed three-channel image. The segmentation process begins by uniformly distributing a set of cluster centers on the image plane. The distribution density of these centers determines the average size of the final generated initial regions. Each cluster center manages pixels within a certain radius around it, typically set to twice the distance between the cluster centers. For each pixel in its neighborhood, its similarity to the cluster center is evaluated. This similarity is determined by two factors: color proximity (the difference between the pixel's three channel values ​​and the center's three channel values) and spatial proximity (the distance between the pixel's coordinates and the center's coordinates).

[0019] Color proximity is measured by summing the squares of the differences between the three channels. A large difference in the red channel indicates that the pixel may not be assigned to the same area as the center in terms of electrode illumination; a large difference in saturation indicates that the pixel may be defective or made of different materials; and a large difference in brightness indicates significant contrast between light and dark areas. Spatial proximity is measured by summing the squares of the horizontal and vertical distances. Pixels farther from the center, even if they are similar in color, should not be assigned to that center.

[0020] Color difference and spatial distance are integrated into a comprehensive metric, and the influence of the two factors is balanced through a weighted approach. The weight coefficient determines whether the algorithm prioritizes color consistency or spatial compactness. In this embodiment, the weight is set biased towards color similarity, ensuring that the region boundaries closely align with the actual edges of the chip's functional areas and defects. After calculating the comprehensive metric for all pixels and their accessible cluster centers, each pixel is assigned to the center with the smallest metric value. After assignment, the position of each cluster center is recalculated, and the new position is taken as the mean of the three-channel color and the mean of the coordinates of all pixels belonging to that center. This process is iterated repeatedly, and the cluster centers gradually move to the positions that best represent their regional characteristics.

[0021] Each initial region consists of a group of spatially adjacent pixels with similar colors. The initial region within the electrode area contains pixels with high red content and high saturation, appearing visually as a yellowish-white tone. The initial region within the luminescent area contains pixels with medium red content and low saturation, appearing visually as a grayish-brown tone. The epitaxial region is similar in color to the luminescent region, but due to its location at the chip edge, it is spatially distinct from the initial region of the luminescent area.

[0022] Furthermore, the step of extracting different channels from the semiconductor chip image in multiple color spaces and combining them across color spaces to obtain a reconstructed three-channel color representation includes: performing RGB color space conversion and HSV color space conversion on the semiconductor chip image; extracting the red channel from the RGB color space; and extracting the saturation channel and the brightness channel from the HSV color space; and combining the red channel, the saturation channel, and the brightness channel to form a reconstructed three-channel color representation.

[0023] Specifically, the raw images of LED chips captured by the camera are typically stored in RGB format, with each pixel containing three color components: red, green, and blue. This representation directly corresponds to the physical structure of the camera sensor, but it is not suitable for direct chip defect detection. This is because there is a strong correlation between the three RGB channels; the green and blue channels carry less effective information in the LED chip image, while the main regional contrast and defect features are concentrated in the red channel.

[0024] The red channel is particularly sensitive to the electrode area of ​​the LED chip. The electrode material is typically metal, which produces strong red reflection under illumination, resulting in a higher pixel value for the electrode area in the red channel. In contrast, the light-emitting area is made of semiconductor material, which reflects red light less strongly, leading to a significantly lower pixel value in the red channel. This strong contrast makes the red channel a key indicator for distinguishing between the electrode and light-emitting areas. While the green and blue channels also exhibit some contrast, it is far less pronounced than that of the red channel, and retaining these two channels would actually introduce noise interference.

[0025] However, relying solely on the red channel is insufficient to handle all types of defects. Pinhole defects on the electrode surface are not prominent in the red channel because the difference in red component between the pinhole area and the surrounding normal electrode area is small. The main characteristic of this type of defect lies in color purity. Due to material loss or oxidation, the color purity at the pinhole is reduced, resulting in a grayish visual effect. Color purity corresponds precisely to the saturation component in the HSV color space. A high saturation value indicates a vivid and pure color, while a low value indicates a color close to gray. Converting an RGB image to the HSV space and extracting its saturation channel can effectively capture these defects with abnormal color purity.

[0026] The HSV color space conversion process involves relatively complex mathematical operations. For each pixel, the maximum and minimum values ​​of its RGB components are first identified. The difference between these values ​​reflects the vibrancy of the color; the larger the difference, the more vibrant the color. Dividing this difference by the maximum value yields the saturation. This normalization process ensures that saturation is unaffected by changes in brightness. The normal portion of the electrode area has high saturation due to its metallic luster; the pinhole area shows a significant decrease in saturation due to material degradation. This difference manifests as localized low-value areas in the saturation channel, facilitating detection algorithms.

[0027] Besides saturation, the luminance channel in the HSV color space also contains important information. The luminance channel is essentially the maximum value of the three RGB components, reflecting the overall brightness of the pixel. The functional area boundaries of an LED chip often exhibit abrupt changes in brightness; the electrode area is typically brighter than the light-emitting area, while the epitaxial area, located at the chip edge, may have lower brightness. These brightness variations create sharp edge contours in the luminance channel. Irradiation defects within the light-emitting area usually manifest as localized brightness anomalies because material structure damage at the irradiation point alters optical properties. The luminance channel is more sensitive to these defects than the red channel.

[0028] After completing the RGB to HSV conversion, the saturation and luminance channels are extracted. At this point, three independent channels are available: a red channel from the RGB space, and a saturation and luminance channel from the HSV space. These three channels characterize the chip image from different perspectives, but each is only a single-channel grayscale image. To allow the detection algorithm to utilize this information simultaneously, they need to be reassembled into a three-channel color representation.

[0029] The assembly process is relatively straightforward. A new three-dimensional array is created, with the same dimensions as the original image and a depth of three. Data from the red channel is filled into the first depth layer, data from the saturation channel into the second depth layer, and data from the brightness channel into the third depth layer. This constructs a new "color" image, which, although visually completely different from the original image, contains richer information for defect detection.

[0030] In this reconstructed color space, the electrode area exhibits a combination of high red value, high saturation, and high brightness; the luminescent area exhibits a combination of medium red value, low saturation, and medium brightness; and the epitaxial area exhibits a combination of medium red value, low saturation, and low brightness. Various defects also display unique characteristic patterns: pinholes exhibit high red value but low saturation, perforations exhibit abnormal brightness values, and scratches exhibit abrupt changes in red and brightness values ​​along a certain direction.

[0031] 102. Extract statistical features from each initial region to obtain the feature set corresponding to each initial region; In this embodiment, all pixels within each initial region are traversed, and their values ​​in the reconstructed three-channel color representation are read. The values ​​of all pixels in the red channel within the same region are summed and then divided by the total number of pixels in that region to obtain the average value of the red channel. The average values ​​of the saturation and brightness channels are calculated using the same method. These three average values ​​constitute the color feature vector of the initial region. Since the pixels within the same initial region are relatively similar in color, the calculated average values ​​can represent the color attributes of the entire region well. The initial region within the electrode area typically has a higher average red value and saturation value, while the initial region within the luminescent area exhibits a lower average saturation value.

[0032] Besides simple mean statistics, other statistics can be calculated to enrich the description of color features. For example, calculating the standard deviation of pixel values ​​within a region reflects the uniformity of color within that region. A smaller standard deviation indicates a concentrated color distribution within the region, suggesting better region quality; a larger standard deviation indicates color fluctuations within the region, potentially containing pixels with different attributes or defective edges. Another example is calculating the median of pixel values ​​within a region. Compared to the mean, the median is more robust to outliers and less susceptible to the influence of a few extreme values. In practical applications, the selection of which statistics to extract depends on the required detection accuracy. More statistics result in a more refined feature description, but also increase the computational load.

[0033] Extracting spatial features requires calculating the geometric properties of the initial region. The most basic spatial feature is the center coordinates of the region. This is obtained by averaging the x-coordinates of all pixels within the region and averaging their y-coordinates. These center coordinates reflect the region's location within the chip image. Regions located at the chip's edge often differ functionally from those in the chip's center; for example, epitaxial regions are always located at the chip's edge. Besides the center coordinates, the area of ​​the region, i.e., the total number of pixels it contains, can also be calculated. Area information helps filter out some noise regions that are too small. The aspect ratio of the region is also a useful shape feature, obtained by calculating the aspect ratio of the region's bounding rectangle. A long, narrow region may correspond to linear defects, while a square region may correspond to regular structures such as electrodes.

[0034] After feature extraction is completed for a single region, the extracted feature values ​​are organized into a feature vector. This feature vector serves as the digital representation of the initial region and is stored in the feature set. By iterating through all initial regions and generating a corresponding feature vector for each region, a feature set containing features from all initial regions is finally obtained.

[0035] 103. Merge multiple initial regions based on the feature sets of each initial region to obtain multiple functional regions of the semiconductor chip; In this embodiment, the step of merging multiple initial regions based on the feature sets of each initial region to obtain multiple functional regions of the semiconductor chip includes: determining the neighborhood relationship between each initial region through a region expansion operation; calculating a similarity metric between adjacent initial regions based on a first feature subset in the feature set; performing preliminary merging of adjacent initial regions based on the similarity metric to obtain multiple preliminary merged regions; performing a classification operation on the multiple preliminary merged regions based on the first feature subset to divide the multiple preliminary merged regions into multiple candidate functional regions according to a preset number of categories; and correcting the boundaries of the multiple candidate functional regions to obtain multiple functional regions of the semiconductor chip based on the corrected boundaries.

[0036] Specifically, the initial number of regions obtained from the preliminary segmentation is usually in the hundreds to thousands. These small regions need to be merged into several large functional regions such as electrode regions, luminescent regions, and epitaxial regions according to their functional attributes. The merging process is divided into three stages: neighborhood determination and preliminary merging, cluster classification, and boundary correction.

[0037] The first stage requires determining which initial regions are adjacent. Morphological dilation is used to determine neighboring regions, extending the boundary of each initial region outwards by a certain number of pixels; this expanded region is called the extended region. The extended region is then checked for overlap with other initial regions. If the extended region overlaps with an initial region, the two regions are considered adjacent. The number of pixels extended determines the strictness of the neighboring region determination; a larger extension range results in more pairs of regions being identified as adjacent. In this embodiment, the extension range is set to three pixels, which captures true adjacency relationships without incorrectly classifying distant regions as adjacent.

[0038] After determining neighborhood relationships, the similarity between each pair of adjacent regions is calculated. The similarity is calculated based on a first feature subset, which includes brightness and spatial location features. Brightness features reflect the lightness or darkness of a region; regions within the same functional area should have similar brightness. Spatial location features consider the relative position of the region within the chip; the epitaxial region is always located at the chip edge, showing a significant difference in positional features compared to the central region. The absolute value of the difference in the mean brightness between two adjacent regions is calculated; the smaller the difference, the more similar the two regions are in brightness. Simultaneously, the distance difference from the center coordinates of the two regions to the chip center is calculated; a smaller distance difference indicates that the two regions are spatially close. The brightness and location differences are combined to obtain a similarity metric; a smaller metric indicates a higher similarity between the two regions.

[0039] The process iterates through all adjacent regions. If the similarity metric of a pair of regions is below a set threshold, the two regions are merged into one. The merging operation combines the pixel sets of the two regions and recalculates the feature values ​​of the merged region. This process is repeated, with each merge reducing the total number of regions while increasing the area of ​​each individual region. After multiple rounds of merging, the initially fragmented regions gradually coalesce into several larger preliminary merged regions. The pixels within these preliminary merged regions are relatively consistent in brightness and position, but they do not yet have clear functional area labels.

[0040] The second stage performs a classification operation on the initially merged regions. LED chips typically contain three functional areas: electrode area, light-emitting area, and epitaxial area. The k-means clustering algorithm is used to divide the initially merged regions into three categories. The first feature subset, namely brightness feature and spatial location feature, is still used during the clustering process. The algorithm first randomly selects three initially merged regions as initial cluster centers, representing the three functional areas respectively. The feature distance from each initially merged region to the three cluster centers is calculated, and the region is assigned to the nearest cluster center. After assignment, the feature mean of all regions within each category is recalculated as the new cluster center. This process is iterated and updated until the cluster centers gradually stabilize at positions that represent each functional area. After the iteration stops, each initially merged region obtains a category label, and regions with the same label belong to the same functional area.

[0041] However, the functional region boundaries obtained through direct clustering may contain inconsistencies. Some initially merged regions located at the functional region boundaries may be misclassified, resulting in jagged or gap-like boundaries. The third stage addresses this issue through boundary correction. A morphological closing operation is performed on each candidate functional region, consisting of two steps: dilation and erosion. The dilation operation expands the region boundaries outward, filling small gaps; the erosion operation shrinks the boundaries inward, restoring them to their approximate original positions while eliminating minor protrusions. After the closing operation, the boundaries become smoother and more continuous.

[0042] The smooth boundary obtained from the closing operation is used as a reference boundary and compared with the boundary obtained from the original clustering. If the deviation between the original boundary and the reference boundary at a certain point exceeds a preset threshold, it indicates that there may be a classification error at that boundary. In this case, the original boundary at that point is replaced with the reference boundary, and the pixels that originally belonged to the wrong category are reassigned to the correct functional area. This correction operation mainly targets small areas near the boundary and has little impact on large areas inside the functional area. After correction, the boundary of each functional area conforms to the actual visual characteristics and maintains geometric regularity.

[0043] 104. Select corresponding feature subsets for different functional regions, perform anomaly identification on the functional regions, and obtain the defect detection results of the semiconductor chip image.

[0044] In this embodiment, the step of selecting corresponding feature subsets for different functional regions, performing anomaly identification on the functional regions, and obtaining the defect detection result of the semiconductor chip image includes: determining the target feature subset corresponding to each functional region from the second and third feature subsets of the feature set according to the category identifier of each functional region and a preset feature selection rule; inputting the target feature subset of the initial region within each functional region into a pre-trained anomaly judgment model corresponding to the category of the functional region, and calculating the anomaly judgment value; based on the comparison result of the anomaly judgment value and a preset judgment threshold, marking the initial region whose anomaly judgment value exceeds the preset judgment threshold as an anomaly region, and integrating the anomaly regions within all functional regions to obtain the defect detection result of the semiconductor chip.

[0045] Specifically, the types and manifestations of defects differ significantly across different functional areas, thus requiring different detection strategies for each area. The electrode area is prone to pinhole defects, which are prominently reflected in saturation characteristics; the luminescent and epitaxial areas are prone to perforations, foreign matter, and other defects, which are clearly reflected in the red channel and brightness characteristics.

[0046] First, based on the functional region segmentation results, a category identifier for each initial region is assigned. The three functional regions—electrode region, luminescent region, and epitaxial region—are traversed, and the list of initial regions contained in each functional region is read. For the electrode region, according to the feature selection rules, a second feature subset is extracted from the feature set as the target feature subset for that region. The second feature subset contains two feature dimensions: mean saturation and mean brightness. Normal portions of the electrode region typically maintain a high level of saturation due to their metallic luster; at pinhole defects, material loss or oxidation significantly reduces saturation. Brightness features reflect changes in reflective intensity on the electrode surface, aiding in the determination of defect locations.

[0047] For the luminescent and epitaxial regions, a third feature subset is extracted according to the feature selection rules. This third feature subset includes the mean red channel value and the mean brightness value. The luminescent region is composed of semiconductor material, and the normal portion exhibits a relatively uniform medium intensity value in the red channel. Due to material breakdown, perforation defects cause the red channel value in localized areas to deviate from the normal range. Foreign matter defects, depending on the foreign matter material, may manifest as excessively high or low red channel values. Brightness characteristics also play an important role in these two functional regions, as defects are often accompanied by abnormal changes in local brightness.

[0048] After feature subset selection, the target feature subsets of the initial regions within each functional area are input into the corresponding pre-trained anomaly detection models. These models, trained using a large number of normal chip samples, learn the feature distribution patterns of normal regions. During the detection phase, the model receives the feature vector of the region to be detected as input and calculates the degree to which the region deviates from the normal distribution. The electrode region uses a model specifically trained for the electrode region, which has mastered the distribution range of normal electrode regions in the saturation and brightness feature spaces. The luminescent and epitaxial regions each use specially trained models, which are familiar with the normal distribution of their respective functional areas in the red channel and brightness feature spaces.

[0049] The model's output is an anomaly determination value, the magnitude of which reflects the probability of an anomaly in the initial region. A larger determination value indicates that the region's features deviate further from the distribution center of normal samples, and the higher the probability of a defect. A smaller determination value indicates that the region's features fall within the normal distribution range, and the region is more likely to be normal. The model calculates the determination value based on the decision boundary learned during the training phase, which separates normal regions from potentially anomalous regions in the feature space.

[0050] After obtaining the anomaly determination value for each initial region, a determination threshold needs to be set to distinguish between normal and abnormal conditions. The choice of threshold affects the detection sensitivity. A lower threshold increases detection sensitivity and can capture more suspected defects, but may also generate more false alarms; a higher threshold reduces the false alarm rate, but may miss some subtle defects. In this embodiment, different determination thresholds are set for different functional areas based on the quality requirements in actual production. Due to the strict requirements for pinhole defects in the electrode area, the threshold is set relatively low; the thresholds for the light-emitting area and the epitaxial area can be appropriately relaxed.

[0051] The process iterates through the anomaly detection values ​​of each initial region, marking areas where the values ​​exceed the threshold of the corresponding functional area as abnormal regions. During marking, the pixels of the abnormal regions are highlighted in a specific color on the image for easy viewing by inspectors. Anomalies marked in the electrode region primarily correspond to pinhole defects, those in the light-emitting region may be perforations or foreign objects, and those in the epitaxial region may be material defects or contamination. All marked abnormal regions in the electrode, light-emitting, and epitaxial regions are integrated to form the defect detection results for the entire LED chip.

[0052] In this embodiment, the semiconductor chip image to be inspected is processed by extracting and combining channels in multiple color spaces. Based on the combined color channels, the chip image is segmented to obtain multiple initial regions. Statistical features are extracted from each initial region to obtain a feature set corresponding to each initial region. The statistical features include color features and spatial features. The multiple initial regions are merged according to the feature sets of each initial region to obtain multiple functional regions of the chip. For different functional regions, corresponding feature subsets are selected to identify anomalies in the functional regions, resulting in the defect detection result of the chip image. This invention avoids the problem of insufficient consideration of different regional characteristics by a unified detection strategy by identifying different functional regions of the chip, thus improving the accuracy and reliability of defect detection.

[0053] Please see Figure 2 Another embodiment of the semiconductor chip defect detection method in this application includes: 201. Extract and combine channels in multiple color spaces from the semiconductor chip image to be detected, and segment the semiconductor chip image based on the combined color channels to obtain multiple initial regions; In this embodiment, step 201 is similar to step 101 in the first embodiment, and will not be described again here.

[0054] 202. For each initial region, the corresponding color features are obtained by calculating the statistical values ​​of all pixels in the initial region on the preset color channel; In this embodiment, an LED chip is used as an example to illustrate the color feature extraction method. Each initial region obtained after segmentation contains several pixels, and these pixels each have a corresponding value in the reconstructed three-channel color representation. The goal of color feature extraction is to summarize the color attributes of the entire region using only a few values.

[0055] For red channel feature extraction, all pixels within the initial region are traversed, and their red channel values ​​are read one by one. All read pixel values ​​are summed and then divided by the total number of pixels in the region to obtain the arithmetic mean of the red channel values. This mean reflects the overall intensity level of the region in the red channel. Due to the reflective properties of the metal material, the electrode region typically has a higher red channel mean; the luminescent region has a moderate red channel mean. The difference in red channel mean between different regions constitutes an important basis for distinguishing functional areas.

[0056] The arithmetic mean of the saturation and luminance channels were calculated using the same method. The mean of the saturation channel reflects the purity of the color in the region. The normal part of the electrode area has a high mean saturation value, presenting a bright golden yellow color; the mean saturation value at pinhole defects drops significantly, and the color leans towards grayish-white. The mean of the luminance channel reflects the brightness of the region. Variations in the mean luminance value often correspond to functional area boundaries or local defects. Combining the mean values ​​of the red channel, saturation channel, and luminance channel together constitutes the basic color feature vector of this initial region.

[0057] Besides the mean, other statistics can be calculated to supplement the description of color features. For example, the variance of pixel values ​​in each channel within a region can be calculated. Variance reflects the color fluctuation within the region. A smaller variance indicates a uniform color distribution within the region, belonging to a high-quality area; a larger variance indicates color differences within the region, which may be because the region crosses the functional area boundary or contains defect edges. Another example is calculating the maximum and minimum pixel values; the difference between the two is called the range. A larger range indicates stronger color contrast within the region. These supplementary statistics can be selectively calculated according to the required detection accuracy. In routine detection, the mean feature is sufficient, while in high-precision detection, features such as variance and range can be added to enhance descriptive capabilities. The above calculation process is repeated for each initial region to obtain the color feature vectors for each initial region.

[0058] 203. For each initial region, the corresponding spatial features are obtained by calculating the normalized distance between the center coordinates of the initial region and the center coordinates of the semiconductor chip image; In this embodiment, the step of obtaining the corresponding spatial feature by calculating the normalized distance between the center coordinates of the initial region and the center coordinates of the semiconductor chip image for each initial region includes: for each initial region, determining the first center coordinates of the initial region and the second center coordinates of the semiconductor chip image; calculating the coordinate differences between the first center coordinates and the second center coordinates in the horizontal and vertical directions; normalizing the coordinate differences with the dimensions of the semiconductor chip image in the corresponding directions to obtain normalized coordinate differences; performing a power operation on the normalized coordinate differences, and determining a distance adjustment coefficient based on the positional relationship between the initial region and the chip boundary; and calculating the normalized distance using the power operation result and the distance adjustment coefficient as the spatial feature.

[0059] Specifically, the epitaxial region of an LED chip is always distributed at the chip's edge, the light-emitting region is located in a large central area, and the electrode region is located in a specific position within the light-emitting region. Color characteristics alone are insufficient to accurately distinguish between the epitaxial region and the light-emitting region, as their colors are quite similar. Spatial location characteristics can compensate for this deficiency by quantifying the initial region's position within the chip to aid in the identification of functional areas.

[0060] For a given initial region, the first step is to determine its center position. The x-coordinate of the region's center is obtained by averaging the x-coordinates of all pixels within that region; the y-coordinate of the region's center is obtained by averaging the y-coordinates of all pixels within that region. These two coordinate values ​​form the first center coordinate of the initial region. The same method is applied to the entire chip image: the x-coordinate of the image center is obtained by dividing the width of the chip image by two, and the y-coordinate of the image center is obtained by dividing the height of the chip image by two. These two coordinate values ​​form the second center coordinate of the chip image.

[0061] Calculate the difference between the first center coordinates and the second center coordinates. Horizontally, subtract the x-coordinate of the chip image center from the x-coordinate of the initial region center, and take the absolute value of the difference to obtain the horizontal coordinate difference. This difference reflects the horizontal distance of the region from the chip's centerline. Vertically, subtract the y-coordinate of the chip image center from the y-coordinate of the initial region center, and again take the absolute value to obtain the vertical coordinate difference. The vertical difference reflects the vertical distance of the region from the chip's horizontal centerline. Regions located at the chip's edge will have a larger coordinate difference in either the horizontal or vertical direction; regions located at the chip's center will have smaller coordinate differences in both directions.

[0062] Directly using pixel-level coordinate differences presents a problem because different chip images may have different sizes. A difference of the same pixel value might represent a greater distance in a small image and a closer distance in a large image. To eliminate the influence of image size, normalization is required. Dividing the horizontal coordinate difference by the width of the chip image yields the normalized horizontal coordinate difference. This normalized value represents the proportion of the area's deviation from the center line relative to the image width, ranging from 0 to 0.5. Dividing the vertical coordinate difference by the height of the chip image yields the normalized vertical coordinate difference, also ranging from 0 to 0.5. After normalization, the spatial features of images of different sizes become comparable.

[0063] While normalized coordinate differences eliminate the influence of size, their numerical changes are relatively gradual, resulting in insufficient differentiation between regions with similar locations. To enhance the distinguishability between different regions, a power operation is performed on the normalized coordinate differences. A power operation is equivalent to multiplying the normalized difference by itself a certain number of times; for example, an eighth power operation involves multiplying the value by itself eight times consecutively. The effect of the power operation is to compress smaller values ​​into smaller ones and amplify larger values ​​into larger ones, thereby widening the gap between different values. After the power operation, the value obtained by the region located in the center of the chip is close to zero, the value obtained by the region located at the edge of the chip is significantly larger, and the value obtained by the region in the middle is somewhere in between. Power operations are performed on the normalized coordinate differences in both the horizontal and vertical directions, and the larger of the two results is taken as the final result for that direction. The reason for choosing the larger value is that if a region deviates significantly from the center in either direction, it should be reflected in the spatial characteristics.

[0064] Besides the distance from the chip center, whether the initial region is adjacent to the chip boundary is also important location information. The epitaxial region is not only far from the chip center but also directly contacts the chip's outer boundary. The system detects whether the pixels in the initial region coincide with the edge pixels of the chip image; if there is overlap, the region is considered to be adjacent to the boundary. Regions adjacent to the boundary need to be given higher location weights to distinguish them from regions that are also far from the center but do not contact the boundary. A distance adjustment coefficient is set to achieve this distinction: a coefficient of two for regions adjacent to the boundary and a coefficient of one for regions not adjacent to the boundary. The value obtained by the exponentiation operation is multiplied by the corresponding distance adjustment coefficient to obtain the final normalized distance value. This value serves as the spatial feature of the initial region, accurately reflecting its location within the chip.

[0065] The epitaxial region, located at the chip edge and close to the boundary, typically has the largest spatial characteristic value among all regions. The light-emitting region, situated in a large area in the center of the chip, has a smaller spatial characteristic value. Although the electrode region is also located inside the chip, its relatively fixed position and proximity to one side result in a medium-level spatial characteristic value. By combining color and spatial characteristics, each functional region can be accurately identified.

[0066] 204. Based on the type of subsequent processing task, the color features and the spatial features are selectively combined according to a predetermined combination rule to obtain the feature set corresponding to each initial region; In this embodiment, the purpose of the region merging task is to group numerous initial regions into several major categories such as electrode regions, light-emitting regions, and epitaxial regions according to their functional attributes. The key to this task lies in distinguishing the locational distribution characteristics of different functional regions. Epitaxial regions are always close to the chip edge, while light-emitting regions occupy a large central area; although their colors are similar, their spatial positions are drastically different. Therefore, the region merging task needs to use both color features and spatial features. The average brightness channel value is selected from the color features; this feature reflects the brightness of the region, and different functional regions have certain differences in brightness. The normalized distance value is selected from the spatial features; this feature can accurately quantify the location attributes of the region. The average brightness value and the normalized distance are combined into a first feature subset, which serves as the input features for the region merging task.

[0067] For anomaly detection tasks, the goal is to identify defective regions within predefined functional areas. Different functional areas have different defect types, requiring different detection features. The main defect in the electrode area is pinholes, which are clearly visible in saturation features, with pinholes showing significantly lower saturation than normal electrode areas. Using brightness features in conjunction further confirms the presence of defects. The average saturation and brightness values ​​are combined into a second feature subset specifically for anomaly detection in the electrode area. Defects in the luminescent and epitaxial regions mainly include perforations and foreign objects, which are sensitive to red channel and brightness features. The average red channel and brightness values ​​are combined into a third feature subset for anomaly detection in the luminescent and epitaxial regions.

[0068] In practice, a complete feature set is generated for each initial region, containing all extracted color and spatial features. When performing region merging, data from the first feature subset extracted from the complete feature set is used in the calculation. When performing electrode region anomaly detection, a second feature subset is extracted from the feature set of the initial regions belonging to the electrode region. When performing luminescent or epitaxial region anomaly detection, a third feature subset is extracted. This on-demand feature selection avoids interference from irrelevant features, allowing each task to make judgments based on the most relevant information, significantly improving detection accuracy.

[0069] The predetermined combination rules are determined based on experience and experimental results during the algorithm design phase and stored as configuration parameters. During the actual detection process, the algorithm automatically searches for the corresponding feature combination method from the configuration parameters according to the current task type and the category identifier of the region to be processed, thus completing the extraction and assembly of feature subsets.

[0070] 205. Merge multiple initial regions based on the feature sets of each initial region to obtain multiple functional regions of the semiconductor chip; 206. Select corresponding feature subsets for different functional regions, perform anomaly identification on the functional regions, and obtain the defect detection results of the semiconductor chip image.

[0071] In this embodiment, In this embodiment, structured spatial data is obtained by semantic recognition and geometric analysis of residential building drawings. A wiring channel network is constructed based on the wall locations and room boundaries in the structured spatial data. The path cost from the power distribution equipment to each switch and socket device is calculated based on the wiring channel network and preset wiring constraints, and multiple candidate wiring schemes are generated based on the path cost. Rule conflict detection is performed on the multiple candidate wiring schemes, and path adjustment processing is applied to conflicting candidate wiring schemes based on the detection results, resulting in multiple optimized wiring schemes and corresponding engineering documents. This invention automatically weighs various engineering constraints through path cost calculation, generates multiple optional schemes for comparison and decision-making, and ensures that the schemes meet the specifications through automatic conflict detection and adjustment processing, significantly shortening the design cycle.

[0072] The semiconductor chip defect detection method in the embodiments of the present invention has been described above. The semiconductor chip defect detection device in the embodiments of the present invention will be described below. Please refer to [link to relevant documentation] for details. Figure 3 One embodiment of the semiconductor chip defect detection device in this invention includes: The image segmentation module 301 is used to extract and combine channels in multiple color spaces from the semiconductor chip image to be detected, and to segment the semiconductor chip image based on the combined color channels to obtain multiple initial regions. Feature extraction module 302 is used to extract statistical features from each initial region to obtain the feature set corresponding to each initial region; The region merging module 303 is used to merge multiple initial regions according to the feature set of each initial region to obtain multiple functional regions of the semiconductor chip. The defect detection module 304 is used to select corresponding feature subsets for different functional areas, perform anomaly identification on the functional areas, and obtain the defect detection results of the semiconductor chip image.

[0073] In this embodiment of the invention, the semiconductor chip defect detection device operates the aforementioned semiconductor chip defect detection method. The device extracts and combines channels from multiple color spaces in the image of the semiconductor chip to be detected. Based on the combined color channels, it segments the chip image to obtain multiple initial regions. Statistical features are extracted from each initial region to obtain a feature set corresponding to each initial region. These statistical features include color features and spatial features. The multiple initial regions are merged according to their feature sets to obtain multiple functional regions of the chip. For different functional regions, corresponding feature subsets are selected to identify anomalies and obtain the defect detection result of the chip image. This invention, by identifying different functional regions of the chip and selectively selecting detection features, avoids the problem of a unified detection strategy failing to adequately address the characteristics of different regions, thus improving the accuracy and reliability of defect detection.

[0074] above Figure 3 The semiconductor chip defect detection device in the embodiments of the present invention will be described in detail from the perspective of unitized functional entities. The semiconductor chip defect detection equipment in the embodiments of the present invention will be described in detail from the perspective of hardware processing.

[0075] Figure 4 This is a schematic diagram of a semiconductor chip defect detection device 400 provided in an embodiment of the present invention. The semiconductor chip defect detection device 400 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 410 (e.g., one or more processors) and a memory 420, and one or more storage media 430 (e.g., one or more mass storage devices) storing application programs 433 or data 432. The memory 420 and storage media 430 can be temporary or persistent storage. The program stored in the storage media 430 may include one or more units (not shown in the diagram), each unit may include a series of instruction operations on the semiconductor chip defect detection device 400. Furthermore, the processor 410 may be configured to communicate with the storage media 430 and execute the series of instruction operations in the storage media 430 on the semiconductor chip defect detection device 400 to implement the steps of the aforementioned semiconductor chip defect detection method.

[0076] The semiconductor chip defect detection device 400 may also include one or more power supplies 440, one or more wired or wireless network interfaces 450, one or more input / output interfaces 460, and / or one or more operating systems 431, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 4 The illustrated semiconductor chip defect detection device structure does not constitute a limitation on the semiconductor chip defect detection device provided by the present invention. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0077] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the semiconductor chip defect detection method.

[0078] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0079] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0080] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting defects in a semiconductor chip, characterized in that, The semiconductor chip defect detection method includes: The semiconductor chip image to be detected is subjected to multi-color space channel extraction and combination, and the semiconductor chip image is segmented based on the combined color channels to obtain multiple initial regions; Statistical features are extracted from each initial region to obtain the feature set corresponding to each initial region; Based on the feature sets of each initial region, multiple initial regions are merged to obtain multiple functional regions of the semiconductor chip; By selecting appropriate feature subsets for different functional regions, anomaly identification is performed on the functional regions to obtain the defect detection results of the semiconductor chip image.

2. The semiconductor chip defect detection method according to claim 1, characterized by, The semiconductor chip image to be detected undergoes multi-color space channel extraction and combination. Based on the combined color channels, the semiconductor chip image is segmented to obtain multiple initial regions, including: The semiconductor chip image is processed by extracting different channels in multiple color spaces and combining them across spaces to obtain a reconstructed three-channel color representation. For the three-channel color representation, multiple initial cluster centers are set on the image plane, and the comprehensive distance between the initial cluster centers and the surrounding pixels within a preset range is calculated; The pixels are clustered and assigned according to the comprehensive distance, and the pixels belonging to the same cluster center form an initial region, thus obtaining the multiple initial regions.

3. The semiconductor chip defect detection method according to claim 2, characterized in that, The process of extracting different channels from the semiconductor chip image in multiple color spaces and combining them across color spaces to obtain a reconstructed three-channel color representation includes: The semiconductor chip image is converted to RGB color space and HSV color space. The red channel is extracted from the RGB color space, and the saturation and brightness channels are extracted from the HSV color space. The red channel, the saturation channel, and the brightness channel are combined to form a reconstructed three-channel color representation.

4. The semiconductor chip defect detection method according to claim 1, characterized in that, The step of extracting statistical features from each initial region to obtain the feature set corresponding to each initial region includes: For each initial region, the corresponding color features are obtained by calculating the statistical values ​​of all pixels in the initial region on the preset color channel; For each initial region, the corresponding spatial features are obtained by calculating the normalized distance between the center coordinates of the initial region and the center coordinates of the semiconductor chip image; Depending on the type of subsequent processing task, the color features and the spatial features are selectively combined according to a predetermined combination rule to obtain the feature set corresponding to each initial region.

5. The semiconductor chip defect detection method according to claim 4, characterized in that, For each initial region, the corresponding spatial features are obtained by calculating the normalized distance between the center coordinates of the initial region and the center coordinates of the semiconductor chip image, including: For each initial region, determine the first center coordinates of the initial region and the second center coordinates of the semiconductor chip image, and calculate the coordinate differences between the first center coordinates and the second center coordinates in the horizontal and vertical directions; The coordinate difference is normalized to the size of the semiconductor chip image in the corresponding direction to obtain the normalized coordinate difference. The normalized coordinate difference is subjected to a power operation, and a distance adjustment coefficient is determined based on the positional relationship between the initial region and the chip boundary. The normalized distance is calculated using the power operation result and the distance adjustment coefficient, and is used as the spatial feature.

6. The semiconductor chip defect detection method according to claim 1, characterized in that, The step of merging multiple initial regions based on the feature sets of each initial region to obtain multiple functional regions of the semiconductor chip includes: The neighborhood relationship between each initial region is determined by the region expansion operation, and the similarity measure between adjacent initial regions is calculated based on the first feature subset in the feature set. Based on the similarity metric, adjacent initial regions are initially merged to obtain multiple initially merged regions; Based on the first feature subset, a classification operation is performed on the multiple preliminary merged regions, and the multiple preliminary merged regions are divided into multiple candidate functional regions according to the preset number of categories; The boundaries of the multiple candidate functional regions are corrected, and the multiple functional regions of the semiconductor chip are obtained based on the corrected boundaries.

7. The semiconductor chip defect detection method according to claim 1, characterized in that, The step of selecting corresponding feature subsets for different functional regions, performing anomaly identification on the functional regions, and obtaining defect detection results for the semiconductor chip image includes: Based on the category identifier of each functional area, the target feature subset corresponding to each functional area is determined from the second feature subset and the third feature subset of the feature set according to the preset feature selection rules; Input the target feature subset of the initial region within each functional region into the pre-trained anomaly detection model corresponding to the category of the functional region, and calculate the anomaly detection value; Based on the comparison between the anomaly determination value and the preset determination threshold, the initial region where the anomaly determination value exceeds the preset determination threshold is marked as an anomaly region. The anomaly regions in all functional regions are integrated to obtain the defect detection result of the semiconductor chip.

8. A semiconductor chip defect detection device, characterized in that, The semiconductor chip defect detection device includes: The image segmentation module is used to extract and combine channels in multiple color spaces from the semiconductor chip image to be detected, and to segment the semiconductor chip image based on the combined color channels to obtain multiple initial regions. The feature extraction module is used to extract statistical features from each initial region to obtain the feature set corresponding to each initial region. The region merging module is used to merge multiple initial regions according to the feature set of each initial region to obtain multiple functional regions of the semiconductor chip. The defect detection module is used to select corresponding feature subsets for different functional areas, perform anomaly identification on the functional areas, and obtain the defect detection results of the semiconductor chip image.

9. A semiconductor chip defect detection device, characterized in that, The semiconductor chip defect detection device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the semiconductor chip defect detection device to perform the steps of the semiconductor chip defect detection method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the semiconductor chip defect detection method as described in any one of claims 1-7.

Images