A crystal seed detection method and system based on image recognition

Abstract

The present application relates to the technical field of crystal seed detection, and particularly relates to a crystal seed detection method and system based on image recognition. The crystal seed detection method and system based on image recognition comprises a crystal seed image sequence acquisition module, a defect distribution extraction module, a feature enhancement module, a shadow processing module and a crystal seed defect recognition module. The present application can comprehensively represent the outer contour of the crystal seed and the characteristics of different types of defects by fixing the crystal seed on a rotating platform and combining radial, tangential and oblique light modes to obtain multi-angle image sequences, thereby ensuring that the detection covers multiple defect morphologies such as cracks, edge collapse and depressions. The high gradient area is extracted based on the edge features of the image and is uniformly projected to the edge defect distribution map, thereby realizing focused detection of the concentrated area of defects and improving the accuracy and efficiency of defect positioning.

Description

Technical Field

[0001] This invention relates to the field of seed crystal detection technology, and specifically to a seed crystal detection method and system based on image recognition. Background Technology

[0002] As a fundamental raw material for semiconductors, the quality of monocrystalline silicon directly determines the stability and reliability of subsequent silicon wafer processing and device fabrication. Seed crystals, as the starting material for monocrystalline silicon preparation, have a significant impact on the nucleation and growth process of the single crystal due to their surface condition. Defects such as cracks, chipping, and pitting at the seed crystal edges not only lead to stress concentration during crystal growth, increasing the risk of fracture, but may also extend into more severe structural damage during subsequent slicing and grinding processes, affecting yield and material properties.

[0003] Currently, the detection of seed crystal defects mainly relies on manual visual inspection or conventional image acquisition methods. Manual inspection is inefficient, highly subjective, and easily affected by the experience and fatigue of the inspectors; while conventional image acquisition often uses a single lighting condition, which makes it difficult to fully reveal the defect characteristics of the seed crystal under different lighting conditions, leading to missed detections or misjudgments in the detection results. Summary of the Invention

[0004] This invention, by fixing the seed crystal on a rotating platform and acquiring multi-angle image sequences using radial, tangential, and oblique illumination modes, comprehensively characterizes the seed crystal's outer contour and different types of defect features, ensuring detection coverage of various defect morphologies such as cracks, edge chipping, and dents. High-gradient regions are extracted based on image edge features and uniformly projected onto the edge defect distribution map, enabling focused detection of concentrated defect areas and improving the accuracy and efficiency of defect localization. Multi-scale feature enhancement significantly amplifies the difference between defects and the background, effectively suppressing false features and improving the stability and robustness of detection results. Furthermore, virtual light and shadow projection is introduced to give defects a three-dimensional effect with contrasting light and dark areas, further enhancing the model's sensitivity to defect morphology and reducing the risk of missed and false detections. Finally, enhanced defect distribution maps under different illumination conditions are fused through multiple channels and input into a YOLOv7-based recognition model for classification and localization, thereby achieving intelligent and refined detection of seed crystal defects, significantly improving detection accuracy, real-time performance, and engineering application value.

[0005] This invention provides a seed crystal detection method based on image recognition, comprising:

[0006] Angle-stepping imaging operations are performed on the seed crystals under different lighting conditions to obtain a seed crystal image sequence. The seed crystal image sequence includes several seed crystal image sets, and each seed crystal image set includes seed crystal images from all angles under the same lighting conditions.

[0007] For each seed image set, the following steps are performed: high gradient edge regions are determined based on the edge features of the seed images, and then all high gradient edge regions corresponding to the seed images are projected onto the same edge defect distribution map using polar coordinates;

[0008] A feature enhancement operation is performed on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map;

[0009] Virtual light and shadow projection processing is performed on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map;

[0010] All shadow-processed edge defect distribution maps are stitched together according to channels to form a seed analysis map. The seed analysis map is then fed into the seed defect recognition model for processing, and a seed detection label is output.

[0011] As a preferred aspect, high-gradient edge regions are determined based on the edge features of the seed images, and then all high-gradient edge regions corresponding to the seed images are projected onto the same edge defect distribution map using polar coordinates. This specifically includes the following steps:

[0012] Canny edge detection is performed on the seed image to extract the outer boundary contour. Then, least squares circular arc fitting is applied to the outer boundary contour to obtain the outer contour curve. Based on the outer contour curve, the inner distance r is extended to obtain the ring zone region.

[0013] Within the annular region, the gradient values ​​corresponding to all pixels are calculated, and then gradient histogram abrupt change point detection is performed on the gradient values ​​corresponding to all pixels to determine the dynamic threshold T.

[0014] Within the annular region, pixels with gradient values ​​higher than the dynamic threshold T are retained, while pixels with gradient values ​​lower than the dynamic threshold T are removed, resulting in a high-gradient edge region.

[0015] Each high-gradient edge region is then mapped to a two-dimensional image in polar coordinates to obtain a local edge defect distribution map. All local edge defect distribution maps are then stitched together end to end in descending order of the seed rotation angle to obtain the edge defect distribution map.

[0016] As a preferred aspect, a feature enhancement operation is performed on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map, specifically including the following steps:

[0017] Wavelet transform is performed on the edge defect distribution map to divide it into high-frequency sub-band and low-frequency sub-band. High-pass filtering and Laplacian enhancement are then performed on the high-frequency sub-band to obtain the first enhanced feature map.

[0018] For the i-th pixel in the first enhanced feature map, calculate the mean square error between the corresponding data of the i-th pixel in all first enhanced feature maps to obtain the gray-level consistency coefficient. Determine whether the gray-level consistency coefficient is higher than the consistency threshold. If the gray-level consistency coefficient is higher than the consistency threshold, retain the i-th pixel in all first enhanced feature maps. If the gray-level consistency coefficient is not higher than the consistency threshold, delete the i-th pixel in all first enhanced feature maps.

[0019] As a preferred aspect, virtual light and shadow projection processing is performed on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map, specifically including the following steps:

[0020] Set up a virtual light source, which is generally selected by radial lighting. Determine the light source direction vector corresponding to the virtual light source, calculate the normal vector of each pixel in the enhanced edge defect distribution map, and then calculate the dot product of the normal vector of each pixel and the light source direction vector and the illumination intensity coefficient, which is recorded as the shadow intensity value.

[0021] The shadow intensity values ​​corresponding to all pixels are combined to form a shadow layer, and the enhanced edge defect distribution map and the shadow layer are added together to obtain the shadow processing edge defect distribution map.

[0022] As a preferred approach, training the seed crystal defect identification model includes the following steps:

[0023] Several seed defect recognition training samples are obtained, including a distribution map of edge defects with shadow processing. The seed defect recognition training samples are labeled with seed detection tags. All labeled seed defect recognition training samples are then combined into a seed defect recognition training set. The seed defect recognition model is trained using the seed defect recognition training set, with the labeled seed detection tags used as the training target during training.

[0024] As a preferred aspect, the seed defect identification model is based on the YOLOv7 model.

[0025] This invention provides a seed crystal detection and system based on image recognition, comprising:

[0026] The seed image sequence acquisition module is used to perform angle step-by-angle shooting operations on the seed crystals under different lighting conditions to obtain a seed image sequence. The seed image sequence includes several seed image sets, and each seed image set includes seed images from all angles under the same lighting conditions.

[0027] The defect distribution extraction module is used to perform the following steps for each seed image set: determine the high gradient edge region based on the edge features of the seed image, and then project the high gradient edge regions corresponding to all seed images onto the same edge defect distribution map in polar coordinates.

[0028] The feature enhancement module is used to perform feature enhancement operations on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map;

[0029] The shadow processing module is used to perform virtual light and shadow projection processing on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map.

[0030] The seed defect identification module is used to stitch together all the shadow-processed edge defect distribution maps according to channels to form a seed analysis map. The seed analysis map is then sent to the seed defect identification model for processing and outputs a seed detection label.

[0031] The present invention has the following advantages:

[0032] This invention, by fixing the seed crystal on a rotating platform and acquiring multi-angle image sequences using radial, tangential, and oblique illumination modes, comprehensively characterizes the seed crystal's outer contour and different types of defect features, ensuring detection coverage of various defect morphologies such as cracks, edge chipping, and dents. High-gradient regions are extracted based on image edge features and uniformly projected onto the edge defect distribution map, enabling focused detection of concentrated defect areas and improving the accuracy and efficiency of defect localization. Multi-scale feature enhancement significantly amplifies the difference between defects and the background, effectively suppressing false features and improving the stability and robustness of detection results. Furthermore, virtual light and shadow projection is introduced to give defects a three-dimensional effect with contrasting light and dark areas, further enhancing the model's sensitivity to defect morphology and reducing the risk of missed and false detections. Finally, enhanced defect distribution maps under different illumination conditions are fused through multiple channels and input into a YOLOv7-based recognition model for classification and localization, thereby achieving intelligent and refined detection of seed crystal defects, significantly improving detection accuracy, real-time performance, and engineering application value. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the image recognition-based seed detection system used in an embodiment of the present invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0035] Example 1: A seed crystal detection method based on image recognition, comprising:

[0036] The seed crystal was fixed on a rotating stage, and an adjustable ring light source was used. An angle-stepping imaging operation was performed on the seed crystal under different lighting conditions to obtain a sequence of seed crystal images. It should be noted that the seed crystal here refers to a small single-crystal silicon rod used for single-crystal silicon preparation. Different lighting conditions included radial illumination, tangential illumination, and oblique illumination. Radial illumination refers to light shining radially from the adjustable ring light source towards the center of the seed crystal, which can highlight the overall outline of the seed crystal and facilitate the extraction of the seed crystal edges. Tangential illumination involves light incident at an angle of 0-5° along the edge of the seed crystal. Because tangential illumination produces strong light and shadow differences at the edges, it can enhance the appearance of abrupt edge structures such as cracks and chipping. The illumination is directed at an angle of 20-45° along the edge of the seed crystal, which can enhance the three-dimensionality of pitting and other concave defects. These illumination conditions are all set using a ring-shaped adjustable light source. The specific shooting steps are as follows: the seed crystal fixed on the rotating stage is rotated clockwise by 5° once, and then images are taken under radial illumination, tangential illumination, and oblique illumination conditions respectively. The seed crystal is then rotated again, and so on, until the seed crystal has been rotated one full circle. All the captured images are combined into a seed crystal image sequence. The seed crystal image sequence includes several seed crystal image sets, and each seed crystal image set includes seed crystal images from all angles under the same illumination conditions.

[0037] For each seed image set, the following steps are performed: high gradient edge regions are determined based on the edge features of the seed images, and then all high gradient edge regions corresponding to the seed images are projected onto the same edge defect distribution map using polar coordinates. It should be noted that the edge features here generally refer to the pixel features within the outer contour curve of the seed, while the edge defect distribution map can present the overall defect distribution of the seed, enabling more comprehensive detection of seed defects.

[0038] It should be added that the main body of the seed crystal has a complete structure and uniform stress, and defects should not occur during the seed crystal preparation process. However, the edge region is the main stress-bearing area, and the defects of the seed crystal are mainly concentrated at the edge position. The high gradient edge region is a defect concentration area determined based on the edge characteristics. Detection of these defect concentration areas can improve the accuracy of seed crystal defect detection.

[0039] A feature enhancement operation is performed on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map. The feature enhancement operation can significantly amplify the difference between the defect and the background and effectively remove pseudo features under single illumination conditions.

[0040] Virtual lighting and shadow projection processing is applied to the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map. Applying virtual lighting and shadow projection processing can make defects such as cracks, chipped edges, and pits exhibit a three-dimensional effect of "light and dark contrast". This not only improves the model's sensitivity to defect morphology and reduces the probability of it being confused with background noise, but also makes the defects more visually clear and prominent.

[0041] All shadow-processed edge defect distribution maps are stitched together according to channels to form a seed analysis map. The seed analysis map is then sent to the seed defect recognition model for processing and seed detection labels are output. The seed detection labels include normal and abnormal. When an abnormal label is output, the defect type and defect location are also output. The seed defect recognition model is based on the YOLOv7 model.

[0042] This application acquires multi-angle image sequences by fixing the seed crystal on a rotating platform and combining radial, tangential, and oblique illumination modes. This comprehensively characterizes the outer contour of the seed crystal and the features of different types of defects, ensuring that the detection covers various defect morphologies such as cracks, edge chipping, and depressions. High-gradient regions are extracted based on image edge features and uniformly projected onto the edge defect distribution map, enabling focused detection of concentrated defect areas and improving the accuracy and efficiency of defect localization. Multi-scale feature enhancement processing significantly amplifies the difference between defects and the background, effectively suppressing false features and improving the stability and robustness of the detection results. Furthermore, virtual light and shadow projection is introduced to give defects a three-dimensional effect of light and dark contrast, further enhancing the model's sensitivity to defect morphology and reducing the risk of missed and false detections. Finally, the enhanced defect distribution maps under different illumination conditions are fused in multiple channels and input into a YOLOv7-based recognition model for classification and localization, thereby achieving intelligent and refined detection of seed crystal defects, significantly improving detection accuracy, real-time performance, and engineering application value.

[0043] High-gradient edge regions are identified based on the edge features of the seed crystal images. Then, all high-gradient edge regions corresponding to the seed crystal images are projected onto the same edge defect distribution map using polar coordinates. The specific steps include the following:

[0044] Canny edge detection is performed on the seed image to extract the outer boundary contour. Then, least squares circular arc fitting is applied to the outer boundary contour to obtain the outer contour curve. Based on the outer contour curve, the inward extension is made by a distance r to obtain the ring zone region. The value of r is determined by the operator according to the size of the seed, which is generally 1-3% of the seed diameter. The ring zone region represents the edge region of the seed, where the probability of defects is relatively high.

[0045] Within the annular region, the gradient values ​​corresponding to all pixels are calculated using the Sobel operator. Then, gradient histogram abrupt change point detection is performed on the gradient values ​​corresponding to all pixels to determine the dynamic threshold T. Specifically, gradient histogram abrupt change point detection involves plotting a gradient histogram of the gradient values ​​of all pixels, then taking the second difference of the histogram curve corresponding to the gradient histogram to determine the peak value as the dynamic threshold T.

[0046] Within the annular region, pixels with gradient values ​​higher than the dynamic threshold T are retained, while pixels with gradient values ​​lower than the dynamic threshold T are removed, resulting in a high-gradient edge region. It should be noted that when defects appear at the edge of the seed crystal, a local grayscale abrupt change occurs at the edge, i.e., a high gradient. Pixels corresponding to the high gradient can more significantly highlight the defects.

[0047] Each high-gradient edge region is then mapped to a two-dimensional image in polar coordinates to obtain a local edge defect distribution map. All local edge defect distribution maps are then stitched together end to end in descending order of the seed rotation angle to obtain the edge defect distribution map. It should be noted that the edge defect distribution map actually flattens out the entire defect region of the seed, which facilitates subsequent seed defect detection.

[0048] The process of mapping a two-dimensional image to polar coordinates involves fitting the outer contour curve corresponding to the high gradient edge region to a circle or ellipse, determining the center and radius, and then converting each pixel within the high gradient edge region to polar coordinates based on the center and radius.

[0049] A feature enhancement operation is performed on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map. The specific steps include the following:

[0050] Wavelet transform is performed on the edge defect distribution map to divide it into a high-frequency sub-band and a low-frequency sub-band. High-pass filtering and Laplace enhancement are then performed on the high-frequency sub-band to obtain the first enhanced feature map. It should be noted that the high-frequency sub-band contains information about seed defects, and the difference between defects and the background can be significantly amplified by high-pass filtering and Laplace enhancement.

[0051] For the i-th pixel in the first enhanced feature map, calculate the mean square error among the corresponding data of the i-th pixel in all first enhanced feature maps to obtain the grayscale consistency coefficient. Determine whether the grayscale consistency coefficient is higher than the consistency threshold. The consistency threshold is set by the operator. If the grayscale consistency coefficient is higher than the consistency threshold, it means that the pixel performs stably under different lighting conditions and should be a true feature. In this case, the i-th pixel is retained in all first enhanced feature maps. If the grayscale consistency coefficient is not higher than the consistency threshold, it means that the feature information is greatly affected by lighting and is unstable. It can be regarded as a false feature and the i-th pixel is deleted in all first enhanced feature maps.

[0052] Next, virtual lighting and shadow projection processing is performed on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map. The specific steps include the following:

[0053] Set up a virtual light source, typically using radial illumination, and determine the light source direction vector. Calculate the normal vector of each pixel in the enhanced edge defect distribution map. Then, calculate the dot product of the normal vector and the light source direction vector of each pixel, multiplied by the illumination intensity coefficient, and record this as the shadow intensity value. The shadow intensity value reflects the brightness of the corresponding pixel under virtual illumination conditions, thus making the features corresponding to the defects more obvious. The illumination intensity coefficient is set by the operator.

[0054] The shadow intensity values ​​corresponding to all pixels are combined to form a shadow layer. The enhanced edge defect distribution map and the shadow layer are then added together to obtain the shadow processing edge defect distribution map. In practice, a shadow depth coefficient is also set for the shadow layer to control the size of the shadow layer when it is added to the enhanced edge defect distribution map.

[0055] Training the seed crystal defect identification model involves the following steps:

[0056] Several seed defect recognition training samples are obtained. These samples include a shadowed edge defect distribution map, which is constructed by the operator based on actual seed detection experiments. The seed defect recognition training samples are then labeled with seed detection tags, typically manually indicating the defect type and location. All labeled seed defect recognition training samples are then combined into a seed defect recognition training set. The seed defect recognition model is trained using this training set. During training, the labeled seed detection tags are used as the training target to determine if the training conditions are met. The training conditions generally require the accuracy of the seed defect recognition model to meet expectations. If the training conditions are met, the trained seed defect recognition model is output; otherwise, the seed defect recognition model is trained again using the seed defect recognition training set.

[0057] Example 2: A seed crystal detection and system based on image recognition, see [link to example]. Figure 1 ,include:

[0058] The seed crystal image sequence acquisition module is used to fix the seed crystal on a rotating stage and, using a ring-shaped adjustable light source, performs angle-stepping imaging operations on the seed crystal under different lighting conditions to obtain a seed crystal image sequence. It should be noted that the seed crystal here refers to a small single-crystal silicon rod used for single-crystal silicon preparation. Different lighting conditions include radial illumination, tangential illumination, and oblique illumination. Radial illumination refers to light shining radially from the ring-shaped adjustable light source towards the center of the seed crystal, which can highlight the overall outline of the seed crystal and facilitate the extraction of the seed crystal edges. Tangential illumination involves light incident at an angle of 0-5° along the edge of the seed crystal. Because tangential illumination produces a strong light and shadow difference at the edge, it can enhance the appearance of cracks and chipped edges. The edge abrupt structure, with oblique lighting (light incident at an angle of 20-45° along the edge of the seed crystal), can enhance the three-dimensionality of pitting and concave defects. These lighting conditions are all set using a ring-shaped adjustable light source. The specific imaging steps are as follows: the seed crystal fixed on the rotating stage is rotated clockwise by 5° once, and then images are taken under radial lighting, tangential lighting, and oblique lighting conditions respectively. The seed crystal is then rotated again, and so on, until the seed crystal has been rotated one full circle. All the captured images are combined into a seed crystal image sequence. The seed crystal image sequence includes several seed crystal image sets, and each seed crystal image set includes seed crystal images from all angles under the same lighting conditions.

[0059] The defect distribution extraction module performs the following steps for each seed image set: determining high-gradient edge regions based on the edge features of the seed images, and then projecting the high-gradient edge regions corresponding to all seed images onto the same edge defect distribution map using polar coordinates; it should be noted that the edge features here generally refer to the pixel features within the outer contour curve of the seed, while the edge defect distribution map can present the overall defect distribution of the seed, enabling more comprehensive detection of seed defects;

[0060] The feature enhancement module is used to perform feature enhancement operations on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map. The feature enhancement operation can significantly amplify the difference between the defect and the background and can effectively remove false features under single illumination conditions.

[0061] The shadow processing module is used to perform virtual light and shadow projection processing on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map. Performing virtual light and shadow projection processing can make defects such as cracks, chipped edges, and pits exhibit a three-dimensional effect of "light and dark contrast". This not only improves the model's sensitivity to defect morphology and reduces the probability of it being confused with background noise, but also makes the defects more visually clear and prominent.

[0062] The seed defect identification module is used to stitch together all the shadow-processed edge defect distribution maps according to channels to form a seed analysis map. The seed analysis map is then sent to the seed defect identification model for processing and outputs seed detection labels. The seed detection labels include normal and abnormal, and when an abnormal label is output, the defect type and defect location will also be output.

[0063] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A seed crystal detection method based on image recognition, characterized in that, include: Angle-stepping imaging operations are performed on the seed crystals under different lighting conditions to obtain a seed crystal image sequence. The seed crystal image sequence includes several seed crystal image sets, and each seed crystal image set includes seed crystal images from all angles under the same lighting conditions. For each seed image set, the following steps are performed: high gradient edge regions are determined based on the edge features of the seed images, and then all high gradient edge regions corresponding to the seed images are projected onto the same edge defect distribution map using polar coordinates; A feature enhancement operation is performed on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map; Virtual light and shadow projection processing is performed on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map; All shadow-processed edge defect distribution maps are stitched together according to channels to form a seed analysis map. The seed analysis map is then fed into the seed defect recognition model for processing, and a seed detection label is output.

2. The seed crystal detection method based on image recognition according to claim 1, characterized in that, High-gradient edge regions are identified based on the edge features of the seed crystal images. Then, all high-gradient edge regions corresponding to the seed crystal images are projected onto the same edge defect distribution map using polar coordinates. The specific steps include the following: Canny edge detection is performed on the seed image to extract the outer boundary contour. Then, least squares circular arc fitting is applied to the outer boundary contour to obtain the outer contour curve. Based on the outer contour curve, the inner distance r is extended to obtain the ring zone region. Within the annular region, the gradient values ​​corresponding to all pixels are calculated, and then gradient histogram abrupt change point detection is performed on the gradient values ​​corresponding to all pixels to determine the dynamic threshold T. Within the annular region, pixels with gradient values ​​higher than the dynamic threshold T are retained, while pixels with gradient values ​​lower than the dynamic threshold T are removed, resulting in a high-gradient edge region. Each high-gradient edge region is then mapped to a two-dimensional image in polar coordinates to obtain a local edge defect distribution map. All local edge defect distribution maps are then stitched together end to end in descending order of the seed rotation angle to obtain the edge defect distribution map.

3. The seed crystal detection method based on image recognition according to claim 2, characterized in that, A feature enhancement operation is performed on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map. The specific steps include the following: Wavelet transform is performed on the edge defect distribution map to divide it into high-frequency sub-band and low-frequency sub-band. High-pass filtering and Laplacian enhancement are then performed on the high-frequency sub-band to obtain the first enhanced feature map. For the i-th pixel in the first enhanced feature map, calculate the mean square error between the corresponding data of the i-th pixel in all first enhanced feature maps to obtain the gray-level consistency coefficient. Determine whether the gray-level consistency coefficient is higher than the consistency threshold. If the gray-level consistency coefficient is higher than the consistency threshold, retain the i-th pixel in all first enhanced feature maps. If the gray-level consistency coefficient is not higher than the consistency threshold, delete the i-th pixel in all first enhanced feature maps.

4. The seed crystal detection method based on image recognition according to claim 3, characterized in that, Perform virtual lighting and shadow projection processing on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map. The specific steps include the following: Set up a virtual light source and determine the light source direction vector corresponding to the virtual light source. Calculate the normal vector of each pixel in the enhanced edge defect distribution map. Then calculate the dot product of the normal vector of each pixel and the light source direction vector and the illumination intensity coefficient, and record it as the shadow intensity value. The shadow intensity values ​​corresponding to all pixels are combined to form a shadow layer, and the enhanced edge defect distribution map and the shadow layer are added together to obtain the shadow processing edge defect distribution map.

5. The seed crystal detection method based on image recognition according to claim 4, characterized in that, Training the seed crystal defect identification model involves the following steps: Several seed defect recognition training samples are obtained, including a distribution map of edge defects with shadow processing. The seed defect recognition training samples are labeled with seed detection tags. All labeled seed defect recognition training samples are then combined into a seed defect recognition training set. The seed defect recognition model is trained using the seed defect recognition training set, with the labeled seed detection tags used as the training target during training.

6. The seed crystal detection method based on image recognition according to claim 5, characterized in that, The seed defect identification model is based on the YOLOv7 model.

7. A seed crystal detection system based on image recognition, characterized in that, The system employs a seed detection method based on image recognition as described in any one of claims 1-6, comprising: The seed image sequence acquisition module is used to perform angle step-by-angle shooting operations on the seed crystals under different lighting conditions to obtain a seed image sequence. The seed image sequence includes several seed image sets, and each seed image set includes seed images from all angles under the same lighting conditions. The defect distribution extraction module is used to perform the following steps for each seed image set: determine the high gradient edge region based on the edge features of the seed image, and then project the high gradient edge regions corresponding to all seed images onto the same edge defect distribution map in polar coordinates. The feature enhancement module is used to perform feature enhancement operations on the edge defect distribution map corresponding to the seed image set to obtain an enhanced edge defect distribution map; The shadow processing module is used to perform virtual light and shadow projection processing on the enhanced edge defect distribution map to obtain a shadow-processed edge defect distribution map. The seed defect identification module is used to stitch together all the shadow-processed edge defect distribution maps according to channels to form a seed analysis map. The seed analysis map is then sent to the seed defect identification model for processing and outputs a seed detection label.

Images