Improved YOLOv10-based solder ball defect detection method for BGA (Ball Grid Array) packaged chip

Abstract

The invention provides an improved YOLOv10-based solder ball defect detection method for a BGA packaged chip, and the method comprises the steps: collecting a solder ball image of the BGA packaged chip, carrying out the morphological preprocessing, and carrying out the analysis and classification of defects; carrying out defect generation on solder ball burr defects with less data; performing similarity measurement on the generated defects and the real defects, and selecting defect data with relatively high similarity to construct a pre-training data set; a YOLOv10 deep learning network is improved, a C2f module in the network is improved, an up-sampling module is replaced with a lightweight dynamic up-sampling operator DySample, an original loss function is replaced with a Focal-EIoU loss function, the improved network is used for pre-training, and a defect detection model is obtained; performing real-time defect detection on the solder ball of the BGA packaging chip by using the defect detection model to obtain defect information; according to the invention, the detection precision and speed of the solder ball defect of the BGA packaging chip are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and particularly relates to a method for detecting solder ball defects of BGA packaged chips based on improved YOLOv10 Background Art

[0002] In recent years, the detection of BGA packaged chips has increasingly become an important link in the semiconductor industry. Among them, the detection of solder ball defects is particularly crucial, and this work is of great significance for ensuring the stability of chip quality, reducing mass production costs, and improving production efficiency. At present, the detection of solder ball defects of BGA packaged chips is mainly completed manually, which not only restricts production efficiency but also cannot guarantee the accuracy of inspection. With the wide application of object detection methods based on deep learning in the field of defect detection, YOLOv10 has become a widely used and relatively advanced object detection model

[0003] However, for the detection of solder ball defects of BGA packaged chips, due to the extremely small size of the solder balls, YOLOv10 faces certain difficulties in feature extraction and processing, and there is still room for improvement in the accuracy of defect detection and classification. How to improve the detection ability of the deep learning network for small targets and achieve a balance between detection speed and accuracy has become an urgent problem to be solved. At the same time, in actual industrial production, especially in the manufacturing process of solder balls of BGA packaged chips, due to strict requirements for production quality, defective samples rarely appear, so it is extremely difficult and costly to obtain defective samples. How to reasonably construct a high-quality dataset sufficient to support model training is also an important problem Summary of the Invention

[0004] 1. Technical problems to be solved

[0005] The existing collected dataset is not sufficient to construct a high-quality dataset sufficient to support model training. At the same time, the method for detecting solder ball defects of BGA packaged chips cannot guarantee the accuracy of inspection and is inefficient, and cannot meet the actual detection needs

[0006] 2. Technical solutions

[0007] To solve the above problems, the present invention provides a method for detecting solder ball defects of BGA packaged chips based on improved YOLOv10, which is characterized by including the following steps

[0008] Step S01: Collect images of solder balls of BGA packaged chips, perform morphological preprocessing on the images, and analyze and classify the defects

[0009] Step S02: Generate defects for the solder ball burr defects with less data

[0010] Step S03: Perform feature similarity measurement on the generated defect data and the real defect data, and select the generated defects and real defects with higher similarity to construct a dataset together.

[0011] Step S04: Improve the YOLOv10 deep learning network and perform pre-training to obtain a defect detection model.

[0012] Step S05: Use the pre-trained defect detection model to perform real-time defect detection on the solder balls of the BGA package chip and obtain defect information.

[0013] Further, the specific method of step S01 includes the following steps:

[0014] Step S11: Convert the solder ball image of the BGA package chip collected by the camera into a grayscale image, and then convert the grayscale image into a binary image;

[0015] Step S12: Select the shape of the structural element according to the shapes of the objects such as noise and dirt existing in the collected image, and adjust the size of the structural element to make it close to the size of the target.

[0016] Step S13: Use the opening operation, first perform morphological erosion, and then perform morphological dilation to remove small dirt or noise points in the image.

[0017] Step S14: Use the closing operation, first perform morphological dilation, and then perform morphological erosion to fill the small holes in the foreground while maintaining the shape of the foreground area, as follows.

[0018] Step S14: Combine the binary image after morphological processing with the original image to generate a final image with noise and dirt removed.

[0019] Step S15: Analyze the defects in the collected image and classify them, which are solder ball bridging, solder ball deformation, solder ball missing, and solder ball burr in sequence.

[0020] Further, the specific method of step S02 includes the following steps:

[0021] Step S21: Input a group of defect-free images, randomly locate a solder ball in the image, and locate the center of the solder ball.

[0022] Step S22: Obtain the color channels of the central region of the target box, calculate the maximum value C max and the average value C avg , and perform weighted extraction of the color from the maximum value to the average value through Gaussian weighting to obtain the weighted value C smooth (x), as follows:

[0023]

[0024] Wherein, C i is the color value of the i-th pixel, n is the number of pixels in the region, x is the distance from the center point of the solder ball, and σ is the width controlling the transition.

[0025] Step S23: Randomly generate a line with specified restrictions on the image. The starting point (x start , y start ) of the line is the center of the solder ball, and the end point of the line is (x end , y end ), and it is filled with the extracted color, specifically as follows:

[0026] x end = x start + L·cos(θ)

[0027] y end = y start + L·sin(θ)

[0028] Wherein, L is an arbitrary line length.

[0029] Step S24: Perform mosaic processing and Gaussian smoothing transition on the generated line, specifically as follows:

[0030]

[0031] Wherein, (i, j) are the coordinates of the Gaussian kernel, G(i, j) is the weight of the Gaussian kernel, I(x + i, y + j) is the pixel value of the neighborhood, σ is the smoothness degree, k is the radius of the kernel, and I output (x, y) is the value of the processed image at the pixel position (x, y).

[0032] Furthermore, the specific method of the said Step S03 is:

[0033] Step S31: Set real defect samples as target domain samples Generate defect samples as source domain samples

[0034] Wherein, t and s represent the target domain and the source domain, i and j represent the sample numbers of the target domain and the source domain, represent the samples of the target domain and the source domain, and n and m represent the total numbers of the target domain samples and the source domain samples

[0035] Step S32: Input the target domain and source domain samples into the feature extraction network, and use the feature vectors of the intermediate layer of the network to represent the sample features, specifically as follows:

[0036]

[0037]

[0038] In the formula, are the feature representations of the target domain and source domain samples respectively, and F(·) represents the feature extraction network.

[0039] Step S33: Calculate the similarity between samples using cosine similarity And for each source domain sample, calculate its similarity score S with all target domain samples j , specifically as follows:

[0040]

[0041] Step S34: Set the similarity threshold to τ, and retain the generated samples that satisfy S j ≥τ. Mix the generated samples and the real samples, and after annotation, divide the dataset into a training set and a test set according to a ratio of 8:2.

[0042] Furthermore, the specific method of the said step S04 is:

[0043] Step S41: Design a cross-scale efficient feature fusion network to replace the feature fusion network of the original network.

[0044] Step S42: Improve the C2f modules in the backbone network and the neck network respectively.

[0045] Step S43: In the neck network, use the lightweight dynamic upsampling operator DySample to replace the original upsampling module UpSample.

[0046] Step S44: Replace the original loss function CIoU with the Focal-EIoU loss function.

[0047] Step S45: Use the labeled dataset to perform pre-training in the improved YOLOv10 deep learning network, and take the result with the best training effect as the defect detection model;

[0048] Furthermore, the specific method of the said step S41 is:

[0049] Feature outputs are added after the 2nd, 4th, and 6th layers of the original network respectively. The feature output of the 2nd layer goes to the 15th layer, the feature output of the 4th layer goes to the 12th layer, and the feature output of the 6th layer goes to the 21st layer. The Concat module of the 12th layer in the improved network performs deep feature fusion on the outputs of the 4th, 6th, and 11th layers of the network. The Concat module of the 15th layer performs deep feature fusion on the outputs of the 2nd, 4th, and 14th layers of the network. The Concat module of the 18th layer performs deep feature fusion on the 13th and 17th layers of the network. The Concat module of the 21st layer performs deep feature fusion on the 6th, 10th, and 20th layers of the network.

[0050] Furthermore, the specific method of step S42 is as follows:

[0051] In the backbone network, replace Bottleneck with DA_Bottleneck, and improve the C2f module of the 8th layer to C2f_DA module. In the neck network, replace all C2f modules of the 13th, 16th, and 19th layers with C2fCIB modules, and keep the C2fCIB module of the 22nd layer unchanged.

[0052] Furthermore, the specific method of step S43 is as follows:

[0053] Replace the upsampling modules UpSample of the 11th and 14th layers in the improved network with the lightweight dynamic upsampling operator DySample.

[0054] Furthermore, the Focal-EIoU loss function replacing the original loss function in step S44 is specifically as follows:

[0055]

[0056] L Focal-EIOU = IOU γ L EIOU

[0057] In the formula, L EIOU is the EIOU loss function, L IOU is the intersection over union loss function, L dis is the distance loss function, L asp is the aspect ratio loss function. IOU is the intersection over union, b is the center point coordinate of the predicted bounding box, b gt is the center point coordinate of the ground truth bounding box, w is the width of the predicted bounding box, w gt is the width of the ground truth bounding box, h is the height of the predicted bounding box, h gt is the height of the ground truth bounding box, w c is the width of the smallest bounding box enclosing the predicted and ground truth bounding boxes, h cis the height of the smallest circumscribed rectangle enclosing the predicted bounding box and the ground truth bounding box, ρ is the Euclidean distance between two points, and γ is the focal adjustment factor.

[0058] Furthermore, the specific method of step S05 is as follows:

[0059] Input a set of real-time acquired BGA package chip solder ball images into the defect detection model configured with pre-trained weights, detect the BGA package chip solder ball images, and obtain the positions of the BGA package chips with solder ball defects and the types of BGA package chip solder ball defects.

[0060] 3. Beneficial effects:

[0061] After data generation, the BGA package chip solder ball defect data of the present invention meets the training requirements of the defect detection model, and at the same time, the detection accuracy of the BGA package chip solder ball defect detection method based on the improved YOLOv10 is higher; Description of the drawings

[0062] Figure 1 is the flowchart of the BGA package chip solder ball defect detection based on the improved YOLOv10 of the present invention.

[0063] Figure 2 is the flowchart of the image preprocessing of the present invention.

[0064] Figure 3 is the flowchart of the solder ball burr defect generation of the present invention.

[0065] Figure 4 is the flowchart of the dataset construction based on similarity measurement of the present invention.

[0066] Figure 5 is the network structure diagram of the improved YOLOv10 deep learning network of the present invention.

[0067] Figure 6 is the network structure diagram of the C2f_DA module of the present invention. Detailed implementation manners

[0068] The present invention will be described in detail below with reference to the drawings and embodiments.

[0069] As Figure 1As shown in the figure, a method for detecting solder ball defects of BGA packaged chips based on improved YOLOv10 includes step S01: collecting solder ball images of BGA packaged chips, performing morphological preprocessing on the images, and then performing data analysis and classification on the defects; step S02: generating defects for solder ball burr defects with less data; step S03: performing similarity measurement on the generated defects and real defects, and selecting the generated defects and real defects with higher similarity to construct a dataset together; step S04: improving the YOLOv10 deep learning model and performing pre-training to obtain a defect detection model; step S05: using the pre-trained defect detection model to perform real-time defect detection on the solder balls of BGA packaged chips to obtain defect information;

[0070] Step S01: As Figure 2 shown in the figure, collect solder ball images of BGA packaged chips, perform morphological preprocessing on the images, and classify the defects.

[0071] Step S01 includes the following steps: Step S11, use a high-definition industrial camera to collect solder ball images of BGA packaged chips, convert the collected images into grayscale images, and then convert the grayscale images into binary images.

[0072] Step S12, according to the shapes and sizes of objects such as noise and dirt existing in the collected images, select different structural element shapes and adjust the sizes of the structural elements to be close to the sizes of the targets.

[0073] Step S13, in one embodiment, the method for using opening operation to remove dirt or noise in the image is: use opening operation to perform morphological erosion to eliminate small connected domains of objects such as noise and dirt in the image, and then compensate for the area reduction caused by erosion in larger connected domains through dilation operation.

[0074] Step S14, in one embodiment, the method for using closing operation to fill small holes in the foreground and keep the shape of the foreground area is: use closing operation, first perform morphological dilation, then perform morphological erosion to fill small holes in the connected domain, expand the boundary of the connected domain, connect two adjacent connected domains, and then reduce the expansion of the connected domain boundary and the increase in area caused by the dilation operation through erosion operation.

[0075] Step S15, combine the binary image after morphological processing with the original image to generate a final image with noise and dirt removed.

[0076] Step S16, perform data analysis on the defects in the collected images and then classify them, which are solder ball bridging, solder ball abnormal shape, solder ball missing, and solder ball burr in sequence.

[0077] Step S02: As Figure 3As shown, defect generation is performed on the burr defects of solder balls with less data: classify the defects in the collected images, analyze the number and distribution of defects in the data, and find that the data of burr defects of solder balls is too small to support the training of the network. Imitating the characteristics of burr defects of solder balls, defect generation is performed on the data without defects. It mainly includes the following steps;

[0078] Step S02 includes the following steps: Step S21, input a group of defect-free images, randomly locate a solder ball in the image, and locate the center of the solder ball.

[0079] Step S22, in one embodiment, the method for obtaining the weighted value C smooth (x) of the color channel of the solder ball center is: obtain the color channels in the central region of the target box, calculate the maximum value C max and the average value C avg of the color channels, and perform weighted extraction of the color from the maximum value to the average value through Gaussian weighting to obtain the weighted value C smooth (x), specifically as follows:

[0080]

[0081] In the formula, C i is the color value of the i-th pixel, n is the number of pixels in this region, x is the distance from the center point of the solder ball, and σ is the width controlling the transition.

[0082] Step S23, in one embodiment, the specific method for generating burr defect features is: randomly generate a straight line with specified restrictions on the image, and the starting point (x start , y start ) of the straight line is the center of the solder ball, and the end point of the straight line is (x end , y end ), and fill it with the extracted color, specifically as follows:

[0083] x end =x start +L·cos(θ)

[0084] y end =y start +L·sin(θ)

[0085] In the formula, L is an arbitrary line length within the specified restriction range.

[0086] Step S24: In one embodiment, the method for image fusion transition is: perform mosaic processing and Gaussian smoothing transition on the edges of the generated straight line so that it can be better combined with the background, specifically as follows:

[0087]

[0088] Wherein, (i, j) are the coordinates of the Gaussian kernel, G(i, j) is the weight of the Gaussian kernel, I(x + i, y + j) is the pixel value of the neighborhood, σ is the smoothness degree, k is the radius of the kernel, and I output (x, y) is the value of the processed image at the pixel position (x, y).

[0089] Step S03: As Figure 4 shown, perform similarity measurement on the generated defects and the real defects, and select the generated defects and real defects with higher similarity to construct a dataset together.

[0090] Step S03 includes the following steps: Step S31: Set the real defect samples as the target domain samples Generate defect samples as the source domain samples

[0091] Wherein, t and s represent the target domain and the source domain respectively, i and j represent the sample numbers of the target domain and the source domain, represent the samples of the target domain and the source domain respectively, and n and m represent the total numbers of the target domain samples and the source domain samples

[0092] Step S32: In one embodiment, the specific method for extracting the feature vectors of the target domain and source domain samples is as follows: Input the target domain and source domain samples into the feature extraction network, and use the feature vectors of the intermediate layer of the network to represent the sample features, specifically as follows:

[0093]

[0094] In the formula, are the feature representations of the target domain and source domain samples respectively, and F(·) represents the feature extraction network,

[0095] Step S33: In one embodiment, the specific method for performing similarity measurement is as follows: Use cosine similarity to calculate the similarity between samples For each source domain sample, calculate its similarity score S with all target domain samples j , specifically as follows:

[0096]

[0097]

[0098] Step S34: Set the similarity threshold as τ, and retain the generated samples that satisfy S j ≥τ. Mix the generated samples and the real samples, label them, and divide the dataset into a training set and a test set according to the ratio of 8:2.

[0099] Step S04: As Figure 5As shown in the figure, the YOLOv10 deep learning network is improved and pre-trained to obtain a defect detection model:

[0100] Step S04 includes the following steps: Step S41: Design a cross-scale efficient feature fusion network to replace the feature fusion network of the original network.

[0101] Step S42: Improve the C2f modules in the backbone network and the neck network respectively.

[0102] Step S43: In the neck network, use the lightweight dynamic upsampling operator DySample to replace the original upsampling module.

[0103] Step S44: Replace the original loss function CIoU with the Focal-EIoU loss function.

[0104] Step S45: Use the labeled dataset to pre-train in the improved YOLOv10 deep learning network, and take the result with the best training effect as the defect detection model.

[0105] Furthermore, in one embodiment, the specific method of designing a cross-scale efficient feature fusion network to replace the feature fusion network of the original network is as follows: Feature outputs are added after the 2nd, 4th, and 6th layers of the original network respectively. The feature output of the 2nd layer is connected to the 15th layer, the feature output of the 4th layer is connected to the 12th layer, and the feature output of the 6th layer is connected to the 21st layer. The Concat module at the 12th layer after improving the network performs deep feature fusion on the outputs of the 4th, 6th, and 11th layers of the network. The Concat module at the 15th layer performs deep feature fusion on the outputs of the 2nd, 4th, and 14th layers of the network. The Concat module at the 18th layer performs deep feature fusion on the 13th and 17th layers of the network. The Concat module at the 21st layer performs deep feature fusion on the 6th, 10th, and 20th layers of the network.

[0106] Furthermore, in one embodiment, the specific method of improving the C2f modules in the backbone network and the neck network respectively is as follows: In the backbone network, as Figure 6 shown, add the DAttention attention mechanism after two layers of Conv, improve the Bottleneck to DA_Bottleneck, and improve it to the C2f_DA module by replacing all the Bottlenecks in the 8th layer C2f module with DA_Bottleneck. In the neck network, replace all the C2f modules in the 13th, 16th, and 19th layers with the C2fCIB module, and keep the 22nd layer C2fCIB module unchanged.

[0107] Further, in one embodiment, in the neck network, the specific method of replacing the original upsampling module with the lightweight dynamic upsampling operator DySample is as follows: Replace the 11th and 14th upsampling modules UpSample in the improved network with the lightweight dynamic upsampling operator DySample.

[0108] Further, in one embodiment, the Focal-EIoU loss function for replacing the original loss function is specifically as follows:

[0109]

[0110] L Focal-EIOU = IOU γ L EIOU

[0111] In the formula, L EIOU is the EIOU loss function, L IOU is the intersection over union loss function, L dis is the distance loss function, L asp is the aspect ratio loss function. IOU is the intersection over union, b is the center point coordinate of the predicted bounding box, b gt is the center point coordinate of the ground truth bounding box, w is the width of the predicted bounding box, w gt is the width of the ground truth bounding box, h is the height of the predicted bounding box, h gt is the height of the ground truth bounding box, w c is the width of the smallest bounding box enclosing the predicted bounding box and the ground truth bounding box, h c is the height of the smallest bounding box enclosing the predicted bounding box and the ground truth bounding box, ρ is the Euclidean distance between two points, and γ is the focal adjustment factor.

[0112] Step S05: Use the pre-trained defect detection model to perform real-time defect detection on the solder balls of the BGA package chip and obtain defect information: Input a set of real-time collected solder ball pictures of the BGA package chip into the defect detection model configured with pre-trained weights, detect the solder ball pictures of the BGA package chip, and obtain the positions of the BGA package chips with solder ball defects and the types of solder ball defects of the BGA package chips.

Claims

1. A method for detecting solder ball defects of BGA packaged chips based on improved YOLOv10, characterized in that: It includes the following steps: Step S01: Collect the solder ball images of the BGA package chip, perform morphological preprocessing on the images, and then conduct data analysis and classification on the defects; Step S02: Generate defects for the solder ball burr defects with less data; Step S03: Perform feature similarity measurement on the generated defect data and the real defect data, and select the generated defects and real defects with higher similarity to construct a data set together; Step S04: Improve the YOLOv10 deep learning network and conduct pre-training to obtain a defect detection model; Step S05: Use the defect detection model to conduct real-time defect detection on the solder balls of the BGA package chip to obtain defect information.

2. The method for detecting solder ball defects of BGA packaged chips based on the improved YOLOv10 according to claim 1, wherein: The specific method of the said Step S01 includes the following steps: Step S11: Convert the solder ball image of the BGA package chip collected by the camera into a grayscale image, and then convert the grayscale image into a binary image; Step S12: Select the shape of the structural element according to the shapes of the noise and dirt existing in the collected image, and adjust the size of the structural element to make it close to the size of the target; Step S13: Use opening operation, first perform morphological erosion, and then perform morphological dilation to remove small dirt or noise points in the image; Step S14: Use closing operation, first perform morphological dilation, and then perform morphological erosion to fill the small holes in the foreground while maintaining the shape of the foreground area; Step S14: Combine the binary image after morphological processing with the original image to generate a final image with noise and dirt removed; Step S15: Analyze the defects in the collected image and classify them, which are solder ball bridging, solder ball abnormal shape, solder ball missing, and solder ball burr in sequence.

3. The BGA package chip solder ball defect detection method based on the improved YOLOv10 according to claim 1, characterized in that: The specific method of the said Step S02 includes the following steps: Step S21: Input a group of defect-free images, randomly locate a solder ball in the image, and locate the center of the solder ball; Step S22: Obtain the color channels of the central region of the target box, and calculate the maximum value C and the average value C of the color channels. Then, perform weighted extraction of the color from the maximum value to the average value through Gaussian weighting to obtain the weighted value C(x) of the color channel, as follows: max and the average value C avg , and perform weighted extraction of the color from the maximum value to the average value through Gaussian weighting to obtain the weighted value C smooth (x), specifically as follows: where C i is the color value of the i-th pixel, n is the number of pixels in the region, x is the distance from the center point of the solder ball, and σ is the width controlling the transition; Step S23: Randomly generate a line with specified restrictions on the image. The starting point (x start , y start ) of the line is the center of the solder ball, and the end point of the line is (x end , y end ), and fill it with the extracted color, specifically as follows: x end = x start + L·cos(θ) y end = y start + L·sin(θ) In the formula, L is an arbitrary line length; Step S24: Perform mosaic processing and Gaussian smooth transition on the generated straight line, specifically as follows: Where (i,j) are the coordinates of the Gaussian kernel, G(i,j) is the weight of the Gaussian kernel, I(x+i,y+j) is the pixel value of the neighborhood, σ is the smoothness level, k is the radius of the kernel, and I output (x,y) is the value of the processed image at the pixel position (x,y).

4. The method for detecting solder ball defects of BGA packaged chips based on the improved YOLOv10 according to claim 1, wherein: The specific method in Step S03 includes the following steps: Step S31: Set real defect samples as target domain samples Generate defect samples as source domain samples Among them, t and s represent the target domain and the source domain, and i and j represent the sample numbers of the target domain and the source domain. They represent the samples of the target domain and the source domain, and n and m represent the total numbers of the target domain samples and the source domain samples. Step S32: Input the target domain and source domain samples into the feature extraction network, and use the feature vectors of the intermediate layer of the network to represent the sample features, specifically as follows: wherein, are the feature representations of the target domain and source domain samples respectively, and F(·) represents the feature extraction network, Step S33: Calculate the similarity between samples using cosine similarity And for each source domain sample, calculate its similarity score S with all target domain samples j , specifically as follows: Step S34: Set the similarity threshold as τ, and retain the generated samples that satisfy S j ≥ τ. Mix the generated samples and the real samples, and after annotation, divide the dataset into a training set and a test set according to the ratio of 8:

2.

5. The BGA package chip solder ball defect detection method based on the improved YOLOv10 according to claim 1, wherein: The specific method in the said Step S04 is: Step S41: Design a cross-scale efficient feature fusion network to replace the feature fusion network of the original network; Step S42: Improve the C2f modules in the backbone network and the neck network respectively; Step S43: In the neck network, adopt the lightweight dynamic upsampling operator DySample to replace the original upsampling module UpSample; Step S44: Replace the original loss function CIoU with the Focal-EIoU loss function; Step S45: Use the labeled data set to conduct pre-training in the improved YOLOv10 deep learning network, and take the result with the best training effect as the defect detection model.

6. The method for detecting solder ball defects of BGA packaged chips based on the improved YOLOv10 as claimed in claim 5, wherein: In step S41, feature outputs are added respectively after the 2nd, 4th, and 6th layers of the original network. The feature output of the 2nd layer is sent to the 15th layer, the feature output of the 4th layer is sent to the 12th layer, and the feature output of the 6th layer is sent to the 21st layer. The Concat module of the 12th layer in the improved network performs deep feature fusion on the outputs of the 4th, 6th, and 11th layers of the network. The Concat module of the 15th layer performs deep feature fusion on the outputs of the 2nd, 4th, and 14th layers of the network. The Concat module of the 18th layer performs deep feature fusion on the 13th and 17th layers of the network. The Concat module of the 21st layer performs deep feature fusion on the 6th, 10th, and 20th layers of the network.

7. The method for detecting solder ball defects of BGA packaged chips based on the improved YOLOv10 according to claim 5, wherein: In step S42, by replacing Bottleneck with DA_Bottleneck, the C2f module of the 8th layer in the backbone network is improved to the C2f_DA module. In the neck network, all C2f modules of the 13th, 16th, and 19th layers are replaced with C2fCIB modules, and the C2fCIB module of the 22nd layer remains unchanged.

8. The method for detecting solder ball defects of BGA packaged chips based on the improved YOLOv10 according to claim 5, characterized in that: In step S43, the UpSample modules of the 11th and 14th layers in the improved network are replaced with the lightweight dynamic upsampling operator DySample.

9. The BGA package chip solder ball defect detection method based on the improved YOLOv10 according to claim 5, characterized in that: In step S44, the specific replacement of the Focal-EIoU loss function for the original loss function is as follows: L Focal-EIOU = IOU γ L EIOU where L EIOU is the EIOU loss function, L IOU is the Intersection over Union (IOU) loss function, L dis is the distance loss function, L asp is the aspect ratio loss function, IOU is the Intersection over Union, b is the center coordinate of the predicted bounding box, b gt is the center coordinate of the ground truth bounding box, w is the width of the predicted bounding box, w gt is the width of the ground truth bounding box, h is the height of the predicted bounding box, h gt is the height of the ground truth bounding box, w c is the width of the smallest bounding box enclosing the predicted and ground truth bounding boxes, h c is the height of the smallest bounding box enclosing the predicted and ground truth bounding boxes, ρ is the Euclidean distance between two points, and γ is the focal adjustment factor.

10. The method for detecting solder ball defects of BGA packaged chips based on the improved YOLOv10 as described in claim 1, wherein: The specific method in step S05 is as follows: A set of real-time collected BGA package chip solder ball pictures are input into the defect detection model configured with pre-trained weights to detect the BGA package chip solder ball pictures, and the positions of the BGA package chips with solder ball defects and the types of BGA package chip solder ball defects are obtained.

Images