Construction method and application of improved YOLOv8 target detection model for small-size defects

Abstract

The invention discloses an improved YOLOv8 target detection model construction method and application for small-size defects, and the method comprises the steps: constructing an improved YOLOv8 target detection model structure which comprises a Backbone module, a Neck module and a Head detection head; the Backbone module is formed by alternately connecting a plurality of CBS network structures and C2f network structures and finally connected with an SPPF network structure, and is used for carrying out feature extraction on an input image to obtain a plurality of effective feature layers; the Neck module is composed of a Conv network structure, a C2f network structure, an up-sampling network structure, a Concat network structure, a ConcatBiFPN and a Bi-Former attention mechanism network structure, bidirectional multi-scale feature fusion is carried out on an effective feature layer by using the ConcatBiFPN, and local spatial features are extracted and detail information is reserved by using the Bi-Former attention mechanism network structure; the Head detection head is used for outputting a target detection result; carrying out training and pruning processing on the built model structure, and deleting redundant parameters and neurons; the target detection model constructed by the method can perform high-precision detection on small-size defects, and the detection efficiency can be ensured.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to an improved YOLOv8 target detection model construction and a chip surface defect detection method. Background Art

[0002] In the semiconductor manufacturing process, defect detection on the chip surface is crucial. Tiny defects on the chip surface will directly affect its performance and service life, so high-precision defect detection during the manufacturing process is essential. Traditional detection methods rely on manual or simple automated tools, which are difficult to effectively identify defects with small sizes or complex shapes. This deficiency not only reduces the efficiency of detection, but may also cause unqualified products to enter the market, thereby affecting the performance of chips in high-end electronic devices.

[0003] However, due to the complex surface structure of the chip, small defect size, and diverse shapes, traditional detection methods have great limitations in accuracy and speed. In recent years, with the development of deep learning technology, defect detection methods based on target detection models have gradually become mainstream. As an advanced target detection algorithm, YOLOv8 has the advantages of real-time and high precision, but when dealing with small-sized defects, such as small-sized defects with an area of ​​less than 32*32 pixels, there will be problems such as insufficient feature extraction and low detection accuracy, as well as the high computational complexity and memory usage of existing target detection algorithms. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, this application proposes an improved YOLOv8 target detection model construction and chip surface defect detection method, which can not only meet the accuracy of small-size chip defect detection, but also improve the efficiency of detection.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for constructing an improved YOLOv8 target detection model for small-size defects includes the following steps:

[0007] Step 1: Build an improved YOLOv8 target detection model structure, where the improved YOLOv8 target detection model includes a Backbone module, a Neck module, and a Head detection head;

[0008] The Backbone module is composed of a plurality of CBS network structures and C2f network structures alternately connected and finally connected to a SPPF network structure; the Backbone module performs feature extraction on the input image to obtain a plurality of effective feature layers;

[0009] The Neck module is composed of Conv, C2f network structure, upsampling, Concat network structure, Concat_BiFPN, and Bi-Former attention mechanism network structure, wherein Concat_BiFPN is used to perform bidirectional multi-scale feature fusion on the effective feature layer, and the Bi-Former attention mechanism network structure is used to extract local spatial features and retain detail information;

[0010] The Head detection head is used to output target detection results;

[0011] Step 2: Train and prune the built improved YOLOv8 target detection model structure to delete redundant parameters and neurons.

[0012] Further, the network structure of the Backbone module is: a first CBS network structure, a second CBS network structure, a first C2f network structure, a third CBS network structure, a second C2f network structure, a fourth CBS network structure, a third C2f network structure, a fifth CBS network structure, a fourth C2f network structure and an SPPF network structure are connected in sequence; wherein the second C2f network structure, the third C2f network structure and the SPPF network structure output the first effective feature layer, the second effective feature layer and the third effective feature layer respectively.

[0013] Further, the network structure of the Neck module is: a first upsampling, a first Concat_BiFPN, a fifth C2f network structure, a second upsampling, a second Concat_BiFPN, a sixth C2f network structure, a first Bi-Former attention mechanism network structure, a first Conv network structure, a seventh C2f network structure, a second Bi-Former attention mechanism network structure, a second Conv network structure, a third Concat_BiFPN, an eighth C2f network structure, and a third Bi-Former attention mechanism network structure are connected in sequence; wherein, the first upsampling and the third Concat_BiFPN are connected to the SPPF network structure to receive the third effective feature layer; the first Concat_BiFPN is connected to the third C2f network structure to receive the second effective feature layer; the second Concat_BiFPN is connected to the second C2f network structure to receive the first effective feature layer.

[0014] Furthermore, the Bi-Former attention mechanism network structure includes a deep separable convolution, a first normalization layer, a Bi-Level Routing Attention module, a second normalization layer, and a multi-layer perceptron module; the deep separable convolution performs a convolution operation on the input features; the first normalization layer normalizes the features after the convolution operation; the Bi-Level Routing Attention module screens and adaptively weights the features at the coarse-grained and fine-grained levels, retaining key information while suppressing noise information; the second normalization layer performs LN normalization on the feature map output by the Bi-Level Routing Attention module; and the multi-layer perceptron module performs linear transformation and activation operations on the feature map.

[0015] Furthermore, the feature processing process in the Bi-Level Routing Attention module is as follows: extract features from the image, generate three matrices from the extracted features, the Q matrix is ​​the feature that needs to be paid attention to, and then generate the K matrix and the V matrix through the gather operation, and calculate the similarity between Q and K through GPU-friendly dense matrix multiplication to form a similarity matrix; obtain the attention map A through normalization processing, and finally perform dense matrix multiplication operation on the matrix A and the matrix V to obtain the final output feature O.

[0016] Furthermore, the process of implementing bidirectional multi-scale feature fusion in Concat_BiFPN is expressed as:

[0017] F i fused =Conv i (ReLU(Conv1(F i +Upsample(F i+1 ))+Downsample(F i )+F i+1 ))

[0018] in, is the final feature map, F i is the low-level feature map, F i+1 is a high-level feature map, Upsample(F i+1 ) is the high-level feature map F i+1 Upsample to low-level feature map F i Resolution, Downsample(F i ) is the low-level feature map F i Downsample to high-level feature map F i+1 Conv1 is a convolution operation between the upsampled feature map and the original feature map, and ReLU is the activation function; Convi A convolution operation is performed on the result of a nonlinear transformation.

[0019] Furthermore, the method for training and pruning the target detection model in step 2 is:

[0020] Step 2.1, during training, use L1 regularization to impose sparse constraints on the weights W of each layer in the improved YOLOv8 object detection model;

[0021] Step 2.2, after the sparse training, a threshold θ is set, and weights smaller than the threshold are removed or set to zero to reduce redundant connections in the model;

[0022] Step 2.3: After pruning is completed, retrain the model to restore some connections in the model.

[0023] Furthermore, the L1 regularization formula is:

[0024]

[0025] Among them, L total is the total loss function, L detection is the original loss function of YOLOv8, λ is the sparsification weight factor, W i is the weight of the i-th layer.

[0026] A chip surface defect detection method based on an improved yolov8 model comprises the following steps:

[0027] Step 1: Use an industrial camera to collect images of surface defects of the chip to be identified;

[0028] Step 2: Based on the chip surface defect image data, the above-mentioned improved YOLOv8 target detection model construction method for small-size defect detection is used to build a target detection model, and the image of the chip surface defects to be identified is input into the trained target detection model to predict the chip surface defects and output the information of chip surface defects in the industrial production process.

[0029] Furthermore, pictures of chip surface defects are collected and the defects in the pictures are annotated to form a data set; the data set is divided according to the ratio of training set: validation set: test set = 8:1:1; and used for training the target detection model.

[0030] Beneficial effects of the present invention:

[0031] (1) Adding the Bi-Former attention mechanism network structure to the Neck module of the YOLOv8 target detection model can capture more global features while effectively suppressing noise and irrelevant information, allowing the model to focus on key target areas. Its multi-head self-attention mechanism can learn the correlation between different regions more finely, thereby improving the recognition accuracy of the model. In addition, Bi-Former can be seamlessly integrated with other network modules to enhance the detection and classification performance of the overall model, which is especially suitable for high-precision target detection tasks.

[0032] (2) The Bi-Level Routing Attention module in the Bi-Former attention mechanism network structure can filter and adaptively weight features at coarse-grained and fine-grained levels, retaining key information while suppressing irrelevant noise information. Therefore, it can handle the problem that small-sized defects have weak features and are easily confused with the background or other irrelevant areas, resulting in false detection or missed detection.

[0033] (3) After the improved YOLOv8 target detection model is built, the network structure of the model is pruned and redundant parameters and neurons are deleted to improve the reasoning speed and computational efficiency of the model while maintaining high-precision detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the structure of the improved YOLOv8 target detection model of the present invention.

[0035] Figure 2 It is a structural diagram of the Bi-Former attention mechanism module.

[0036] Figure 3 This is a schematic diagram of the Bi-Level Routing Attention structure.

[0037] Figure 4 This is a schematic diagram of the Concat_BiFPN structure.

[0038] Figure 5 It is a schematic diagram of the pruning structure.

[0039] Figure 6 This is a flow chart of chip surface defect detection using the improved YOLOv8 target detection model. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] Combined with Figure 1-4 , the present application provides a method for constructing an improved YOLOv8 target detection model for small-size defect detection, comprising the following steps:

[0042] Step 1: Build an improved YOLOv8 target detection model structure. Figure 2 The improved YOLOv8 target detection model includes Backbone module, Neck module and Head detection head; the image is input into the Backbone module for feature extraction, and three effective feature layers of different scales are extracted; in the Neck module, three deeply fused output feature maps are finally generated through multi-level feature fusion and superposition of attention mechanism, providing multi-scale global information and details for target detection. The Head detection head outputs the detection result based on the feature map output by the Neck module. The specific structure is:

[0043] The Backbone module is mainly composed of 5 CBS network structures, 4 C2f network structures and one SPPF network structure; by adding the CBS network structure, the features extracted by the CBS module will be spliced ​​and fused. The specific connection relationship of the Backbone module is as follows: it is composed of the first CBS network structure, the second CBS network structure, the first C2f network structure, the third CBS network structure, the second C2f network structure, the fourth CBS network structure, the third C2f network structure, the fifth CBS network structure, the fourth C2f network structure and the SPPF network structure connected in sequence; among which, the second C2f network structure, the third C2f network structure and the SPPF network structure output the first effective feature layer, the second effective feature layer and the third effective feature layer respectively.

[0044] The Neck module is composed of 2 Conv, 4 C2f network structures, 2 upsampling, 1 Concat network structure, 3 Concat_BiFPN, and 3 Bi-Former attention mechanism network structures, which fuse multiple effective feature layers. The specific connection relationship of the Neck module is as follows: the first upsampling, the first Concat_BiFPN, the fifth C2f network structure, the second upsampling, the second Concat_BiFPN, the sixth C2f network structure, the first Bi-Former attention mechanism network structure, the first Conv network structure, the seventh C2f network structure, the second Bi-Former attention mechanism network structure, the second Conv network structure, the third Concat_BiFPN, the eighth C2f network structure, and the third Bi-Former attention mechanism network structure are connected in sequence; among them, the first upsampling and the third Concat_BiFPN are connected to the SPPF network structure to receive the third effective feature layer; the first Concat_BiFPN is connected to the third C2f network structure to receive the second effective feature layer; the second Concat_BiFPN is connected to the second C2f network structure to receive the first effective feature layer. The output input head part of the first Bi-Former attention mechanism network structure, the second Bi-Former attention mechanism network structure and the third Bi-Former attention mechanism network structure outputs the final feature information.

[0045] More specifically, the Bi-Former attention mechanism network structure can be referenced Figure 2In the workflow of the Bi-Former attention mechanism network structure, the input features first pass through the depthwise separable convolution (DWConv 3x3), which is an important convolution operation used to extract local spatial features and retain more detailed information. After that, the feature map is normalized by the normalization layer (LN) module to ensure data stability and avoid gradient explosion or disappearance during training. Next, the feature map enters the Bi-Level Routing Attention module, where the feature map can filter and adaptively weight features at coarse and fine granularity levels, retaining key information while suppressing irrelevant noise information. The feature map output by the Bi-Level Routing Attention module is LN normalized again to ensure stability before entering the multi-layer perceptron (MLP) module. In the MLP, the feature map undergoes further linear transformation and activation operations to enhance the nonlinear expression ability of the model. Throughout the process, the residual connection adds the output of each part to the initial input to achieve information transmission and fusion, so that the model can not only capture important features, but also effectively suppress noise, improving the detection performance and expression ability of the model. Therefore, the Bi-Former attention mechanism network structure can capture more global features while effectively suppressing noise and irrelevant information, allowing the model to focus on key target areas. Its multi-head self-attention mechanism can learn the correlation between different regions more finely, thereby improving the recognition accuracy of the model. In addition, Bi-Former can be seamlessly integrated with other network modules to enhance the detection and classification performance of the overall model, especially for high-precision target detection tasks.

[0046] like Figure 3 The specific implementation of the Bi-Level Routing Attention module is as follows: First, extract features from the image and generate three matrices from the extracted features. The Q matrix is ​​the feature that needs to be paid attention to. Then, the K matrix (all features that may be paid attention to, used for feature indexing and matching) and the V matrix (storing specific feature details related to K) are generated through the gather operation. Then, the similarity between Q and K is calculated through GPU-friendly dense matrix multiplication (mm) to form a similarity matrix. The attention map A is obtained through normalization processing (softmax). Finally, the matrix A and the matrix V are operated by mm to obtain the final output feature O.

[0047] In the Neck module, Concat_BiFPN is used to achieve bidirectional multi-scale feature fusion, the formula is:

[0048]

[0049] in, is the final feature map, F i is the low-level feature map, F i+1 is a high-level feature map, Upsample(F i+1 ) is the high-level feature map F i+1 Upsample to low-level feature map F i Resolution, Downsample(F i ) is the low-level feature map F i Downsample to high-level feature map F i+1 Conv1 is a convolution operation performed on the upsampled feature map and the original feature map. ReLU is an activation function used to introduce nonlinear transformation. i A convolution operation is performed on the result of a nonlinear transformation.

[0050] like Figure 3 First, the three feature maps of different scales output by Backbone enter the Neck part. After the first feature map is processed by the C2f module, it enters the Bi-Former attention mechanism for feature enhancement. The second feature map is fused by C2f and Concat_BiFPN, and then the feature size is aligned through upsampling, and it is fused with the features of the previous level, and finally input into the Bi-Former module. Similarly, the third feature map is fused with the features from the previous network layer through a Conv layer and Concat operation, and then further processed by the C2f module, and finally enters the Bi-Former mechanism for enhancement. Through this multi-level feature fusion and the superposition of the attention mechanism, the Neck part finally generates three deeply fused output feature maps, providing multi-scale global information and details for target detection.

[0051] The Head detection head is used to output the target detection results, including convolutional layer, bounding box regression, category prediction, anchor box generation, loss function, and non-maximum suppression.

[0052] Step 2: Train and prune the built improved YOLOv8 target detection model, delete redundant parameters and neurons, so as to improve the reasoning speed and computing efficiency of the model while maintaining high-precision detection performance.

[0053] Since the present invention adds Bi-Former attention mechanism and Concat_BiFPN in the Neck module of the target detection model, the global information and details of multiple scales can be obtained through the superposition of feature fusion and attention mechanism, but at the same time, the number of parameters and the amount of calculation of the model will be increased. Therefore, the present invention prunes the network structure of the model, as follows:

[0054] Step 2.1, during training, use L1 regularization to impose sparse constraints on the weights W of each layer in the improved YOLOv8 target detection model, the formula is:

[0055]

[0056] Among them, L total is the total loss function, L detection is the original loss function of YOLOv8, λ is the sparsification weight factor, W i is the weight of the i-th layer.

[0057] Step 2.2, after sparse training, set a threshold θ, and weights less than the threshold will be removed or set to zero to reduce redundant connections in the model. The basic formula for pruning is:

[0058] W′=W*I

[0059] Among them, W is the weight matrix before pruning, W′ is the weight matrix after pruning, and I is an indicator function.

[0060] In this embodiment, if the threshold θ is zero, when the absolute value of W is greater than the threshold θ, the weight is retained, otherwise it is pruned.

[0061] Step 2.3, after pruning is completed, the model usually has a certain performance degradation. Therefore, it is necessary to fine-tune the pruned model. By retraining the model, some performance can be restored and further optimized to ensure that high accuracy can be maintained with fewer parameters. The goal of fine-tuning is to balance the pruned lightweight model between inference efficiency and accuracy.

[0062] The data set used in the training process of step 2 above is prepared as follows:

[0063] 1000 pictures of surface defects of optoelectronic chips are taken with an industrial camera and annotated using Labelimg with the label name "defect". In this embodiment, the annotation format adopts the PascalVOC format.

[0064] More specifically, the images in the dataset are preprocessed and data enhancement is performed on the images through translation, rotation, cropping, flipping, etc. Then, all images are adjusted to the size required for model training, with a unified resolution of 640*640, and normalized. The processed images are divided into a ratio of training set: validation set: test set = 8:1:1.

[0065] More specifically, the model training and pruning process is as follows:

[0066] The improved YOLOv8 target detection model is trained using the training set and validation set of the dataset. L1 regularization is used to impose sparse constraints on the weights W of each layer. The training epoch is set to 600 times. The backbone feature extraction network is frozen in the first 200 epochs and the initial learning rate is set to 0.01. The backbone feature extraction network is unfrozen in the last 400 epochs and the initial learning rate is set to 0.0001. After the training is completed, the model is pruned and fine-tuned to ensure recognition speed while ensuring recognition accuracy. Target detection is verified on the test set of the constructed chip surface defect dataset to complete the accurate and efficient recognition of chip surface defects.

[0067] A chip surface defect detection method based on an improved yolov8 model comprises the following steps:

[0068] Step 1: Use an industrial camera to collect images of surface defects of the chip to be identified;

[0069] Step 2: Input the image of the chip surface defects to be identified into the improved YOLOv8 target detection model constructed by the above-mentioned improved YOLOv8 target detection model construction method for small-size defect detection, predict the chip surface defects, and output the information of the chip surface defects in the industrial production process, including the coordinates and size (x, y, w, h) of the bounding box, the category confidence, the probability distribution of the target category, and the inference speed and accuracy of the model after pruning optimization.

[0070] It has been verified that the method of the present invention can be used to detect small-sized defects, and its accuracy and speed are improved compared with the original yolov8 network.

[0071] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for constructing an improved YOLOv8 target detection model for small-size defects, characterized in that: The steps include: Step 1: Build an improved YOLOv8 target detection model structure, where the improved YOLOv8 target detection model includes a Backbone module, a Neck module, and a Head detection head; The Backbone module is composed of a plurality of CBS network structures and C2f network structures alternately connected and finally connected to a SPPF network structure; the Backbone module performs feature extraction on the input image to obtain a plurality of effective feature layers; The Neck module is composed of Conv, C2f network structure, upsampling, Concat network structure, Concat_BiFPN, and Bi-Former attention mechanism network structure, wherein Concat_BiFPN is used to perform bidirectional multi-scale feature fusion on the effective feature layer, and the Bi-Former attention mechanism network structure is used to extract local spatial features and retain detail information; The Head detection head is used to output target detection results; Step 2: Train and prune the built improved YOLOv8 target detection model structure to delete redundant parameters and neurons.

2. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 1, characterized in that: The network structure of the Backbone module is: a first CBS network structure, a second CBS network structure, a first C2f network structure, a third CBS network structure, a second C2f network structure, a fourth CBS network structure, a third C2f network structure, a fifth CBS network structure, a fourth C2f network structure and an SPPF network structure are connected in sequence; wherein the second C2f network structure, the third C2f network structure and the SPPF network structure output the first effective feature layer, the second effective feature layer and the third effective feature layer respectively.

3. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 1, characterized in that: The network structure of the Neck module is: the first upsampling, the first Concat_BiFPN, the fifth C2f network structure, the second upsampling, the second Concat_BiFPN, the sixth C2f network structure, the first Bi-Former attention mechanism network structure, the first Conv network structure, the seventh C2f network structure, the second Bi-Former attention mechanism network structure, the second Conv network structure, the third Concat_BiFPN, the eighth C2f network structure, and the third Bi-Former attention mechanism network structure are connected in sequence; wherein, the first upsampling and the third Concat_BiFPN are connected to the SPPF network structure to receive the third effective feature layer; the first Concat_BiFPN is connected to the third C2f network structure to receive the second effective feature layer; the second Concat_BiFPN is connected to the second C2f network structure to receive the first effective feature layer.

4. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 1, characterized in that: The Bi-Former attention mechanism network structure includes a deep separable convolution, a first normalization layer, a Bi-Level Routing Attention module, a second normalization layer, and a multi-layer perceptron module; the deep separable convolution performs a convolution operation on the input features; the first normalization layer performs normalization processing on the features after the convolution operation; The Bi-Level Routing Attention module screens and adaptively weights features at coarse-grained and fine-grained levels, retaining key information while suppressing noise information; the second normalization layer performs LN normalization on the feature map output by the Bi-Level RoutingAttention module; and the multi-layer perceptron module performs linear transformation and activation operations on the feature map.

5. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 1, characterized in that: The feature processing process in the Bi-Level Routing Attention module is as follows: extract features from the image, generate three matrices from the extracted features, the Q matrix is ​​the feature that needs to be paid attention to, and then generate the K matrix and the V matrix through the gather operation. The similarity between Q and K is calculated through GPU-friendly dense matrix multiplication to form a similarity matrix; the attention map A is obtained through normalization processing, and finally the matrix A and the matrix V are subjected to dense matrix multiplication to obtain the final output feature O.

6. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 1, characterized in that: The process of implementing bidirectional multi-scale feature fusion in Concat_BiFPN is expressed as: in, is the final feature map, F i is the low-level feature map, F i+1 is a high-level feature map, Upsample(F i+1 ) is the high-level feature map F i+1 Upsample to low-level feature map F i Resolution, Downsample(F i ) is the low-level feature map F i Downsample to high-level feature map F i+1 Conv1 is a convolution operation between the upsampled feature map and the original feature map, and ReLU is the activation function; Conv i A convolution operation is performed on the result of a nonlinear transformation.

7. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 1, characterized in that: The method for training and pruning the target detection model in step 2 is: Step 2.1, during training, use L1 regularization to impose sparse constraints on the weights W of each layer in the improved YOLOv8 object detection model; Step 2.2, after the sparse training, a threshold θ is set, and weights smaller than the threshold are removed or set to zero to reduce redundant connections in the model; Step 2.3: After pruning is completed, retrain the model to restore some connections in the model.

8. The method for constructing an improved YOLOv8 target detection model for small-size defects according to claim 7, characterized in that: The L1 regularization formula is: Among them, L total is the total loss function, L detection is the original loss function of YOLOv8, λ is the sparsification weight factor, W i is the weight of the i-th layer.

9. A chip surface defect detection method based on an improved yolov8 model, characterized in that: The steps include: Step 1: Use an industrial camera to collect images of surface defects of the chip to be identified; Step 2: Based on the chip surface defect image data, a target detection model is constructed using the improved YOLOv8 target detection model construction method for small-size defect detection described in claim 1, and the image of the chip surface defects to be identified is input into the trained target detection model to predict the chip surface defects, and output the information of chip surface defects in the industrial production process.

10. A chip surface defect detection method based on an improved yolov8 model according to claim 9, characterized in that: Collect pictures of chip surface defects and annotate the defects in the pictures to form a data set; divide the data set in the ratio of training set: validation set: test set = 8:1:1; and use it to train the target detection model.

Images