An improved FCOS algorithm-based target detection method for an unmanned vehicle-mounted camera

Abstract

The application relates to an unmanned vehicle-mounted camera target detection method based on an improved FCOS algorithm, and relates to the field of computer vision. An image is collected by an unmanned vehicle, the image is pretreated, and then is put into an improved FCOS network model for training. In the training process, the model performs feature extraction, prediction, loss calculation and parameter updating on the image. After multiple iterations, a trained detection model file can be obtained. After model conversion, the model can be applied and deployed on terminal equipment such as an unmanned vehicle. The application has stronger feature extraction capability, and the two-stage model constructed has better detection effect on small targets, effectively improves the recognition accuracy of the model, and improves the missed detection and false detection of the model.

Description

Technical Field

[0001] This invention relates to the field of computer vision image processing technology, and in particular to a target detection method for unmanned vehicles equipped with cameras based on an improved FCOS algorithm. Background Technology

[0002] The automotive industry has developed for over 200 years, undergoing multiple rounds of technological iteration and industrial upgrading. Automobiles have evolved from mechanization to electrification and electronicization, and today, intelligent vehicles have become a new theme in automotive development. Intelligent vehicles, through installed sensors, perceive surrounding dynamic and static obstacles in real time, enabling timely hazard warnings and driver assistance. Autonomous intelligent vehicles can completely free the driver's hands, achieving environmental perception, self-localization, and autonomous vehicle planning and control. Image processing and target detection are key issues in environmental perception.

[0003] Patent No. CN202111633725.0, Patent Title: A Target Detection Method for Unmanned Vehicles Equipped with Cameras Based on YOLOv5, discloses a target detection method for unmanned vehicles equipped with cameras based on YOLOv5. The method includes an unmanned vehicle body and a camera installed at the front end of the unmanned vehicle body. The camera acquires images, which are then pre-processed and categorized before being fed into a YOLOv5 network model for training. During training, the model extracts feature points from the images in the training set. After multiple iterations, a trained detection model file is obtained (in .pt format). The best-performing detection model is then deployed to embedded devices and applied. This invention can extract more feature information during training, greatly enhancing the YOLO algorithm's performance in detecting highly concealed targets, effectively improving recognition accuracy, achieving good training results, and having a low false negative rate.

[0004] To address the issue of the large number of parameters in the original YOLOv5 detection model, this paper replaces the BottleneckCSP module with the GhostBottleneck module and replaces the Conv module in the Neck with the DWConv module. This solves the problems of YOLOv5's low performance in recognizing well-concealed targets and its large number of model parameters.

[0005] This application proposes a detection model based on FCOS. Its anchor-free two-stage structure, compared to existing anchor-based models, exhibits stronger generalization ability, simpler model structure, and higher detection accuracy. Furthermore, its multi-branch structure and reparameterization technique enhance the model's feature extraction capability without increasing the number of parameters during model inference. The two-stage structure also improves the detection capability of the original FCOS model for small targets.

[0006] Patent No. CN201810530668.5, Patent Title: A Driving Scene Object Detection Method Based on Deep Learning and Multi-layer Feature Fusion, discloses a driving scene object detection method based on deep learning and multi-layer feature fusion. This method includes the following steps: 1) acquiring video images as a training dataset and preprocessing them; 2) building a training network; 3) initializing the training network to obtain a pre-trained model; 4) training the pre-trained model obtained in step 3) on the training dataset to obtain a trained model; 5) acquiring images of the front using an in-vehicle camera and inputting these images into the trained model obtained in step 4) to obtain detection results. This invention employs a multi-layer feature fusion method based on feature pyramids to enhance the semantic information of low-level feature maps, thereby improving the feature extraction quality of the network and achieving higher detection accuracy.

[0007] To address the slow detection speed of SSD detection algorithms, a lightweight feature extraction network, MobileNet, is employed as the feature extraction network. A multi-layer feature fusion method based on feature pyramids is used to enhance the semantic information of the low-level feature maps. This solves the problems of slow SSD detection speed and low small target detection capability in traditional SSD detection algorithms.

[0008] This application proposes a detection model based on FCOS. Its anchor-free two-stage structure, compared to existing anchor-based models, eliminates the need for anchor design, simplifying the model's usage and improving its generalization ability. It employs a lightweight backbone network, VargNet, and incorporates multi-branch and reparameterized structures. This structure allows the invention to improve accuracy without increasing the network's computational cost. Furthermore, the constructed two-stage structure addresses the low performance of FCOS in small target detection.

[0009] In the field of computer image vision inspection, deep learning neural network models can be continuously trained to acquire the ability to distinguish and identify objects of different categories, and these models can exhibit good performance. The FCOS algorithm is an anchor-free detection algorithm. Compared with anchor-based detection networks, it does not require setting anchor-related parameters, greatly increasing the generalization performance and portability of the detection model and reducing the design burden on users. The FCOS network has a simple structure, no complex hyperparameter limitations, and is highly scalable. The FCOS network can provide excellent detection performance while maintaining speed. Summary of the Invention

[0010] To address the above issues, this invention proposes a target detection method for unmanned vehicles equipped with cameras based on an improved FCOS algorithm. This method offers enhanced feature extraction capabilities, and the constructed two-stage model demonstrates superior detection performance on small targets, effectively improving the model's recognition accuracy and reducing false positives and false negatives.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] A target detection method for unmanned vehicles equipped with cameras based on an improved FCOS algorithm, the specific steps of which are as follows:

[0013] (1) Install the camera at the front of the vehicle so that the camera can capture the front of the vehicle; while the vehicle is in motion, the camera captures video of the vehicle's movement.

[0014] (2) Extract keyframes from the acquired video stream information, preprocess the extracted images, including filtering duplicate images, deleting images without targets, and labeling all images. The label file contains the image category information and target location information. Divide the preprocessed images into a pre-training image dataset for training, a verification image dataset for verification, and a test image dataset for testing.

[0015] (3) Build the improved FCOS model, set the training parameters of the model, use the preprocessed image in step (2) to set the parameters of the improved FCOS model, and train the model. During the training process, feature extraction will be performed on the images in the training set. After multiple iterations, the trained detection model is obtained.

[0016] (4) Check the effect of the model, analyze the final saved model, and check whether the model has overfitting or underfitting. If overfitting or underfitting occurs, adjust the number of training times and hyperparameters according to the experimental results to obtain a reasonable detection model.

[0017] (5) Set the score threshold and IOU threshold, decode the output of the model, perform non-maximum suppression on the decoded prediction boxes according to the set threshold, delete redundant target boxes, and finally output the target detection results.

[0018] In the post-processing stage, the confidence scores of the candidate boxes in the first stage and the classification confidence scores predicted in the second stage are combined as the final output classification confidence score.

[0019] As a further improvement of the present invention, the specific method of key frame extraction in step (2) is as follows: the video image data collected in step (1) is extracted as key frames, and one frame is extracted every 10 frames as an image in the dataset.

[0020] As a further improvement of the present invention, the specific method of image preprocessing in step (2) is as follows:

[0021] First, all extracted images are sorted out, and duplicate, invalid, and irrelevant images are filtered out. Second, the remaining images after the filtering are sorted out and divided into three parts in a ratio of 7:2:1, which are used as the training set for model training, the validation set for model validation, and the test set for testing the model. Finally, the labeling tool is used to label and process the target objects to be detected on the training set and the validation set.

[0022] As a further improvement of the present invention, the improved FCOS network training model described in step (3) is built in the PyTorch environment, the training set, validation set and test set paths obtained in step (2) are added to the model configuration file, and the corresponding parameters are set in the improved FCOS network training model.

[0023] A channel-space attention mechanism module is added after feature maps of different sizes to weight the feature maps in both channel and spatial directions, improving the model's accuracy. A lightweight VargNet backbone network is used to reduce the model's computational cost. A multi-branch structure is proposed based on the original model to enhance its feature representation capabilities, and parameter recalculation is employed to reduce computational cost during inference. A BiFPN structure is added to the feature fusion module. The convolutional modules in the original FCOS detection module are replaced with depthwise separable convolutional modules, and a proposed center-prior-based sampling method and a new Centerness calculation method are added to the FCOS sampling stage. A refine module is added after the FCOS detection module to further improve the model's accuracy. The improved FCOS network model includes a backbone feature extraction network, a feature fusion module, an FCOS Proposal module, and an FCOSRefine module.

[0024] As a further improvement of the present invention, the backbone feature extraction network uses the VargNet network, which uses variable group convolution to replace traditional convolution and incorporates a multi-branch structure and a reparameterization structure. The variable group convolution is similar to depthwise separable convolution and consists of two parts. The first part is the variable group convolution part, which does not use a fixed group form, making it more conducive to the operation of the underlying hardware. The second part is the pointwise convolution part, which is used to ensure the matching of the number of channels. The multi-branch structure consists of multiple 3*3 convolutions, which improves the feature representation capability of the model. During model inference, reparameterization adjusts the model to a single-path structure, maintaining model performance while reducing model resource consumption.

[0025] As a further improvement to this invention, the feature fusion module adopts a BiFPN structure. This structure eliminates nodes with only one input, simplifying the computational load. Furthermore, it adds connection channels between the input and output, improving the model's feature fusion performance. This module simultaneously employs both top-down and bottom-up fusion methods, adding semantic information from the top layer and localization information from the bottom layer, significantly enhancing the model's feature fusion capability. This module also considers that feature maps of different sizes contribute differently to the model, therefore assigning different weights to feature maps of different sizes.

[0026] As a further improvement to this invention, the FCOS Refine module uses the FCOS detection head as the candidate box extraction network in the first stage, incorporating a center prior rule during FCOS positive and negative sample sampling. A square sampling region of a certain side length is set for the location of the ground truth box center, and positive samples are sampled only within this region. For targets of different scales, a scaling factor of the side length is set to control the center sampling region, and a new centrality branch is used to evaluate sample quality. Simultaneously, a Refine module is proposed for the second stage of feature extraction and training. Between the first and second stages, the ROI Align module is used to select candidate boxes from the feature map. This module avoids quantization operations in the coordinate mapping process and uses bilinear interpolation to calculate the target position value, significantly improving localization accuracy.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] (1) This invention constructs a two-stage network, which differs from ordinary two-stage networks that require a large number of candidate boxes during both training and forward inference. The candidate box network based on FCOS can achieve model training convergence without requiring a large number of candidate boxes in the second stage.

[0029] (2) This invention employs an object detection method based on an improved FCOS algorithm, adding a channel spatial attention mechanism. It also utilizes a VargNet backbone network with multi-branch and reparameterized structures to enhance the model's feature representation capabilities and reduce computational complexity. A BiFPN module is used to improve the model's feature fusion capabilities. Finally, FCOS Refine and ROIAlign modules are added. In the data flow of the first and second stages, the number of candidate boxes is reduced, balancing computational speed and model accuracy.

[0030] (3) Compared with the one-stage FCOS model, the present invention has a two-stage detection process, which has better performance. In the second stage, the candidate boxes of the first stage are used to select more detailed feature maps from the large feature maps, which is more conducive to the detection of small targets, so the detection performance of small targets will be better. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the training process of this invention;

[0032] Figure 2 This is a diagram of the improved FCOS network model architecture in this invention;

[0033] Figure 3 This is a diagram showing the detection effect of the improved FCOS network model in a practical application. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0035] like Figure 2 As shown, this invention introduces the Refine module into the FCOS network to construct a two-stage anchor-free network; it adopts a lightweight backbone network, incorporating multi-branch ideas and reparameterized structures; it adds a channel attention mechanism after the feature map; it uses BiFPN, which has stronger fusion capabilities, as the feature fusion module; it uses the ROIAlign module to extract candidate boxes from the feature map; and it proposes a new center-prior-based sampling method and a new Centerness calculation method in FCOS. Finally, it trains the network using images captured by an in-vehicle camera. The improved FCOS network model's detection performance in practical applications is shown in the image below. Figure 3 As shown, the specific implementation steps are as follows, and the flowchart of the invention during the training process is as follows. Figure 1 As shown:

[0036] (1) Install the camera at the front of the vehicle so that the camera can capture the front of the vehicle; during the vehicle's movement, the camera collects video information of the vehicle's movement.

[0037] (2) Keyframes are extracted from the acquired video stream information. One frame is extracted every 10 frames as one image in the dataset. The extracted images are then preprocessed. The preprocessed images are organized and divided into a pre-training image dataset for training, a validation image dataset for validation, and a test image dataset for testing, in a 7:2:1 ratio. Image preprocessing includes filtering duplicate images and deleting images without targets. It also includes labeling all valid images with the required categories. The labeling file contains the image category information and target location information. Occluded objects are labeled as hard markers. During training, it is optional to use these occluded target data for training.

[0038] (3) Build the improved FCOS model and set the training parameters of the model. Use the preprocessed images from step (2) to train the improved FCOS model with complete parameter settings. During the training process, feature extraction will be performed on the images in the training set. After multiple iterations, the trained detection model can be obtained.

[0039] (4) The feature extraction backbone network adopts the VargNet network, which uses variable grouped convolutions to replace traditional convolutions. Variable grouped convolutions are similar to depthwise separable convolutions, consisting of two parts. The first part is the variable grouped convolution part. Variable grouped convolutions do not use fixed grouping, which is more conducive to the operation of the underlying hardware. The second part is the pointwise convolution part, used to ensure the matching of the number of channels; it adopts a multi-branch idea and uses reparameterization techniques. The multi-branch idea means that using multiple branches in the feature extraction stage can improve the model's feature representation ability and feature extraction ability. However, the disadvantage of using a multi-branch structure is the increase in the number of model parameters. Therefore, adding reparameterization structures reduces the computational cost of the model. Reparameterization means that in the model inference stage, the parameters of the convolutions of multiple branches are reconstructed, absorbing all convolutions and BN structures into one convolution, making the structure a single path. This maintains the same accuracy while reducing the computational cost. After inputting a 576*704*3 image into the backbone network, it first undergoes a regular convolution to increase the dimensionality, raising the number of channels to 16 dimensions and reducing the feature map size to 288*352. The feature map is then sequentially passed through modules consisting of stacked BasicVarBlocks of varying numbers (1, 3, 7, and 4). Each large block undergoes downsampling, reducing the feature map to half its original size. The output of each module is fed into the next module while simultaneously stored in a list, serving as input to the feature fusion module.

[0040] (5) Add a channel spatial attention mechanism module after feature maps of different sizes to make the feature maps weighted average in both channel and spatial directions, thereby improving the accuracy of the model;

[0041] (6) The feature fusion module adopts a BiFPN structure, which eliminates nodes with only one input, simplifying the computational load. Furthermore, it adds connection channels between the input and output, improving the model's feature fusion performance. This module uses both top-down and bottom-up fusion methods, simultaneously fusing semantic information from the top layer and localization information from the bottom layer, significantly enhancing the model's feature fusion capability. The feature fusion module finally outputs p2, p3, p4, p5, and p6, which are feature maps upsampled by 4, 8, 16, 32, and 64 times respectively, from the input image. The number of output channels is fixed at 32.

[0042] (7) The FCOS Refine module uses the FCOS detection head as the candidate box extraction network in the first stage, and adds the Refine module to perform the feature extraction in the second stage.

[0043] In the first stage, FCOS is used as the candidate box extraction network, and convolutional structures are used for prediction. The convolutional structures share parameters across the four feature layers. During the sample sampling stage, a center prior rule is incorporated. This rule considers sample points in the center region of the ground truth bounding box as positive samples because sample points in the center of the object generally appear on the object, and their receptive field is more suitable for covering the entire object, thus usually outputting higher confidence. Therefore, when sampling positive and negative samples with FCOS, a square sampling region of a certain side length is set at the center of the ground truth bounding box, and only samples within this region are sampled as positive samples. A scaling factor is set for the center sampling region to control the size of the target at different scales. This improvement ensures that the positive sample points used for calculation are all near the center of the target, which better conforms to the center prior rule, while avoiding sampling at the edges, thus reducing the number of low-quality samples. The prediction head has three branches: classification, regression, and centrality branches.

[0044] The classification branch uses Focal loss to mitigate the imbalance between positive and negative samples. The formula is as follows:

[0045] L fl =-α(1-p t ) γ log p t

[0046]

[0047] Where α is a weighting factor, used to adjust the weights of positive and negative samples. γ is a moderating factor, used to adjust the weights of hard samples; y is the true label, and p is the predicted value for classification. Because this classification task is a binary task, i.e., either the target class is true or it is not, a true label of 1 indicates the target class is true (positive sample), and a true label of 0 indicates it is not true (negative sample). t p reflects the degree of proximity to the ground truth, i.e., category y. t The larger the value, the closer it is to category y, meaning the more accurate the classification.

[0048] The regression branch uses CIoU loss, which considers the overlap area of ​​the bounding boxes, the distance, and the aspect ratio. The formula is as follows:

[0049]

[0050]

[0051]

[0052] Where IOU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box, ρ 2 (b,b gt ) is the Euclidean distance between the center points of the two boxes; w gt and h gt is the length and width of the truth box, and w and h are the length and width of the prediction box.

[0053] The centerness branch is used to measure the quality of sample points, calculated using cross-entropy loss. This invention proposes a new method for calculating the centerness branch: using IOU and classification score to evaluate the quality of the predicted bounding box. Specifically: for all sample points within the ground truth bounding box, the IOU between the predicted and ground truth bounding boxes is calculated, and the IOU value is combined with the predicted class confidence score as a whole. The label value of the centerness branch has two cases: if it is within the positive sample sampling area set above, the label value is 1; if it is within the negative sample sampling area, the label value is 0. For a high-quality sample point, its predicted bounding box IOU and confidence score are both high, while for low-quality sample points, either the predicted bounding box IOU or the class confidence score may be low, so these boxes need to be filtered out. The formula for calculating the centerness branch is:

[0054]

[0055] Where IOU is the overlap area between the predicted bounding box and the ground truth bounding box, and score is the confidence score of the predicted bounding box.

[0056] Between the first and second stages, the ROIAlign module is used to select candidate boxes from the feature map output by the candidate box extraction network. The steps of ROI Align are as follows:

[0057] a) First, a series of candidate boxes are obtained from the candidate box extraction network and then input into the feature maps of each layer of the feature fusion network;

[0058] b) Calculate the square root of the area of ​​each candidate box, using the formula:

[0059] c) According to the formula Calculate the level corresponding to the candidate box, where 5 is a fixed parameter;

[0060] d) After finding the level corresponding to the candidate box, use the bilinear interpolation algorithm to calculate the feature map corresponding to the candidate box;

[0061] In the second stage, regarding the selection of the number of candidate boxes used in the first stage: during training, the top 30 candidate boxes with an IOU threshold of 0.7 and a score threshold of 0 are selected for training to ensure the recall and convergence speed of the model; during the inference stage, the top 15 candidate boxes with an IOU threshold of 0.7 and a score threshold of 0 are also used for prediction.

[0062] In the Refine module, residual connections and separable group convolutions are used. The number of separable group convolutions is set to 16, and separable group convolutions are added to the residual connections. Finally, convolutions are used for prediction.

[0063] (8) When performing post-processing, a score threshold of 0.4 and an IOU threshold of 0.5 are selected. The post-processing steps are as follows:

[0064] a) Extract each prediction from the Refine module, including predictions for classification and regression.

[0065] b) After processing the predicted classification values ​​with the sigmoid function, combine them with the confidence scores of the candidate boxes from the first stage to obtain the final output classification confidence score.

[0066] c) Combine the predicted location information with the candidate boxes from the first stage to calculate the accurate location information.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. An improved FCOS algorithm-based unmanned vehicle-mounted camera target detection method, the specific steps being as follows, characterized in that: (1) the camera is installed at the front end of the vehicle to facilitate the camera to capture the front side of the vehicle; during the driving of the vehicle, the camera collects video during the driving of the vehicle; (2) the key frame of the collected video stream information is intercepted, the intercepted image is preprocessed, the image preprocessing includes screening of repeated images, deleting images without targets, and further including labeling all images, the labeling file containing class information and target position information of the image; the preprocessed image is divided into a pretraining image dataset for training, a verification image dataset for verification, and a test image dataset for testing; (3) an improved FCOS model is built, the training parameters of the model are set, the preprocessed image in step (2) is used for parameter setting of the improved FCOS model, the training of the model is performed, during the training process, the images in the training set are extracted, after multiple iterations, the trained detection model is obtained; the improved FCOS network training model in step (3) is built in a Pytorch environment, the training set, verification set and test set paths obtained in step (2) are added to the model configuration file, and corresponding parameters are set in the improved FCOS network training model; a channel space attention mechanism module is added after feature maps of different sizes, so that the feature maps are weighted and averaged in the channel and space directions, and the precision of the model is improved; a VargNet lightweight backbone network is used to reduce the calculation amount of the model, a multi-branch structure is proposed on the basis of the original model to improve the feature representation capability of the model, and a re-parameter method is used to reduce the calculation amount of the model during inference; a BiFPN structure is added in the feature fusion module; the convolution module in the original FCOS detection module is replaced with a depth separable convolution module, and a proposed sampling method based on center prior and a new Centerness calculation method are added in the FCOS sampling stage; wherein the calculation formula of the center degree branch is: ； wherein the iou is the overlapping area between the prediction box and the real box, and the score is the confidence score of the prediction box; a refine module is added after the FCOS detection module to further improve the precision of the model; the improved FCOS network model includes a backbone feature extraction network, a feature fusion module, an FCOS Proposal module and an FCOS Refine module. The backbone feature extraction network uses a VargNet network, which uses a variable grouping convolution to replace a traditional convolution, and adds a multi-branch structure and a reparameterization structure, the variable grouping convolution is similar to a depth separable convolution, and is composed of two parts, the first part is a variable group convolution part, the variable grouping convolution does not adopt a fixed grouping form, and is more conducive to the operation of the underlying hardware, the second part is a point-by-point convolution part, which is used to ensure the matching of the number of channels, the multi-branch structure is composed of multiple 3*3 convolutions, which improves the feature representation capability of the model, during model inference, the reparameterization adjusts the model to a single path structure, while maintaining the performance of the model, the resource consumption of the model is reduced; The feature fusion module uses a BiFPN structure, which deletes nodes with only one input, simplifies the structure operation amount; and increases the connection channels between the input and the output, improves the feature fusion performance of the model; the module simultaneously uses two fusion methods from top to bottom and from bottom to top, simultaneously increases the semantic information of the top layer and the positioning information of the bottom layer, greatly improves the feature fusion capability of the model, and the module also considers that the contribution of different size feature maps to the model is different, so different weights are given to feature maps of different sizes; The FCOS Refine module takes the detection head of FCOS as a candidate frame extraction network in the first stage, adds a center prior rule when sampling positive and negative samples of FCOS, sets a square sampling area with a certain side length according to the position of the center of the real frame, and only samples positive samples in the sampling area, sets a scaling coefficient with a side length to control the center sampling area for targets of different scales, and uses a new center degree branch to evaluate the quality of the samples, and simultaneously proposes a Refine module for feature extraction and training in the second stage; between the first stage and the second stage, the ROI Align module is used to select the candidate frame from the feature map, the module does not have a quantization operation in the coordinate mapping process, and a bilinear interpolation algorithm is used to calculate the value of the target position, which obviously improves the positioning accuracy; (4) check the effect of the model, analyze the last saved model, and check whether the model has overfitting or underfitting phenomenon, if the overfitting or underfitting phenomenon occurs, adjust the training times and hyperparameters according to the experimental results to obtain a reasonable detection model; (5) set a score threshold and an IOU threshold, decode the output results of the model, perform non-maximum suppression processing on the decoded prediction frame according to the set threshold, delete the redundant target frame, and finally output the target detection result; In the post-processing stage, the confidence of the first stage candidate frame and the classification confidence predicted by the second stage are combined as the final output classification confidence.

2. The unmanned vehicle camera target detection method based on the improved FCOS algorithm according to claim 1, wherein the specific way of key frame extraction in step (2) is that the video image data collected in step (1) is extracted as a key frame every 10 frames. ​ 3. The unmanned vehicle-mounted camera target detection method based on the improved FCOS algorithm according to claim 1, characterized in that: The specific operation mode of the image preprocessing in step (2) is: First, all the extracted images are sorted, and the repeated, invalid and irrelevant images are deleted; secondly, the remaining images after the selection operation are sorted, and are divided into three parts according to the ratio of 7:2:1, which are respectively used as the training set for model training, the verification set for model verification and the test set for model testing; finally, the label processing tool is used to label and process the labels of the detection target of the training set and the verification set.

Images