Image recognition model training method, system, equipment and medium
By performing multi-scale processing and prediction on images, the training process of object recognition model is simplified, the problems of object recognition complexity and slow training speed in the prior art are solved, and efficient object detection is achieved.
Patent Information
- Application Number
- CN202410965576.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-12
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, object recognition in the image requires the pre-finding of several areas where objects may exist, resulting in complex processing flow, slower model training speed and greater difficulty.
Using an image recognition model training method, multi-scale processing of multiple images to be trained can obtain multi-scale feature maps, and obtain the prediction recognition box and prediction category of the target object based on the prediction vector, so as to identify the network in the preset model. Adjust the parameters to simplify the training process and reduce the difficulty of model training.
There is no need to find the target object area in advance. Just make a prediction of each grid once to obtain the prediction recognition box and category of the target object, which simplifies the model training process, improves detection speed and accuracy, and reduces the misjudgment rate.
Smart Images

Figure CN118941890A_ABST
Abstract
Description
Background Art
[0002] Urban governance faces a series of sudden abnormal events involving social security, production and living safety, traffic safety, etc. For example, abnormal events include crowd gathering, road occupation, fighting, pushing, vehicle road occupation, traffic congestion, and garbage dumping. Timely discovery and handling of these abnormal events can reduce the loss of life and property, improve the efficiency of urban operation, maintain the stable development of the city, and improve the level of urban governance. With the rapid economic development, the rapid expansion of urban space, the difficulty of urban governance has soared, and abnormal events have overwhelmed managers. Modern urban governance urgently needs to monitor and respond to abnormal events in a comprehensive, rapid, and reliable manner to avoid serious economic losses and adverse social impacts caused by improper handling and delays of abnormal events.
[0003] Computer vision target (object) detection is to find certain specific targets that trigger abnormal events in an image of video images taken by a camera, identify the types of these targets, and mark the locations of these targets. The target objects can be vehicles or people, etc.
[0004] Since the position of the target is uncertain, there is no way to guarantee that the target is in the center of the image, the size of the target is also uncertain, and the number and types of targets cannot be guaranteed. Therefore, target detection has always been a difficult problem.
[0005] The simplest method is the sliding window method, which is to design windows of different sizes, let the windows slide at the smallest step length, and put all the images in the window into the classifier for classification. That is, use the sliding window to identify an object. If an object happens to appear in a sliding window, it can be detected, and the position of the sliding window is the position of the detected object. However, this sliding window method may end up with hundreds of thousands of windows, which is too time-consuming and inefficient.
[0006] The typical method is the region-convolutional neural network (R-CNN) method, which first scans and analyzes the image through machine learning and deep learning methods to find several regions where objects may exist (region proposal), and then removes duplicate regions through post-processing to obtain the target region. Finally, these target regions are cropped and put into an image classifier for classification. This method is also called the two-stage method. However, the existing technology of the region-convolutional neural network method needs to find several regions where objects may exist in advance to identify objects in the image, and the processing flow is relatively complicated, resulting in slow model training and greater difficulty. Summary of the invention
[0007] In order to overcome the problem in the prior art that, in order to identify objects in images, it is necessary to find in advance several areas where objects may exist, and the processing flow is relatively complicated, resulting in slow and difficult model training, the present application provides an image recognition model training method, system, device and medium.
[0008] In the first aspect, in order to solve the above technical problems, the present application provides an image recognition model training method, comprising:
[0009] Acquire a plurality of images to be trained, where the images to be trained are grid images including a plurality of grids of the same size;
[0010] Input a plurality of images to be trained into a preset recognition model, perform multi-scale processing on the images to be trained, and obtain corresponding multi-scale feature maps;
[0011] Perform prediction processing on the multi-scale feature map to obtain the prediction vector corresponding to each grid in the image to be trained;
[0012] Based on multiple prediction vectors corresponding to the image to be trained, a predicted recognition box and a predicted category of the target object in the image to be trained are obtained;
[0013] Based on the predicted recognition boxes and predicted categories corresponding to a plurality of images to be trained, the network parameters in the preset recognition model are adjusted to obtain a target recognition model.
[0014] In a second aspect, the present application also provides an image recognition model training system, comprising:
[0015] An acquisition module is used to acquire a plurality of images to be trained, where the images to be trained are grid images including a plurality of grids of the same size;
[0016] A processing module, used to input a plurality of images to be trained into a preset recognition model, perform multi-scale processing on the images to be trained, and obtain corresponding multi-scale feature maps;
[0017] The first prediction module is used to perform prediction processing on the multi-scale feature map to obtain a prediction vector corresponding to each grid in the image to be trained;
[0018] A second prediction module is used to obtain a predicted recognition box and a predicted category of a target object in the image to be trained based on a plurality of prediction vectors corresponding to the image to be trained;
[0019] The parameter adjustment module is used to adjust the network parameters in the preset recognition model based on the predicted recognition boxes and predicted categories corresponding to multiple images to be trained, so as to obtain the target recognition model.
[0020] In a third aspect, the present application also provides a computing device, including a memory, a processor, and a program stored in the memory and running on the processor, and when the processor executes the program, the steps of an image recognition model training method as described above are implemented.
[0021] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores instructions. When the instructions are executed on a terminal device, the terminal device executes the steps of an image recognition model training method.
[0022] The beneficial effect of the present application is: by performing multi-scale processing on the to-be-trained image divided into multiple grids of the same size in the preset recognition model, and predicting the prediction vector corresponding to each grid in the to-be-trained image, the predicted recognition box and predicted category of the target object obtained based on the multiple prediction vectors corresponding to each to-be-trained image are used to adjust the network parameters in the preset model to obtain the target recognition model. In this way, during the model training process, there is no need to find multiple areas where the target object may exist in the to-be-trained image in advance. It is only necessary to perform a prediction once for each grid in the to-be-trained image to obtain the predicted recognition box and predicted category of the target object. Therefore, the process of processing the to-be-trained image during training is simple, and the processing process of each grid in the to-be-trained image is the same, which reduces the difficulty of model training. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 A flowchart of an image recognition model training method for this application;
[0024] Figure 2 A network structure diagram of the preset recognition model for this application;
[0025] Figure 3 This is the network structure diagram of the CBL module in the backbone network of the target recognition model of this application;
[0026] Figure 4 This is the network structure diagram of the CSPX_Y module in the backbone network of the target recognition model of this application;
[0027] Figure 5 This is a network structure diagram of the residual module in the backbone network of the target recognition model of this application;
[0028] Figure 6 This is the network structure diagram of the SPPF module in the backbone network of the target recognition model of this application;
[0029] Figure 7 This is a schematic diagram of the workflow of the SPPF module in the backbone network of the target recognition model of this application;
[0030] Figure 8Another flowchart of an image recognition model training method for this application;
[0031] Fig. 9 A business process diagram for applying the target recognition model for this application;
[0032] Fig.10 This is a recognition flow chart of the target recognition model of this application;
[0033] Fig.11 Another flowchart of an image recognition model training method for this application;
[0034] Fig.12 This is a structural diagram of an image recognition model training system for this application. DETAILED DESCRIPTION
[0035] The following examples are provided to further explain and supplement the present application and do not constitute any limitation to the present application.
[0036] The following describes an image recognition model training method, system, device and medium according to an embodiment of the present application in conjunction with the accompanying drawings.
[0037] An image recognition model training method in an embodiment of the present application is applied to a terminal device. In the present application scheme, the terminal device is used as the execution subject to illustrate the present application scheme. The terminal device is used to execute the steps of an image recognition model training method.
[0038] like Figure 1 As shown, the present application provides an image recognition model training method, comprising:
[0039] Step S1, obtaining a plurality of images to be trained, where the images to be trained are grid images including a plurality of grids of the same size;
[0040] Step S2, inputting a plurality of images to be trained into a preset recognition model, performing multi-scale processing on the images to be trained, and obtaining corresponding multi-scale feature maps;
[0041] Step S3, performing prediction processing on the multi-scale feature map to obtain a prediction vector corresponding to each grid in the image to be trained;
[0042] Step S4, obtaining a predicted recognition box and a predicted category of the target object in the image to be trained based on multiple prediction vectors corresponding to the image to be trained;
[0043] Step S5, based on the predicted recognition boxes and predicted categories respectively corresponding to the multiple images to be trained, the network parameters in the preset recognition model are adjusted to obtain the target recognition model.
[0044] In the present embodiment, an image recognition model training method is provided, which performs multi-scale processing on a to-be-trained image divided into a plurality of grids of the same size in a preset recognition model, and predicts a prediction vector corresponding to each grid in the to-be-trained image, and adjusts the network parameters in the preset model based on the predicted recognition box and predicted category of the target object obtained by the plurality of prediction vectors corresponding to each to-be-trained image, thereby obtaining a target recognition model. In this way, during the model training process, it is not necessary to find in advance the plurality of regions where the target object may exist in the to-be-trained image, and it is only necessary to perform a prediction once for each grid in the to-be-trained image, so that the predicted recognition box and predicted category of the target object can be obtained. Therefore, the process of processing the to-be-trained image during training is simple, and the processing process of each grid in the to-be-trained image is the same, which reduces the difficulty of model training.
[0045] When the target recognition model obtained by the training method of the present application is used to identify the target object in the image, it only needs to browse the image once (you only look once, YOLO) to predict the predicted recognition box and predicted category of the target object in the image. Therefore, the training method of the present application can be called a model training method based on YOLO, which converts the target detection problem into a single regression problem of directly extracting the recognition box (Bounding Boxes) and the object category probability from the image. It only needs to browse the image once to identify the category and position of the object in the image. The training and reasoning process is based on the entire image, and can "see" the overall information of the entire image, which can effectively solve the problems of slow target detection speed and high misjudgment rate of background images. Compared with the R-CNN method, this method does not need to find the target area (Region proposal) where the object may exist in advance, and only requires one "scan" (recognition), which is also called a single-stage (1-stage) method.
[0046] In some embodiments, the prediction vector of the present application needs to be post-processed to obtain the final detection result, and the post-processing is used to remove redundant recognition boxes (Bounding Boxes) and only retain the most representative recognition box (BoundingBox) bounding box.
[0047] Optionally, multiple images to be trained are obtained, including:
[0048] Acquire multiple collected images related to abnormal events in urban governance; abnormal events include crowd gathering, road occupation, fighting, pushing, road occupation, traffic congestion, and garbage dumping;
[0049] Preprocessing the collected image to obtain a corresponding processed image; wherein the preprocessing includes defogging processing, and / or deraining processing, and / or low-light enhancement processing, and / or super-resolution reconstruction processing;
[0050] The processed image is segmented to obtain a to-be-trained image having a plurality of grids of the same size.
[0051] In this embodiment, multiple collected images related to abnormal events such as gathering of people, occupying the road for business, fighting, pushing, vehicles occupying the road, traffic jams, and garbage dumping in urban governance are obtained in a targeted manner, so as to improve the event targeting of the images to be trained for subsequent training. At the same time, the collected images are pre-processed by defogging, and / or deraining, and / or low-light enhancement, and / or super-resolution reconstruction to improve the resolution and image quality of the processed images obtained, thereby improving the event accuracy of the images to be trained for subsequent training. Model training is performed using images to be trained with strong event targeting and high time accuracy, reducing the interference of useless images and images with poor accuracy on training, thereby improving the efficiency and accuracy of model training.
[0052] In some embodiments, focusing on the detection service problem of typical abnormal events in urban governance, digital twin technology is used to digitize more than 200 communities and streets in a province, and the collection of relevant images and video data in all time and domain is realized, and multiple collected images are obtained and stored in a distributed storage system. In order to solve the image degradation and poor image quality problems existing in the collected images collected by video surveillance equipment under adverse imaging conditions such as fog, haze, dust, rain, snow, and low light, deep neural networks are used to pre-process the collected image data to achieve image defogging, deraining, low-light enhancement, super-resolution reconstruction, etc.
[0053] Optionally, multi-scale processing is performed on the training image to obtain a corresponding multi-scale feature map, including:
[0054] The backbone network of the preset recognition model is used to extract features from the training image to obtain feature maps of multiple different scales;
[0055] A fusion operation is performed on multiple feature maps of different scales to obtain a multi-scale feature map corresponding to the image to be trained; wherein the backbone network includes a first feature extraction module, a second feature extraction module, a pooling module and a fusion module, and the fusion operation includes upsampling processing and splicing processing.
[0056] In this embodiment, a backbone network of a preset recognition model is used to extract feature maps of multiple scales in the image to be trained, and features in the image to be trained are repeatedly extracted at different depths. The feature maps of multiple scales are fused to obtain a multi-scale feature map, so that the multi-scale feature map contains global features of the image to be trained, and the number and amount of features in the global features can meet the recognition accuracy requirements of the target object, thereby improving the accuracy of the predicted recognition box and predicted category obtained by subsequent prediction of the multi-scale feature map.
[0057] Figure 2 The network structure diagram of the preset recognition model for this application is as follows: Figure 2 As shown, the preset recognition model includes an input end, a backbone network and a model prediction module. The input end is used to receive multiple images to be trained in the input model, and then the multiple images to be trained are input into the backbone network for multi-scale processing to obtain corresponding multiple multi-scale feature maps, and then the multiple multi-scale feature maps are input into the model prediction module for prediction to obtain the predicted recognition box and predicted category of the target object in each image to be trained. The predicted recognition box and predicted category of the target object in the image to be trained obtained by the preset recognition model are determined based on the target recognition box and target category, and sent to the terminal device display interface such as the information system, portal website and mobile phone APP of the relevant department, so as to realize the display and early warning of abnormal events.
[0058] The input end includes a data expansion module and a data normalization module. The data expansion module uses a mosaic strategy to expand and enhance the input data to alleviate the problem of the lack of number and types of abnormal event samples in actual scenarios and meet the data requirements for deep learning model training. This strategy expands the labeled data set by rotating, scaling, splicing, cropping, and adaptively adjusting anchor frames on the labeled data. The data normalization module adaptively scales the input data according to the model requirements to solve the problem of inconsistent raw data sizes on different devices, and normalizes the width (w) and height (h) of the input image to the range of 0-1.
[0059] The backbone network includes multiple first feature extraction modules, multiple second feature extraction modules, pooling modules and multiple fusion modules. Among them, the first feature extraction module is a CBL (Conv+BN+LeakyReLU) module, the second feature extraction module is a CSPX_Y (CSP Bottleneck with convolutions) module, the pooling module is an SPPF module (Spatial Pyramid Pooling Fast), and the fusion module is a Concat module. As the image to be trained is gradually extracted in the backbone network, the size of the feature map obtained during feature extraction will gradually decrease, but the number of channels will gradually increase to capture more feature information.
[0060] Figure 3 This is the network structure diagram of the CBL module in the backbone network of the target recognition model of this application, as shown Figure 3 As shown in the figure, the CBL module consists of adjacent 2D convolutional layers (Convolutional, Conv) and batch normalization layers (Batch Normalization, BN), as well as activation functions (Leaky Rectified Linear Units, LeakyReLU, corrected linear function). Conv uses multiple different convolution kernels to process the input image to obtain different feature maps. BN shuffles the training data set to prevent training from shifting and effectively avoid overfitting. At the same time, a large learning rate can be used without careful parameter adjustment. A larger learning rate can greatly increase the learning speed.
[0061] The activation formula of the activation function (LeakyReLU) is as follows:
[0062]
[0063] Among them, φ(x) is the activation result, x is the current neuron input, and when x is greater than 0, there is no saturation problem, the calculation efficiency is high, and the network can converge quickly. Adding parameter 0.1 to the part where x is less than or equal to 0 solves the problem that some neuron inputs will fall into the hard saturation area during training, resulting in the inability to update the weights and affecting the network convergence.
[0064] Figure 4 This is the network structure diagram of the CSPX_Y module in the backbone network of the target recognition model of this application, as shown Figure 4 As shown in Figure 1, the CSPX_Y module consists of three CBL modules, one residual module (Res unit) and a concatenation (Concat) module. The residual module (Res unit) is used to allow the network to converge even after the depth of the network construction increases. Figure 5As shown, the residual module consists of 2 CBL modules and an add module.
[0065] The CSPX_Y module first divides the feature map of the base layer into two parts: a direct path and a complex path, and then concatenates them through a cross-stage hierarchy. In the direct path, a part of the feature map passes directly through a CBL module. This path maintains the information of the original feature map without introducing additional computational burden. In the complex path, the other part of the feature map passes through two CBL modules and a Res Unit module. The Res Unit module retains the details and semantic information in the original input data by introducing residual connections, making it easier for information to propagate to subsequent layers and avoiding information loss. This design helps solve the problems of gradient disappearance and model degradation in deep networks, allowing the network to train deeper layers more effectively. The combination of the CBL module and the Res Unit module allows the network to reuse and combine features at multiple scales. By processing features at different levels, the network can learn richer hierarchical representations, which helps improve the generalization and robustness of the model.
[0066] In traditional convolutional neural networks, gradient information is repeatedly transmitted throughout the network, while the CSPX_Y module achieves the diversion of gradient information by splitting and merging feature maps. Before the merging point, the feature map of each branch is independently convolved and generates independent gradient information, which is fused at the merging point. The CSPX_Y module can reduce the amount of calculation while ensuring accuracy, effectively reduce the repeated calculation of gradient information, reduce memory requirements, and enhance the representation ability of the backbone network. Compared with the traditional Cross Stage Partial Network (CSPNet), the CSPX_Y module has the characteristics of simple structure, fast calculation, and easy fusion.
[0067] The SPPF module is placed at the end of the backbone network to extract multi-scale features. Figure 6As shown in the figure, the SPPF module consists of three maximum pooling modules (MaxPool) and a concatenation (Concat) module. The diversity of the sizes of the input training images makes it easier for the network to converge (convergence) than the training images of a single size. The SPPF module uses the same image of different scales (scale) as input to obtain pooling features of the same length. By multi-scale pooling of feature maps, semantic information at different scales is captured, effectively improving the prediction accuracy of the model. Pooling (pooling), that is, based on the idea of local correlation, obtains new element values by sampling or aggregating information from a set of locally related elements. Max pooling (Max pooling) selects the largest element value from the set of locally related elements.
[0068] Figure 7 The following is a schematic diagram of the workflow of the SPPF module in the backbone network of the target recognition model of this application. Figure 7 As shown in the figure, an image is divided into 4×4 non-overlapping blocks of the same size (pooling size). Max pooling is performed in each block, only the largest number (max) is taken, and other nodes are discarded, keeping the original plane structure output as 2×2, which will not damage the recognition result.
[0069] In some embodiments, when the preset recognition model of the present application is trained, the input end is used to receive multiple images to be trained in the input model, and then the multiple images to be trained are input into the backbone network, and the deeper features in the images to be trained are extracted by repeatedly using the CBL module and the CSPX_Y module to obtain an intermediate multi-scale feature map. The obtained multi-scale feature map is pooled using the SPPF module, and the feature image is enlarged by upsampling so that the size of the enlarged feature map is consistent with the size of the corresponding image to be trained. Then, the processed intermediate multi-scale feature images are passed through multiple Concat modules and CSPX_Y modules to achieve the fusion of feature maps of different scales, and finally the multi-scale feature map is obtained by performing Concat operation after processing by three Conv convolutional layers. The accuracy of the feature information in the fused multi-scale feature map meets the set requirements, that is, the backbone network has a good feature representation capability. The model prediction module is then used to predict the multi-scale feature map, and the preset recognition model is adjusted according to the predicted recognition box and predicted category of the target object in the image to be trained to obtain a target recognition model. Finally, the trained target recognition model is used as a detection model for detecting target objects in abnormal event images. According to the feature information learned by the backbone network, the image input into the detection model is predicted to obtain the predicted recognition box (Bounding Box) and predicted category of the target object in the image. The recognition box information of the predicted recognition box includes the center position, width, height and recognition box confidence. According to the recognition box information of the predicted recognition box and its category, the abnormal event in the image can be located and identified.
[0070] In an exemplary embodiment provided by the present application, a plurality of CBL modules, a plurality of CSPX_Y modules, an SPPF module and a plurality of Concat modules included in the backbone network are used to perform multi-scale processing on the training image to obtain the corresponding multi-scale feature map, which is applied to the training of the target recognition model. Figure 8 , Figure 8 Another flowchart of an image recognition model training method for this application is shown in FIG. Figure 8As shown, data is collected to obtain multiple collected images, which are stored in a distributed file system, the collected collected images are preprocessed, and the predicted processed data are annotated. The annotated data set is input into a preset recognition model. When training in the backbone network, the image features of different scales of the image to be trained are extracted by the CBL module in the backbone network, and the CSPX_Y module in the backbone network reduces the amount of calculation and the number of parameters while maintaining the feature extraction capability of features of different scales, thereby improving the training speed and generalization capability of the model. When training the backbone network, the CBL module and the CSPX_Y module are repeatedly used to extract deeper image features in the image to be trained. As the number of feature extractions increases, the number of channels in the obtained feature map increases, and the image size decreases. The SPPF module is used to pool the multi-scale feature map to capture semantic information at different scales, and the pooled feature map is upsampled to enlarge the feature image, so that the size of the enlarged feature map is consistent with the size of the corresponding image to be trained. Then, the processed intermediate multi-scale feature images are passed through multiple Concat modules, CSPX_Y modules, and Conv to achieve splicing and fusion of feature maps of different scales to obtain a multi-scale feature map. Based on the multi-scale feature map, determine whether the preset recognition model meets the accuracy requirements. If the accuracy requirements are not met, adjust the network parameters corresponding to the backbone network in the preset recognition model, retrain the preset recognition model until the preset recognition model meets the accuracy requirements, and use the preset recognition model that meets the accuracy requirements as the target recognition model (detection model); if the accuracy requirements are met, use the preset recognition model at this time as the target recognition model (detection model).
[0071] In some embodiments, Fig. 9 As shown, when the target recognition model of the present application is applied, the collected images and / or collected videos obtained by real-time video monitoring with an intelligent camera are transmitted to the data processing and analysis center for data preprocessing and segmentation processing to obtain the image to be tested, and the image to be tested is input into the target recognition model to obtain the target recognition frame with the location information of the target object and the target category of the target object in the image to be tested, and abnormal events are identified based on the target recognition frame and the target category to obtain the identification result, and the data containing the identification result is pushed to the terminal device (abnormal event analysis and visualization system) for system situation detection, intelligent positioning, real-time monitoring and analysis, cloud video display and storage. Among them, abnormal event identification includes personnel gathering identification, or, road occupation business identification, or, personnel pushing identification, or, vehicle road occupation identification, or, garbage pile identification, or, traffic congestion identification;
[0072] The image to be tested is input into the target recognition model, and the target recognition frame and target category are output; the recognition result includes the location information and type information of the abnormal event in the image to be tested.
[0073] Optionally, based on a plurality of prediction vectors corresponding to the image to be trained, a predicted recognition box and a predicted category of a target object in the image to be trained are obtained, including:
[0074] Identify the prediction vector and obtain the prediction sub-recognition box of the target object in the corresponding grid;
[0075] Determine the predicted probability that the target object in the grid is of a preset category;
[0076] Based on each predicted sub-recognition box in the image to be trained and the corresponding prediction probability, the predicted recognition box and predicted category of the target object in the image to be trained are obtained.
[0077] In this embodiment, the prediction vector contains image information of the location of the corresponding grid. Based on the image information, the location and possible categories of the target object in the grid can be understood. The prediction vector is identified to obtain the prediction sub-recognition frame of the corresponding grid, and the prediction probability of the target object being a preset category is calculated based on the possible categories. This facilitates the prediction recognition frame and prediction category of the target object in the image to be trained to be obtained with high accuracy based on each prediction sub-recognition frame and the corresponding prediction probability in the image to be trained.
[0078] Optionally, the recognition box information of the predicted sub-recognition box includes a center position, a width, a height, and a recognition box confidence;
[0079] Based on each predicted sub-recognition box and the corresponding prediction probability in the image to be trained, the predicted recognition box and predicted category of the target object in the image to be trained are obtained, including:
[0080] Based on the recognition box confidence and prediction probability of the predicted sub-recognition box, the category probability of the target object category appearing in the predicted sub-recognition box is calculated;
[0081] The box composed of multiple predicted sub-recognition boxes with a category probability greater than or equal to a set value and the same preset category is used as the predicted recognition box;
[0082] The preset category corresponding to the predicted recognition box is used as the predicted category.
[0083] In this embodiment, based on the recognition frame confidence and prediction probability of the predicted sub-recognition frame, the category probability of the target object category appearing in the predicted sub-recognition frame can be understood. When the category probability is greater than or equal to the set value, it means that the probability that the target object corresponding to the category probability is the corresponding category is the largest, and the frame composed of multiple predicted sub-recognition frames with category probabilities greater than or equal to the set value and the same preset category is used as the predicted recognition frame, and the preset category corresponding to the predicted recognition frame is used as the predicted category, which can improve the accuracy of the predicted recognition frame and the predicted category of the target object in the training image.
[0084] In some embodiments, first, the image to be trained is segmented to obtain images with s 2 A grid image with grids of equal size. When detecting the target object in the training image, only the part of the target object that falls in the grid is required to detect the target object, rather than the entire target object, thereby improving the detection speed.
[0085] Then, each grid predicts B bounding boxes. The bounding box information of each bounding box has 5 values, which are the center position (x, y) of the bounding box associated with the grid, the width (w) and height (h) of the bounding box associated with the entire image frame (image), and the confidence score of the bounding box predicted this time. The confidence score of the bounding box represents the confidence that the bounding box contains a target object. The confidence score of the bounding box is defined as Pr(Object)*IoU (Intersection-over-Union, that is, the ratio of the intersection and union of the predicted bounding box and the pre-annotated window). If there is no target object, the confidence score is zero.
[0086] Finally, the probability that the target object category appears in the recognition box (Bounding Box) is calculated, that is, the predicted fitness between the recognition box (Bounding Box) and the target object.
[0087] Each grid predicts the category of the target object, and the bounding box in the grid does not require the object category.
[0088] Each grid predicts only one type of target object, and each grid predicts a conditional probability value C(Pr(Classi|Object)). The probability value C represents the probability that a grid contains a target category (object category).
[0089] Each Bounding Box multiplies the category probability of the grid by the confidence of the Bounding Box to obtain the confidence score Pr(Classi)*IoU of a specific target (object) category: Pr(Classi|Object)*Pr(Object)*IoU=Pr(Classi)*IoU;
[0090] This confidence score represents the category probability of the target object appearing in the recognition box (Bounding Box), that is, the predicted fit between the recognition box (Bounding Box) and the target object.
[0091] Fig.10 This is the recognition flow chart of the target recognition model of this application, such as Fig.10 As shown, the image to be tested is segmented to obtain 2The grid on input operation is performed on the image to be tested, and the image to be tested is input into the target recognition model for recognition, and the number of predicted sub-recognition boxes (Bounding Boxes) of the target object in each grid (grid) is obtained. The classifier is used to identify C different categories of objects in the image to be tested, including dogs, bicycles, cars, etc., then the output is a Tensor of S×S(B×5+C), so that the model can accurately classify the objects in the image. Then, a Bounding boxes + confidence (bounding box recognition + confidence calculation) operation is performed on the image to be tested, and the predicted probabilities that the target objects in each grid of the image to be tested are dogs, bicycles, cars, etc. are determined, and the Class probability map (class probability map) corresponding to the image to be tested is obtained; when Final detections are performed on the image to be tested, based on each predicted sub-recognition box in the image to be tested and the corresponding predicted probability, the predicted recognition box and predicted category of the target object in the image to be tested are obtained, and based on the recognition box confidence and predicted probability of the predicted sub-recognition box, the category probability of the categories of dogs, bicycles, cars, etc. appearing in the predicted sub-recognition box is calculated; a box composed of multiple predicted sub-recognition boxes with a category probability greater than or equal to a set value and the same preset category is used as a predicted recognition box to obtain a dog recognition box, a bicycle recognition box, a car recognition box, etc.; the preset category corresponding to the predicted recognition box is used as the predicted category, and a test image having a dog recognition box, a bicycle recognition box and a car recognition box, and each recognition box has a category identifier is obtained. The preset recognition model of the present application can identify key features in the image to be tested, which facilitates the algorithm to be trained by learning the mapping from the input image to be trained to the corresponding annotations, thereby reducing the loss function (loss) error and improving the model quality of the obtained target recognition model.
[0092] Optionally, based on the predicted recognition boxes and predicted categories respectively corresponding to the multiple images to be trained, the network parameters in the preset recognition model are adjusted to obtain the target recognition model, including:
[0093] Determine the box regression loss between the predicted box and the true box corresponding to each image to be trained;
[0094] Determine the binary cross entropy loss for the predicted category for each training image;
[0095] Based on multiple recognition box regression losses and multiple binary cross entropy losses, the network parameters in the preset recognition model are adjusted to obtain the target recognition model.
[0096] In this embodiment, the recognition box regression loss between the predicted recognition box and the real recognition box corresponding to each image to be trained and the binary cross entropy loss of the predicted category corresponding to each image to be trained are used to adjust the network parameters in the preset recognition model, so that the network parameters in the adjusted model can meet the prediction requirements. The model that meets the prediction requirements is used as the target recognition model, which can ensure the prediction accuracy of the target recognition model.
[0097] In some embodiments, in order to improve the prediction accuracy of the model, the recognition box regression loss (Regression Loss) and the recognition box classification loss (Classification Loss) are used in model training to quantitatively describe the position and category errors of the prediction results.
[0098] The loss function (DIoU Loss) used for the recognition box regression loss is:
[0099]
[0100] Among them, L DIoU represents the box regression loss between the predicted box and the corresponding true box, b represents the center position of the predicted box, and b gt represents the center position of the real recognition box corresponding to the predicted recognition box, ρ represents the Euclidean distance between the predicted recognition box and the corresponding real recognition box, c represents the diagonal distance of the closure area of the predicted recognition box and the corresponding real recognition box, and IoU represents the intersection over union ratio between the predicted recognition box and the corresponding real recognition box. This recognition box regression loss function optimizes the goal of reducing the Euclidean distance between the center points of the two rectangular boxes.
[0101] The loss function used for the recognition box classification loss is the binary cross entropy loss (BCE Loss):
[0102]
[0103] Among them, L BCE represents the binary cross entropy loss of the predicted category corresponding to the predicted recognition box, i represents the sequence number of the predicted recognition box, N represents the total number of predicted recognition boxes, and y i represents the true label to which the i-th predicted recognition box belongs, p(y i ) indicates that the i-th predicted recognition box belongs to y i As a classification loss function, BCE Loss aims to improve the accuracy of candidate recognition box category prediction.
[0104] Optionally, the method further comprises:
[0105] Detection task requirements; task requirements include warning requirements for gathering of people, road occupation, fighting, pushing, vehicles occupying the road, traffic congestion, and garbage dumping;
[0106] When a task requirement is detected, a directed event data set is determined based on the task requirement;
[0107] The target recognition model is fine-tuned based on the directional event set, so that the fine-tuned target recognition model is applied to the recognition of the image to be tested corresponding to the task requirements.
[0108] In this embodiment, when a task requirement is detected, it means that the detection accuracy of the target recognition model needs to meet the recognition of the image to be tested corresponding to the task requirement. Then, the target recognition model will be fine-tuned based on the directional event data set determined by the task requirement, and the network parameters in the target recognition model will be further adjusted, so that each module of the backbone network will continuously optimize its weight parameters through the back propagation algorithm and the gradient descent method, thereby further improving the feature representation ability of the backbone network for such event detection tasks, thereby improving the recognition accuracy of the fine-tuned target recognition model when applied to the recognition of the image to be tested corresponding to the task requirement.
[0109] In an exemplary embodiment provided by the present application, after obtaining the trained target recognition model, the model prediction module in the target recognition model is fine-tuned to improve the accuracy of the target recognition model. Fig.11 , Fig.11 Another flowchart of an image recognition model training method for this application is shown in FIG. Fig.11 As shown, data is collected to obtain multiple collected images, which are stored in a distributed file system, and the collected images are preprocessed and segmented, and the real recognition frames and real categories of the target objects in the images to be trained obtained after processing are annotated. Then, the annotated image data set to be trained is input into the preset recognition model for model training. At the same time, a directional event data set is introduced to fine-tune the preset recognition model, and the model obtained by fine-tuning the verification data is used for prediction. If the accuracy of the predicted result does not meet the requirements, the network parameters of the preset recognition model are adjusted, and the preset recognition model is retrained until the preset recognition model meets the accuracy requirements; if the accuracy of the predicted result can meet the requirements, the model with the accuracy that can meet the requirements is used to predict the input image data to obtain the prediction result.
[0110] The target recognition model of this application is a computer vision target detection model, which is based on supervised learning. In supervised learning, the data is labeled, and the supervision information (labeling information) of each image to be trained is the N target objects it contains. Each target object has five pieces of information, namely the center position (x, y) of the target object and its height (h), width (w), and category. For the detection task of specific abnormal events (events with a higher degree of danger) in urban governance, the LabelImg open source tool is used to label the images to be detected, and a special abnormal event labeling dataset is constructed.
[0111] The present invention provides an image recognition model training method that solves target object detection as a regression problem. Based on a single end-to-end network, the system completes the output from the input of the original image to the predicted recognition box of the object in the image (including the object position) and the corresponding predicted category.
[0112] According to the mode of "base model (preset model) + targeted fine-tuning (using targeted event data set for model fine-tuning)", a deep learning-based dynamic detection service method for abnormal events was designed, which realized the automatic detection of abnormal events in monitoring data and could automatically push abnormal event warning lists to relevant departments.
[0113] The target recognition model of the present application integrates the target region prediction and target category prediction into a single neural network model. The prediction process is simple, and fast target detection and recognition are achieved with high accuracy. At the same time, it has a certain robustness and is more suitable for real-time detection of field application environments. The target recognition model of the present application does not have an explicit process for obtaining the target region (Region roposal). The region-based object detection method (R-CNN method) uses a separate module to obtain the candidate box (a rectangular area that may contain an object), so the training process is also divided into multiple modules. The separated module is a selective search (target detection) method independent of the R-CNN network.
[0114] The training method of the present application uses the whole image information for prediction during the training and reasoning process, and can "see" the overall information of the whole image. However, the region-based object detection method (R-CNN method) only "sees" the local image information in the candidate frame during the detection process. If part of the data in the image background (non-object) (such as patches in the background) is included in the candidate frame and sent to the detection network for detection, it is easy to be misdetected as an object because the global image cannot be seen during the detection. The target recognition model obtained by the training method of the present application extracts the global information (overall information) in the image to be tested, and identifies the target object in each grid. According to the predicted sub-recognition frame of the target object in each grid and the predicted probability of the preset category, the predicted recognition frame and predicted category of each target object in the image to be tested are obtained. This method of obtaining the final result through multiple sub-results reduces the misjudgment rate of the image to be tested and improves the accuracy of the detection result.
[0115] The training method of the present application has a certain universality, and can learn the generalizable representation of the target, use natural images for training, and then use artistic images for prediction to train the target recognition model. The detection accuracy of the target object in the image using the target recognition model of the present application is higher than other methods. The target recognition model of the present application solves object detection as a regression problem. The entire detection process is simple, and the detection speed is much higher than other previous real-time target detection methods.
[0116] In some embodiments, an image recognition model training method of the present application is implemented by the following steps:
[0117] (1) Set up high-definition cameras throughout the city to collect relevant images and video data of the city at all times and in all areas, and store them in a distributed storage system;
[0118] (2) Use deep neural networks to pre-process the collected and stored image data, such as defogging, deraining, low-light enhancement, and super-resolution reconstruction; at the same time, perform pre-processing such as adaptive scaling and normalization on the collected and stored image data to solve the problem of inconsistent size of original image data collected by different devices;
[0119] (3) Use the LabelImg open source tool to label the preprocessed data and build a special abnormal event labeling dataset. At the same time, use the Mosaic strategy to rotate, scale, splice, crop, adaptively adjust the anchor frame, and normalize the labeled image data to expand the labeled dataset to obtain the training dataset, so as to alleviate the problem of lack of abnormal event samples in actual scenarios.
[0120] (4) Pre-training a preset recognition model using the data set to be trained to obtain a target recognition model;
[0121] (5) Aiming at the specific abnormal event detection task in urban governance, the target recognition model is fine-tuned using a dedicated data set to improve the model's feature representation capability, obtain a fine-tuned target recognition model, and improve the prediction accuracy of the target recognition model;
[0122] (6) Use the fine-tuned target recognition model to select and identify the captured images of the target object with urban abnormal events to obtain the predicted recognition box (Bounding Box) and predicted category of the target object in the image to be tested. The steps are as follows:
[0123] ① Segment the preprocessed image into 2 A grid image of equal-sized grids;
[0124] ②Perform multi-scale processing on the image to be tested in the backbone network to obtain the corresponding multi-scale feature map;
[0125] ③ In the model prediction module, the multi-scale feature map is predicted and processed to obtain the prediction vector corresponding to each grid;
[0126] ④ In the model prediction module, based on each prediction vector, predict the category probability C of each grid containing the preset object;
[0127] ⑤ In the model prediction module, B bounding boxes are predicted for each grid. Each bounding box has 5 pieces of information, namely the center position (x, y) of the bounding box related to the grid, the width (w) and height (h) of the bounding box related to the entire image frame (image), and the confidence score of the bounding box predicted this time;
[0128] ⑥ In the model prediction module, the category probability of each grid is multiplied by the confidence of the recognition box (Bounding Box) to obtain the confidence score (category probability) of the target object category in the grid, and the box composed of multiple prediction sub-recognition boxes with a category probability greater than or equal to the set value and the same preset category is used as the prediction recognition box, and the preset category corresponding to the prediction recognition box is used as the prediction category to identify the position coordinates of the target object in the image to be tested and the corresponding prediction category. Among them, the position coordinates are the recognition box information of the prediction recognition box, including the position information composed of the center position, width, and height of each prediction sub-recognition box.
[0129] (7) Use the location coordinates in the prediction identification box information and the corresponding prediction category to locate and identify the abnormal event type of the city, automatically push the abnormal event to the relevant departments, and issue targeted warnings.
[0130] like Fig.12 As shown, the present application provides an image recognition model training system, comprising:
[0131] An acquisition module is used to acquire a plurality of images to be trained, where the images to be trained are grid images including a plurality of grids of the same size;
[0132] A processing module, used to input a plurality of images to be trained into a preset recognition model, perform multi-scale processing on the images to be trained, and obtain corresponding multi-scale feature maps;
[0133] The first prediction module is used to perform prediction processing on the multi-scale feature map to obtain a prediction vector corresponding to each grid in the image to be trained;
[0134] A second prediction module is used to obtain a predicted recognition box and a predicted category of a target object in the image to be trained based on a plurality of prediction vectors corresponding to the image to be trained;
[0135] The parameter adjustment module is used to adjust the network parameters in the preset recognition model based on the predicted recognition boxes and predicted categories corresponding to multiple images to be trained, so as to obtain the target recognition model.
[0136] Optionally, obtain a module, specifically for:
[0137] Obtain multiple collected images related to abnormal events in urban governance;
[0138] Preprocessing the collected image to obtain a corresponding processed image;
[0139] The processed image is segmented to obtain a to-be-trained image having a plurality of grids of the same size.
[0140] Optionally, the processing module is specifically configured to:
[0141] The backbone network of the preset recognition model is used to extract features from the training image to obtain feature maps of multiple different scales;
[0142] A fusion operation is performed on multiple feature maps of different scales to obtain a multi-scale feature map corresponding to the image to be trained; wherein the backbone network includes a first feature extraction module, a second feature extraction module, a pooling module and a fusion module.
[0143] Optionally, the second prediction module is specifically used to:
[0144] Identify the prediction vector and obtain the prediction sub-recognition box of the target object in the corresponding grid;
[0145] Determine the predicted probability that the target object in the grid is of a preset category;
[0146] Based on each predicted sub-recognition box in the image to be trained and the corresponding prediction probability, the predicted recognition box and predicted category of the target object in the image to be trained are obtained.
[0147] Optionally, the recognition box information of the predicted sub-recognition box includes a center position, a width, a height, and a recognition box confidence; the second prediction module is specifically used to:
[0148] Based on the recognition box confidence and prediction probability of the predicted sub-recognition box, the category probability of the target object category appearing in the predicted sub-recognition box is calculated;
[0149] The box composed of multiple predicted sub-recognition boxes with a category probability greater than or equal to a set value and the same preset category is used as the predicted recognition box;
[0150] The preset category corresponding to the predicted recognition box is used as the predicted category.
[0151] Optionally, the parameter adjustment module is specifically used to:
[0152] Determine the box regression loss between the predicted box and the true box corresponding to each image to be trained;
[0153] Determine the binary cross entropy loss for the predicted category for each training image;
[0154] Based on multiple recognition box regression losses and multiple binary cross entropy losses, the network parameters in the preset recognition model are adjusted to obtain the target recognition model.
[0155] Optionally, the system further includes a fine-tuning module, the fine-tuning module being specifically used for:
[0156] Detection task requirements;
[0157] When a task requirement is detected, a directed event data set is determined based on the task requirement;
[0158] The target recognition model is fine-tuned based on the directional event set, so that the fine-tuned target recognition model is applied to the recognition of the image to be tested corresponding to the task requirements.
[0159] A computing device according to an embodiment of the present application includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, some or all steps of the above-mentioned image recognition model training method are implemented.
[0160] Among them, the computing device can be a computer, and correspondingly, its program is computer software. The above-mentioned parameters and steps in a computing device of the present application can refer to the parameters and steps in an embodiment of an image recognition model training method above, and will not be repeated here.
[0161] In an embodiment of the present application, a computer-readable storage medium is provided, in which instructions are stored. When the instructions are executed, the steps of the above-mentioned image recognition model training method are executed.
[0162] The computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.
[0163] The technical solution of the embodiment of the present disclosure can be embodied in the form of a software product, which is stored in a storage medium and includes one or more instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method of the embodiment of the present disclosure. The aforementioned computer-readable storage medium may be a non-transitory computer-readable storage medium, including: a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk or an optical disk, and other media that can store program codes, or a transient computer-readable storage medium.
[0164] The flowchart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present application. Wherein, each box in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
[0165] Those skilled in the art know that the present application can be implemented as a system, method or computer program product. Therefore, the present disclosure can be specifically implemented in the following forms, namely: it can be complete hardware, it can be complete software (including firmware, resident software, microcode, etc.), or it can be a combination of hardware and software, generally referred to as "module" or "system" herein. In addition, in some embodiments, the present application can also be implemented in the form of a computer program product in one or more computer-readable media, and the computer-readable medium contains computer-readable program code. Computer-readable storage media can be, for example, but not limited to - electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any combination of the above.
[0166] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.
[0167] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.
Claims
1. A method for training an image recognition model, characterized in that: include: Acquire a plurality of images to be trained, wherein the images to be trained are grid images including a plurality of grids of the same size; Inputting the plurality of images to be trained into a preset recognition model, performing multi-scale processing on the images to be trained, and obtaining corresponding multi-scale feature maps; Performing prediction processing on the multi-scale feature map to obtain a prediction vector corresponding to each grid in the image to be trained; Based on the multiple prediction vectors corresponding to the image to be trained, a predicted recognition box and a predicted category of the target object in the image to be trained are obtained; Based on the predicted recognition boxes and the predicted categories respectively corresponding to the plurality of images to be trained, the network parameters in the preset recognition model are adjusted to obtain a target recognition model.
2. The method according to claim 1, characterized in that The step of obtaining a plurality of images to be trained comprises: Obtain multiple collected images related to abnormal events in urban governance; Preprocessing the acquired image to obtain a corresponding processed image; The processed image is segmented to obtain a to-be-trained image having a plurality of grids of the same size.
3. The method according to claim 1, characterized in that The performing multi-scale processing on the image to be trained to obtain a corresponding multi-scale feature map includes: Using the backbone network of the preset recognition model to extract features from the image to be trained, and obtaining feature maps of multiple different scales; A fusion operation is performed on the feature maps of the multiple different scales to obtain a multi-scale feature map corresponding to the image to be trained; wherein the backbone network includes a first feature extraction module, a second feature extraction module, a pooling module and a fusion module.
4. The method according to claim 1, characterized in that: The step of obtaining a predicted recognition box and a predicted category of a target object in the image to be trained based on the plurality of predicted vectors corresponding to the image to be trained comprises: Identify the prediction vector to obtain a corresponding prediction sub-identification box of the target object in the grid; Determining a predicted probability that the target object in the grid is of a preset category; Based on each of the predicted sub-recognition boxes in the image to be trained and the corresponding prediction probability, a predicted recognition box and a predicted category of the target object in the image to be trained are obtained.
5. The method according to claim 4, characterized in that The recognition frame information of the predicted sub-recognition frame includes a center position, a width, a height, and a recognition frame confidence; The obtaining, based on each of the predicted sub-recognition frames in the image to be trained and the corresponding predicted probability, a predicted recognition frame and a predicted category of the target object in the image to be trained comprises: Based on the recognition frame confidence and the predicted probability of the predicted sub-recognition frame, calculating the category probability that the category of the target object appears in the predicted sub-recognition frame; The frame consisting of a plurality of predicted sub-recognition frames whose category probability is greater than or equal to a set value and whose preset categories are the same is used as a predicted recognition frame; The preset category corresponding to the predicted identification box is used as the predicted category.
6. The method according to claim 1, characterized in that The method of adjusting the network parameters in the preset recognition model based on the predicted recognition boxes and the predicted categories respectively corresponding to the plurality of images to be trained to obtain the target recognition model includes: Determine a recognition frame regression loss between the predicted recognition frame and the real recognition frame corresponding to each of the to-be-trained images; Determine the binary cross entropy loss of the predicted category corresponding to each of the images to be trained; Based on the multiple recognition box regression losses and the multiple binary cross entropy losses, the network parameters in the preset recognition model are adjusted to obtain a target recognition model.
7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Detection task requirements; When the task requirement is detected, determining a directional event data set based on the task requirement; The target recognition model is fine-tuned based on the directional event set, so as to apply the fine-tuned target recognition model to the recognition of the image to be tested corresponding to the task requirement.
8. An image recognition model training system, characterized in that: include: An acquisition module, used for acquiring a plurality of images to be trained, wherein the images to be trained are grid images including a plurality of grids of the same size; A processing module, used for inputting the plurality of images to be trained into a preset recognition model, performing multi-scale processing on the images to be trained, and obtaining corresponding multi-scale feature maps; A first prediction module, used to perform prediction processing on the multi-scale feature map to obtain a prediction vector corresponding to each grid in the image to be trained; A second prediction module, used to obtain a predicted recognition box and a predicted category of a target object in the image to be trained based on the multiple prediction vectors corresponding to the image to be trained; The parameter adjustment module is used to adjust the network parameters in the preset recognition model based on the predicted recognition boxes and the predicted categories respectively corresponding to the multiple images to be trained, so as to obtain a target recognition model.
9. A computing device comprising a memory, a processor, and a program stored in the memory and running on the processor, characterized in that: When the processor executes the program, the steps of an image recognition model training method as described in any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the steps of an image recognition model training method as described in any one of claims 1 to 7.
Citation Information
Patent Citations
Target identification method and device based on improved multi-scale depth model
CN113221956A
Smoking behavior identification method based on optimized YOLOv4 model
CN113807276A
Object prediction model training method and system based on YOLOV5
CN114266883A
Insulator defect detection model generation method and device, equipment and storage medium
CN115829999A
