ROI extraction method suitable for board end environment
By optimizing the structure and extracting features of a lightweight neural network model, and combining it with traditional image processing algorithms, the real-time performance and accuracy issues of ROI extraction in board-side environments are solved, realizing an efficient and lightweight ROI extraction method suitable for scenarios such as intelligent monitoring and industrial inspection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUATAI AURORA INTELLIGENT SENSING TECHNOLOGY (WUXI) CO LTD
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-17
AI Technical Summary
Existing ROI extraction methods struggle to simultaneously meet the requirements of real-time performance, lightweight design, and high accuracy in a board-side environment. Traditional methods exhibit poor robustness, while neural network-based methods suffer from high computational complexity and are difficult to adapt to the limited resources available on the board.
By optimizing the structure of a lightweight neural network model, removing redundant structures, adding a feature aggregation layer, and combining it with traditional image processing algorithms, feature extraction and image post-processing are achieved, resulting in a ROI extraction method suitable for board-side environments.
It achieves stable ROI extraction in complex scenarios, reduces computational overhead and resource consumption, controls inference latency to the millisecond level, adapts to board-side resource constraints, and improves the practicality and cross-platform adaptability of the algorithm.
Smart Images
Figure CN121883803A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer vision technology, and in particular to a method for ROI extraction suitable for board-end environments. Background Technology
[0002] Region of Interest (ROI) extraction is a core technology in computer vision. Its core objective is to accurately locate and segment target regions with practical analytical value from complex images or video streams, providing data support for subsequent tasks such as target recognition, feature extraction, and semantic analysis. This significantly reduces the amount of data required for subsequent processing and improves the overall system efficiency. This technology is widely used in intelligent monitoring, industrial quality inspection, vehicle-mounted assisted driving, portable medical image analysis, and many other fields. Especially in terminal devices based on embedded boards (such as FPGAs, ARM architecture boards, and dedicated AI acceleration boards), the efficiency and reliability of ROI extraction directly determine the real-time performance and practicality of the entire vision system.
[0003] The board-side environment places stringent requirements on ROI extraction algorithms: on the one hand, it needs to meet real-time requirements to adapt to dynamic input scenarios such as video streams; on the other hand, due to the limited computing power, storage resources and strict power consumption constraints of the board-side equipment, the algorithm must have the characteristics of lightweight and low resource consumption. The current mainstream ROI extraction technologies are mainly divided into two categories: traditional image processing algorithms and neural network-based algorithms. These two types of ROI extraction methods show significant differences in principle and performance, and both have shortcomings in adapting to the board-side environment: (1) ROI extraction methods based on traditional image processing algorithms rely on the underlying visual information such as grayscale, edge, and texture of the image to locate the target area. The core advantage is that the principle is simple, the computational overhead is small, and there is no need for a complex model training process. However, the disadvantage is that the robustness is poor, it depends on fixed detection features, is easily affected by noise, and cannot extract the ROI region of a specific target in a complex background. (2) ROI extraction methods based on neural networks are gradually becoming the mainstream technology. They have shown good extraction accuracy and robustness in complex scenarios. However, neural network-based methods are generally complex in structure, slow inference speed, and have high hardware computing power requirements, making them difficult to adapt to the limited computing power resources on the board. It can be seen that the existing two types of ROI extraction methods do not perform well in the board environment. Summary of the Invention
[0004] This application addresses the aforementioned problems and technical requirements by proposing a method for ROI extraction suitable for board-side environments. The technical solution of this application is as follows: A method for ROI extraction suitable for board-side environments, the method comprising: The lightweight neural network model that has completed model training is optimized in terms of model structure while keeping the feature extraction capability unchanged, so as to obtain a feature extraction network for the detection target and deploy it on the target board; The target board acquires the original image and uses the deployed feature extraction network to extract features to obtain an initial feature map A. The larger the pixel value of each pixel in the initial feature map A, the higher the probability that the corresponding pixel in the original image belongs to the ROI of the detection target. Perform image post-processing operations on the initial feature map A to obtain the ROI of the detected target with the same size as the original image.
[0005] A further technical solution involves optimizing the structure of the lightweight neural network model after training to obtain a feature extraction network, including: Redundant structures unrelated to ROI extraction are removed from the lightweight neural network model, and a feature aggregation layer is added to enhance the feature response of the detected target.
[0006] A further technical solution involves optimizing the structure of the lightweight neural network model after training to obtain a feature extraction network, which also includes: Replace the hierarchical structures in lightweight neural network models that pose compatibility risks with general hierarchical structures that have the same functionality.
[0007] The further technical solution involves using the MobileNet_v2 lightweight neural network model, and optimizing the model structure of the trained lightweight neural network to obtain the feature extraction network, which includes: Remove the classification probability output branch after the last Clip layer in the MobileNet_v2 neural network model, and add a ReduceMean layer as a feature aggregation layer after the last Clip layer; replace all Clip layers in the MobileNet_v2 neural network model with ReLU layers.
[0008] A further technical solution involves performing image post-processing operations on the initial feature map A to obtain the ROI, including: The pixel values in the initial feature map A are mapped to the interval [0, 255] to obtain the normalized feature map B; The normalized feature map B is upsampled to restore the original image size, resulting in the upsampled feature map C. The upsampled feature map C is segmented into foreground and background regions to obtain a binary map E; The ROI of the detected target is obtained based on the foreground region in the binary image E.
[0009] A further technical solution involves segmenting the upsampled feature map C into foreground and background regions to obtain a binary map E, including: An improved watershed algorithm is used to segment the foreground and background of the upsampled feature map C to obtain a binary map D; Connectivity separation and filtering are performed on the foreground region in binary graph D, and connected components that reach the area threshold are retained to obtain binary graph E.
[0010] A further technical solution involves using an improved watershed algorithm to segment the upsampled feature map C into foreground and background to obtain a binary map D, including: After Gaussian blurring of the upsampled feature map C, the gradient map G is calculated using the Sobel operator; The threshold T is calculated using the Otsu algorithm. Foreground markers are added to pixels in the gradient map G whose pixel values are greater than T, and background markers are added to pixels in the gradient map G whose pixel values are less than T / 2. Using the gradient map G as input, and adding foreground and background markers as constraints, a watershed transformation is performed to obtain a binary map D.
[0011] A further technical solution involves obtaining the ROI of the detected target based on the foreground region in the binary image E, including: Calculate the covariance matrix for each foreground region. ,in, , , ;in, It is the first in the foreground area The coordinates of each pixel It is the total number of pixels in the foreground region. , ; Calculate the covariance matrix eigenvalues and , ; eigenvalues The corresponding feature vector direction is used as the length direction of the ROI rectangle to determine the length of the ROI rectangle. , eigenvalue The corresponding feature vector direction is used as the width direction of the ROI rectangle, thus determining the width of the ROI rectangle. This yields the parameters of the ROI rectangle, which are used to indicate the position and size of the ROI of the detected target.
[0012] A further technical solution involves upsampling the normalized feature map B to obtain the upsampled feature map C, including: The normalized feature map B is upsampled using a bilinear interpolation algorithm to obtain the upsampled feature map C.
[0013] A further technical solution involves deploying the feature extraction network onto the target board, including: After exporting the feature extraction network to the intermediate format ONNX, a conversion tool that matches the model format of the target board is used to convert the feature extraction network in the intermediate ONNX format. Quantization optimization is performed during the format conversion process to obtain a feature extraction network that matches the model format of the target board and then it is deployed on the target board.
[0014] The beneficial technical effects of this application are: This application discloses a ROI extraction method suitable for board-side environments. This method combines the lightweight characteristics of traditional image processing algorithms with the strong feature learning capabilities of neural network algorithms. Through a staged processing mechanism and technology fusion, it achieves stable extraction in complex scenarios while accelerating processing speed to meet the needs of dynamic scenarios and significantly reducing end-to-end latency. This method significantly reduces the computational overhead and resource consumption of the algorithm while ensuring the accuracy and robustness of ROI extraction. It can achieve stable extraction in complex scenarios while balancing board-side resource consumption, ensuring that the inference latency on board-side devices (such as ARM architecture boards and dedicated AI chips) is controlled at the millisecond level. On the RK3588 platform, the total execution time is <10ms, meeting the real-time processing requirements of video streams. This method is highly accurate, fast, lightweight, and compatible, and can adapt to the resource constraints and deployment requirements of board-side environments. It is conducive to promoting the widespread application of ROI extraction technology in the field of edge computing and is suitable for scenarios with high real-time requirements such as intelligent monitoring and industrial inspection.
[0015] The feature extraction network obtained by optimizing the structure of the MobileNet_v2 neural network model has a computational cost of ≤0.5 GMACs, which is only 1 / 360 of Faster R-CNN (180 GMACs). It can run smoothly on ARM Cortex-A73 architecture boards without dedicated AI acceleration units, avoiding dependence on high-cost GPU boards and reducing the hardware cost of terminal devices.
[0016] This method replaces the hierarchical structures in lightweight neural network models that pose compatibility risks with general hierarchical structures that have the same functionality. Combined with format adaptation optimization, it achieves full coverage of mainstream board-side model formats, enabling the feature extraction network to be directly compatible with mainstream board-side model formats such as om, rknn, nb, and wk, avoiding format conversion conflicts. The conversion success rate on mainstream chips is 100%, reducing deployment complexity and cost, and improving the algorithm's cross-platform adaptability.
[0017] This method improves the algorithm's adaptability to scenarios with varying lighting conditions, complex backgrounds, and diverse target shapes by using neural network adaptive feature learning and traditional algorithm parameter adaptive adjustment mechanisms. It reduces manual debugging costs, achieves "one-time deployment, multi-scenario adaptation", and enhances the algorithm's practicality and promotional value. Attached Figure Description
[0018] Figure 1 This is a flowchart of a method for ROI extraction according to an embodiment of this application.
[0019] Figure 2 This is a schematic diagram illustrating the process of optimizing the MobileNet_v2 neural network model and deploying it on a target board in one embodiment of this application.
[0020] Figure 3 This is a schematic flowchart illustrating the image post-processing operation performed on the initial feature map A in one embodiment of this application. Detailed Implementation
[0021] The specific embodiments of this application will be further described below with reference to the accompanying drawings.
[0022] This application discloses a method for ROI extraction suitable for board-side environments. Please refer to [link / reference needed]. Figure 1 The flowchart shown illustrates the ROI extraction method, which includes: Step 110: Train the lightweight neural network model according to the detection target.
[0023] First, a training dataset for the detection targets is obtained. This dataset contains a large number of sample images, each labeled with the target information. The target information can have different meanings depending on the specific application; for example, it could be industrial parts in an industrial setting, vehicles and pedestrians in a traffic setting, or medical lesions in medical image recognition. Specific model training methods can be found in existing mature methods and will not be elaborated upon here.
[0024] Step 120: Optimize the model structure of the lightweight neural network model that has completed model training while keeping the feature extraction capability unchanged, to obtain a feature extraction network for the detection target.
[0025] The lightweight neural network model itself already possesses lightweight characteristics. In order to further reduce computing power requirements and ensure that the inference latency at the board end is controlled within milliseconds to meet the real-time processing requirements of video streams, the network model is further optimized. This includes removing redundant structures in the lightweight neural network model that are not related to ROI extraction and adding a feature aggregation layer, which is used to enhance the feature response of the detected target.
[0026] In one embodiment, please refer to Figure 2The flowchart illustrates the process, using the lightweight MobileNet_v2 neural network model. This model boasts the advantages of lightweight design with depthwise separable convolutions and inverted residual structures, with ≤4M parameters and ≤0.5 GMACs computational cost. Redundant structures in the MobileNet_v2 model unrelated to ROI extraction include a classification probability output branch. This branch comprises the GlobalAveragePool layer, fully connected layer, and Softmax layer following the last Clip layer in the MobileNet_v2 model. This branch outputs class probabilities and is irrelevant to ROI extraction; therefore, it has been removed.
[0027] To enhance the feature response of the detected target, a ReduceMean layer is added as a feature aggregation layer after the last Clip layer in the MobileNet_v2 neural network model. The ReduceMean layer calculates the mean along the channel dimension of the feature map obtained from the last Clip layer, and uses this as the output feature map of the feature extraction network. The feature map output by the ReduceMean layer indicates the probability that each pixel in the input image belongs to the ROI of the detected target; that is, the feature map output by the ReduceMean layer contains the location information of the ROI to be extracted. The core principle of this operation is that the feature response value of the ROI of the detected target is significantly higher than that of the background region, and the calculation of the channel mean enhances the grayscale difference of the ROI of the detected target.
[0028] In addition, considering that the network needs to be deployed to the board in the future, and that different board formats are different, the operators in the model structure are further standardized in order to reduce the board deployment threshold. The hierarchical structure with compatibility conflict risk in the lightweight neural network model is replaced with a general hierarchical structure with the same function to ensure the compatibility of the feature extraction network on different architecture boards.
[0029] When using the MobileNet_v2 neural network model, all Clip layers that are prone to compatibility conflicts are replaced with ReLU layers. Both ReLU and Clip layers can limit the range of feature values, so the replacement does not affect the model's feature extraction capability, but it is more versatile and avoids compatibility issues.
[0030] Step 130: Deploy the feature extraction network onto the target board.
[0031] To lower the deployment threshold on the target board, one embodiment also includes format adaptation optimization. The feature extraction network is exported to the intermediate ONNX format, ensuring that all network layers use the standard layers commonly used in ONNX. Then, based on the target board type, a conversion tool (officially provided toolchain) matching the target board's model format is used to convert the intermediate ONNX feature extraction network, thereby converting the ONNX model into a feature extraction network that matches the target board's model format. Finally, it is deployed to the target board, avoiding format conversion conflicts, ensuring conversion success on various boards, reducing deployment complexity and cost, and improving cross-platform adaptability. Furthermore, quantization optimization is performed during the format conversion process, quantizing float32 precision to int8, reducing storage usage and computational load.
[0032] The model formats of the boards that this application is compatible with include om, rknn, nb, and wk. Several mainstream boards, their model formats, and conversion tools are shown in the table below:
[0033] Step 140: The target board acquires the original image and uses the deployed feature extraction network to extract features to obtain the initial feature map A.
[0034] As described above, the initial feature map A obtained by the feature extraction network is used to characterize the probability that a pixel at a corresponding position in the original image belongs to the ROI of the target. The larger the pixel value of each pixel in the initial feature map A, the higher the probability that the corresponding pixel in the original image belongs to the ROI of the target. Since the feature extraction network is based on a lightweight neural network model and has undergone further structural optimization, the model inference speed of the feature extraction network is relatively fast. Based on the ARM Cortex-A73 architecture board, the model inference latency can be controlled within 5ms.
[0035] Because the feature extraction network involves multiple convolutional layers during feature extraction from the original image, the size of the output initial feature map A is smaller than the original image. Taking the MobileNet_v2 neural network model as an example, the last Clip layer in the MobileNet_v2 neural network model inputs the feature map into the ReduceMean layer to obtain the initial feature map A. The size of the initial feature map A is the same as the size of the feature map output by the last Clip layer, which is smaller than the original input image. The specific size is determined by the downsampling parameters. For example, in one instance, the size of the input original image is 224×224, while the size of the initial feature map A is 7×7, which is much smaller than the size of the original image.
[0036] Step 150: Taking advantage of the sparsity and low resolution of the initial feature map A, further image post-processing operations are performed on the initial feature map A to obtain the ROI of the detection target with the same size as the original image. Thus, the accurate mapping from the initial feature map A to the ROI of the original image is achieved through a purely hardware-friendly operator. The entire image post-processing process has a delay of ≤3ms, ensuring the end-to-end real-time performance of the algorithm.
[0037] In one embodiment, the image post-processing operation performed on the initial feature map A includes the following steps, please refer to [reference needed]. Figure 3 The flowchart shown below: Step 310: Map the pixel values in the initial feature map A to the interval [0, 255] to obtain the normalized feature map B.
[0038] The pixel value range of the initial feature map A varies greatly in different scenarios. First, normalization is performed to obtain the pixel values of each pixel in the normalized feature map B. The pixel values in the normalized feature map B are of type uint8. This step can be completed by traversing the initial feature map A once, which is computationally efficient.
[0039] Step 320: The size of the normalized feature map B is the same as that of the initial feature map A. The image size is still very small. Therefore, the normalized feature map B is upsampled to restore the size of the original image, resulting in the upsampled feature map C.
[0040] In one embodiment, the normalized feature map B is upsampled using a bilinear interpolation algorithm to obtain the upsampled feature map C. This method has only 1 / 3 the computational cost of bicubic interpolation, good plate-end adaptability, and can ensure that the edges of the upsampled feature map C are smooth, avoiding jagged distortion, which is beneficial to improving the subsequent segmentation accuracy.
[0041] In steps 330 and 320, the foreground and background regions of the upsampled feature map C show a significant grayscale difference. Therefore, the upsampled feature map C is further segmented into foreground and background regions to obtain a binary map E, including: (1) First, the improved watershed algorithm is used to segment the foreground and background of the upsampled feature map C to obtain the binary map D. Specifically: Gaussian blurring is applied to the upsampled feature map C to remove noise interference. Then, the gradient map G is calculated using the Sobel operator to highlight the edges of the detected target.
[0042] Then, the Otsu algorithm is used to automatically calculate the threshold T. Foreground markers are added to pixels in the gradient map G with pixel values greater than T, and foreground markers and background markers are added to pixels in the gradient map G with pixel values less than T / 2. Pixels in the gradient map G with pixel values greater than T / 2 but not greater than T belong to the region to be segmented.
[0043] Finally, using the gradient map G as input and the added foreground and background markers as constraints, a watershed transformation is performed to obtain the binary map D. The specific transformation method can be found in the watershed algorithm, and will not be elaborated here. The improved watershed algorithm, through marker control, effectively avoids over-segmentation of target details by the traditional watershed algorithm, ensuring the integrity of the foreground region.
[0044] (2) The obtained binary image D may contain small-area noisy connected components (such as isolated points or fine edges). Therefore, the foreground region in the binary image D is further separated and filtered to retain connected components that reach the area threshold to obtain the binary image E, and the small-area noisy regions are deleted. In one embodiment, the binary image D can be scanned from left to right and from top to bottom using the scan line method to determine the set of pixels of each connected component, thereby calculating the area of each connected component. The area threshold for this step can be set by customization. In one embodiment, it is taken as 1% of the area of the original image. For example, in one instance, the image size of the original image is 224×224, so the area threshold can be 501.
[0045] Step 340: Obtain the ROI of the detected target based on the foreground region in the binary image E.
[0046] In one embodiment, a rotational caliper algorithm is used to fit the minimum bounding rectangle as the ROI detection box for each connected component of the foreground region in the binary image E, and the parameters of the ROI rectangle are obtained. The parameters of the ROI rectangle are used to indicate the position and size of the ROI of the detected target. This method can adapt to any rotation angle of the detected target and is more accurate than axis-aligned rectangles.
[0047] Specifically, for each connected component of the foreground region in the binary graph E: Calculate the covariance matrix of the foreground region. :
[0048] in,
[0049]
[0050]
[0051] in, It is the first in the foreground area The coordinates of each pixel It represents the total number of pixels in the foreground region. and It is in the foreground area The average coordinates of each pixel, and , .
[0052] Then calculate the covariance matrix. Two eigenvalues and , . Eigenvalues The corresponding feature vector direction is used as the length direction of the ROI rectangle to determine the length of the ROI rectangle. , eigenvalue The corresponding feature vector direction is used as the width direction of the ROI rectangle, thus determining the width of the ROI rectangle. Therefore, the parameters of the ROI rectangle can be obtained, including the center coordinates of the ROI rectangle. ,long ,Width and the rotation angle of the ROI rectangle relative to the coordinates of the original image. Alternatively, one approach is to convert the coordinates of the top-left and bottom-right pixels of the axis-aligned rectangle to meet the input requirements of subsequent detection algorithms.
[0053] In one example, taking vehicles as the target and Rockchip RK3588 board (4-core Cortex-A76 + 4-core Cortex-A55, 6 TOPS computing power) as the target, the performance of the proposed ROI extraction method, the Faster R-CNN-based ROI extraction method, and the traditional edge detection-based ROI extraction method is compared using a traffic monitoring dataset (224×224 resolution) containing 500 vehicle images. The performance metrics are as follows:
[0054] The verification results show that while maintaining accuracy close to Faster R-CNN, this application has significantly optimized end-to-end latency and model size compared to Faster R-CNN, and has excellent board-side compatibility, fully meeting the usage requirements of board-side environments.
[0055] The above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.
Claims
1. A method for ROI extraction suitable for plate-end environments, characterized in that, The ROI extraction method includes: The lightweight neural network model that has completed model training is optimized in terms of model structure while keeping the feature extraction capability unchanged, so as to obtain a feature extraction network for the detection target and deploy it on the target board; The target board acquires the original image and uses the deployed feature extraction network to extract features to obtain an initial feature map A. The larger the pixel value of each pixel in the initial feature map A, the higher the probability that the corresponding pixel in the original image belongs to the ROI of the detection target. Perform image post-processing operations on the initial feature map A to obtain the ROI of the detected target with the same size as the original image.
2. The ROI extraction method according to claim 1, characterized in that, Optimizing the model structure of a lightweight neural network model after training yields a feature extraction network, including: Redundant structures unrelated to ROI extraction are removed from the lightweight neural network model, and a feature aggregation layer is added to enhance the feature response of the detected target.
3. The ROI extraction method according to claim 2, characterized in that, Optimizing the model structure of a lightweight neural network model after training to obtain a feature extraction network also includes: Replace the hierarchical structures in lightweight neural network models that pose compatibility risks with general hierarchical structures that have the same functionality.
4. The ROI extraction method according to claim 3, characterized in that, The lightweight neural network model used is the MobileNet_v2 neural network model. The feature extraction network, obtained by optimizing the model structure of the trained lightweight neural network model, includes: Remove the classification probability output branch after the last Clip layer in the MobileNet_v2 neural network model, and add a ReduceMean layer as a feature aggregation layer after the last Clip layer; replace all Clip layers in the MobileNet_v2 neural network model with ReLU layers.
5. The ROI extraction method according to claim 1, characterized in that, Post-processing operations on the initial feature map A to obtain the ROI include: The pixel values in the initial feature map A are mapped to the interval [0, 255] to obtain the normalized feature map B; The normalized feature map B is upsampled to restore the original image size, resulting in the upsampled feature map C. The upsampled feature map C is segmented into foreground and background regions to obtain a binary map E; The ROI of the detected target is obtained based on the foreground region in the binary image E.
6. The ROI extraction method according to claim 5, characterized in that, After upsampling, the feature map C is segmented into foreground and background regions to obtain a binary map E, which includes: An improved watershed algorithm is used to segment the foreground and background of the upsampled feature map C to obtain a binary map D; Connectivity separation and filtering are performed on the foreground region in binary graph D, and connected components that reach the area threshold are retained to obtain binary graph E.
7. The ROI extraction method according to claim 6, characterized in that, An improved watershed algorithm is used to segment the foreground and background of the upsampled feature map C to obtain a binary map D, which includes: After Gaussian blurring of the upsampled feature map C, the gradient map G is calculated using the Sobel operator; The threshold T is calculated using the Otsu algorithm. Foreground markers are added to pixels in the gradient map G whose pixel values are greater than T, and background markers are added to pixels in the gradient map G whose pixel values are less than T / 2. Using the gradient map G as input, and adding foreground and background markers as constraints, a watershed transformation is performed to obtain a binary map D.
8. The ROI extraction method according to claim 5, characterized in that, The ROIs of the detected targets, derived from the foreground region in the binary image E, include: Calculate the covariance matrix for each foreground region. ,in, , , ;in, It is the first in the foreground area The coordinates of each pixel It is the total number of pixels in the foreground region. , ; Calculate the covariance matrix eigenvalues and , ; eigenvalues The corresponding feature vector direction is used as the length direction of the ROI rectangle to determine the length of the ROI rectangle. , eigenvalue The corresponding feature vector direction is used as the width direction of the ROI rectangle, thus determining the width of the ROI rectangle. This yields the parameters of the ROI rectangle, which are used to indicate the position and size of the ROI of the detected target.
9. The ROI extraction method according to claim 5, characterized in that, Upsampling the normalized feature map B yields the upsampled feature map C, which includes: The normalized feature map B is upsampled using a bilinear interpolation algorithm to obtain the upsampled feature map C.
10. The ROI extraction method according to claim 1, characterized in that, Deploying the feature extraction network to the target board includes: After exporting the feature extraction network to the intermediate format ONNX, a conversion tool that matches the model format of the target board is used to convert the feature extraction network in the intermediate ONNX format. Quantization optimization is performed during the format conversion process to obtain a feature extraction network that matches the model format of the target board and then it is deployed on the target board.