A real-time video semantic segmentation method for key components of unmanned aerial vehicles based on deep learning

Abstract

The present invention discloses a real-time video semantic segmentation method for key components of unmanned aerial vehicles based on deep learning, and belongs to the technical field of semantic segmentation of images. The present invention obtains a semantic segmentation data set of unmanned aerial vehicle videos by collecting all scenes where unmanned aerial vehicles appear on the network, splitting multiple consecutive frames, and performing data enhancement after marking key frames. A multi-branch feature extraction network is used to perform feature learning on video frames, and multi-branch features are fused and supplemented to achieve the effect of deep network feature extraction, so as to achieve the effect of improving network speed under the same accuracy; the low-level features of the segmentation stage are symmetrically fused to the subsequent convolution stage using optical flow information to ensure that edge details are not lost in the convolution network, and the edge segmentation accuracy is ensured, which provides a guarantee for the subsequent calculation of the center point using edge information. According to the video semantic segmentation recognition results, the present invention obtains the contour and center point of the target component based on pixel replacement and binarization processing, thereby improving ease of use.

Description

Technical Field

[0001] The present invention relates to the technical field of image semantic segmentation, and in particular to a real-time video semantic segmentation method of key components of unmanned aerial vehicles based on deep learning. Background Art

[0002] There are three types of image segmentation tasks: semantic segmentation, instance segmentation, and panoptic segmentation. The segmentation usually refers to semantic segmentation, which is a classification task for all pixels in the image. Instance segmentation refers to the segmentation of specific objects, which are instances. Not only are the pixels in the image segmented, but they are also further classified. That is, in semantic segmentation, people are classified into the same category, but in instance segmentation, people are divided into multiple categories. Panoptic segmentation is a synthesis of semantic segmentation and instance segmentation, and all objects in the image are classified and segmented. Semantic segmentation is a key technology in the field of computer vision and pattern recognition. Together with object detection and image classification, it constitutes the three major tasks of environmental perception. In 2014, Shelhamer et al. proposed a semantic method FCN based on a fully convolutional neural network. This work has made significant progress in semantics. It not only answers the question of how CNN can achieve semantic separation with limited training and limited training, but also effectively solves the problem of how to generate semantic prediction output pixels of arbitrary input size.

[0003] With the development of image semantic segmentation, after reaching a bottleneck in accuracy, researchers have turned their attention to the speed of semantic segmentation. However, since deep semantic information often requires a high cost to obtain, such as deeper networks and more effective feature extraction methods, this also limits the processing speed of semantic segmentation. In addition, the quality and posture of dynamic objects in videos are much lower than those of static images. For example, the rapid movement of objects in videos causes blurring of objects, as well as occlusion, etc. The concept of video semantic segmentation has been proposed. Video semantic segmentation is a commonly used method for autonomous driving. Since autonomous driving has high real-time requirements, its processing speed is about 15FPS-30FPS (Frames Per Second). The main difference between video semantic segmentation networks and image semantic segmentation networks is the addition of the use of time dimension information. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time video semantic segmentation method for key components of drones based on deep learning, which solves the problems of lack of scene processing for segmentation of key components of drones in existing neural networks, low edge processing accuracy and slow processing speed.

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

[0006] A real-time video semantic segmentation method for key components of drones based on deep learning includes the following steps:

[0007] Step 1, create a data set:

[0008] Based on network resources, all scene images where drones appear are collected, and after image annotation and data enhancement of target components, several video semantic segmentation datasets are obtained, with images of a specified number of consecutive frames as a scene segment.

[0009] Step 2: Build and train the video semantic segmentation network:

[0010] The video semantic segmentation network includes M feature extraction network branches, each of which includes multiple convolutional layers and an optical flow propagation module for optical flow alignment and upsampling. The output feature maps of the M feature extraction branches are spliced ​​and fused, and then input into a fully connected layer with a Softmax function after passing through an encoding network and an attention module to output probability information of each target component.

[0011] The video semantic segmentation network is trained based on the video semantic segmentation dataset, and the optimal model is obtained after the training is completed;

[0012] Step 3, inputting the video sequence to be identified into the optimal model obtained in step 2, and obtaining the recognition result based on its output, wherein the length of the video sequence to be identified is the same as the number of frames of the scene fragment of the video semantic segmentation dataset;

[0013] Perform key point processing on the recognition results: replace pixels on the recognition results, retain the target component, perform binarization processing, and then obtain the contour of the target component based on the contour acquisition process; calculate the image center moment based on the contour of the target component to obtain the center point of the target component and output it visually.

[0014] Preferably, in step 1, the image annotation of the target component is specifically as follows: the target components of the drone (such as the rotor, main body, and payload of the drone) are framed with continuous points, and different pixel values ​​are used for target components of different categories.

[0015] Preferably, in step 1, data enhancement includes: image flipping, rotation, scaling, cropping and shifting.

[0016] Preferably, in step 2, each feature extraction network branch is specifically set as:

[0017] The first convolutional layer contains a 3*3 convolution and a 3*3 residual structure (bottleneck), and the number of output channels is 16;

[0018] The second convolution layer has two residual structures with a convolution kernel size of 3*3 and one residual structure with a convolution kernel of 5*5, and the number of output channels is 40;

[0019] The third convolutional layer consists of five residual structures with a convolution kernel size of 5*5 and 96 output channels;

[0020] The size of the fourth convolution kernel is 5*5 with 2 residual structures, and the number of output channels is 96;

[0021] The fifth layer is a 3*3 pooling layer;

[0022] The fifth layer of the network is connected to the pyramid pooling network (PPM network), and the output of the third layer of the network and the pyramid pooling network are optically aligned and upsampled to obtain the sixth layer, and the second layer of the network and the sixth layer output are optically aligned and upsampled to obtain the seventh layer, and then the first layer of the network and the seventh layer network output are optically aligned and upsampled, and then the optical flow alignment results of the sixth and seventh layers are fused to obtain the output feature map of the feature extraction network branch.

[0023] Preferably, in step 2, when training the video semantic segmentation network, the back propagation algorithm and the gradient descent method are used for training, and the weight value of each layer of the network is adjusted by calculating the intersection over union (IoU) value of each category of drones, and it is iterated repeatedly until the maximum intersection over union value is obtained.

[0024] Preferably, in step 3, performing key point processing on the recognition result specifically includes:

[0025] Filter the target parts by using pixel replacement based on the recognition results;

[0026] Based on the recognition results, the maximum connected domain algorithm is used for a single object to filter out the largest target and prevent the recognition results from affecting the center point calculation.

[0027] Binarization is performed based on the recognition result to obtain the outline of the target component;

[0028] The image center moment is calculated based on the contour of the target part to obtain the center point of the target part and visualize the output.

[0029] The technical solution provided by the present invention brings at least the following beneficial effects:

[0030] (1) The present invention collects all scenes where drones appear on the network, splits the specified continuous frames, and performs data enhancement after marking the key frames to obtain a drone video semantic segmentation dataset. A multi-branch feature extraction network is used to learn the features of the video frames, and the multi-branch features are fused and supplemented to achieve the deep network feature extraction effect, so as to achieve the effect of improving the network speed with the same accuracy;

[0031] (2) In the present invention, since most drone flight scenes are in the air, the edges between the background and the target are difficult to distinguish, which affects the semantic segmentation effect. In the feature extraction and segmentation network, the present invention uses optical flow information to symmetrically fuse the low-level features of the segmentation stage into the subsequent convolution stage, ensuring that the edge details are not lost in the convolution network, ensuring the accuracy of edge segmentation, and providing a guarantee for the subsequent calculation of the center point using edge information.

[0032] (3) The video semantic segmentation network architecture of the present invention uses time dimension information to ensure that the multi-scale changes of the drone during flight can be effectively segmented and processed, thereby improving the recognition accuracy of common occlusion, motion tearing, and background homogeneity in the video;

[0033] (4) The present invention replaces pixels of the image based on the video semantic segmentation recognition result, retains only the target component, and then performs binarization, uses the corresponding function to obtain the contour, and further uses the calculation of the image center moment to obtain the center point, making the present invention more user-friendly. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 A processing flow chart of a method for real-time video semantic segmentation of key components of a drone based on deep learning provided in an embodiment of the present invention;

[0036] Figure 2 A flowchart of establishing a data set in an embodiment of the present invention;

[0037] Figure 3 A structural diagram of a network model used in an embodiment of the present invention;

[0038] Figure 4 This is a center point recognition result diagram of an embodiment of the present invention;

[0039] Figure 5 Schematic diagram of comparing recognition results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0041] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0042] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0043] An embodiment of the present invention provides a real-time video semantic segmentation method for key components of a drone based on deep learning, which includes three parts: construction of a Phomtom dataset. The dataset includes a Phomtom image dataset and a Phomtom video dataset; for the video semantic segmentation part, a component segmentation network algorithm including an input layer, a fully connected layer, several hidden layers, and an output layer is designed, and the model is trained based on the above-mentioned Phomtom dataset to obtain the optimal model; a semantic segmentation method is designed by combining SFNet and TDNet to improve the accuracy and speed of semantic segmentation of the Phomtom drone. The present invention mainly solves the problems of poor edge processing and multi-scale processing of existing drones.

[0044] like Figure 1 As shown, an embodiment of the present invention provides a method for real-time video semantic segmentation of key components of a drone based on deep learning, comprising the following steps:

[0045] Step 1: Establish a dataset: After collecting, annotating and enhancing the scene images of all drones on the Internet, we obtain several video semantic segmentation datasets with 30 consecutive frames as a scene segment, and divide them into training sets and test sets;

[0046] Step 2: Construct and train a video semantic segmentation network: The video semantic segmentation network includes M (preferably 2) feature extraction network branches, each of which includes multiple convolutional layers and an optical flow propagation module for optical flow alignment and upsampling. The output feature maps of the M feature extraction branches are spliced ​​and fused, and then input into a fully connected layer with a Softmax function after passing through an encoding network and an attention module to output the probability information of each target component. When obtaining the recognition result, the target component to which each pixel belongs is determined based on the maximum probability. The above training set is input into the network model. After the training is completed and the optimal model is obtained, the test set is input into the optimal model to obtain the recognition result;

[0047] Step 3: Perform key point processing on the result graph of the video semantic segmentation network: For different components, the image will be pixel-replaced first, only the target component will be retained, and then binarization will be performed. The target contour will be obtained using the corresponding function, and the center point will be obtained by further calculating the image center moment.

[0048] As a possible implementation, Figure 2 As shown, step 1 includes the following steps:

[0049] Step 1.1: Manually search for relevant drone videos using different keywords on video platforms such as Youku, iQiyi, Tencent Video, Bilibili, Douyin, and Kuaishou, delete small and blurry videos, and use Python scripts to decompose the remaining videos frame by frame, leaving only the video clips with drones;

[0050] Step 1.2: Open the labeling tool Labelme that comes with Python and load the image to be labeled. Select the three categories of the drone's rotor, body, and payload with continuous points, mark the three pixel values ​​0, 1, and 2, and save them. Figure 3-4 , a json file containing the coordinates of the labeled space points will be generated in the set folder path. Use the python script to convert the json file into a png label image. The label image obtained here is a grayscale image with values ​​of 0, 1, and 2;

[0051] Step 1.3: Flip, rotate, scale, crop, and shift the processed video clips and the corresponding annotated images. The final video semantic segmentation dataset of this article has 1,200 images, including 40 videos, with 30 images from each video. 900 of them are used for training, 300 for verification, and the rest are supplemented by image datasets.

[0052] As a possible implementation, step 2 includes the following steps:

[0053] Step 2.1: Construct a network model including a video semantic segmentation network branch, an input layer, the first convolution layer-the Nth convolution layer, a fully connected layer, an optical flow propagation module and a Softmax function. The network model includes M identical feature extraction branch networks. When M is 2, the model includes two identical feature extraction branch networks. Each feature extraction branch network includes 5 convolution layers, which are set as follows: the first convolution layer contains a 3*3 convolution and a 3*3 residual structure (bottleneck), and the number of output channels is 16; the second convolution layer has 2 bottlenecks with a convolution kernel size of 3*3 and a bottleneck with a convolution kernel of 5*5, and the number of output channels is 40; the third convolution layer consists of five bottlenecks with a convolution kernel size of 5*5, and the number of output channels is 96; the fourth convolution kernel size is 2 bottlenecks with a convolution kernel size of 5*5, and the number of output channels is 96; the fifth layer is a 3*3 pooling layer.

[0054] Step 2.2: Establish a feature map optical flow network based on the above network model: connect the PPM (Pyramid Pooling Module) network to collect context information after the fifth layer of the network, perform FAM (optical flow alignment) on the output of the third layer of the network and the PPM module and upsample to obtain the sixth layer. Similarly, perform FAM processing and upsample the output of the second layer of the network and the sixth layer to obtain the seventh layer. Perform FAM upsampling on the output of the first layer of the network and the seventh layer of the network and fuse the FAM of the sixth and seventh layers to obtain the eighth layer, and output the branch network result. Fuse another branch feature map to obtain the output result.

[0055] Step 2.3: Use the back propagation algorithm and gradient descent method to train the above network model. Adjust the weight value of each layer of the network by calculating the IoU (intersection over union) value of each category of drones. Repeat the iteration until the maximum mIoU value is obtained to save the network weight, and input the test set into the optimal network model to obtain the recognition result.

[0056] Preferably, step 3 comprises the following steps:

[0057] Step 3.1: Based on the recognition results, use pixel replacement to screen the key components that need to be processed;

[0058] Step 3.2: Based on the recognition results, the maximum connected domain algorithm is used for a single object to filter out the largest target and prevent the recognition results from affecting the center point calculation;

[0059] Step 3.3: Binarize the recognition result to obtain the target contour;

[0060] Step 3.4: Based on the above target contour, calculate the central moment to obtain the coordinates of the center point of the key component and visualize it.

[0061] As a possible implementation method, a method for real-time video semantic segmentation of key components of a drone based on deep learning provided in this embodiment includes the following steps:

[0062] a: Establish a dataset: After collecting all the existing Internet drone scene videos, splitting the frames and annotating the images, and performing data enhancement, we obtain several video semantic segmentation datasets with 30 consecutive frames as a scene segment, and divide them into training sets and test sets, such as Figure 2 As shown:

[0063] a1: Video collection: We manually searched for relevant drone videos on Youku, iQiyi, Tencent Video, Bilibili, Douyin, Kuaishou and other video platforms using different keywords, deleted small and blurry videos, and used Python scripts to split the remaining videos into 30 frames per frame, leaving only the pictures with drones. We aggregated the remaining pictures, removed duplicate pictures, and finally completed the image collection.

[0064] a2: Image annotation: Open the labeling tool Labelme that comes with Python and load the image to be labeled. Select the three categories of the drone's rotor, body, and payload with continuous points, mark the three pixel values ​​0, 1, and 2, and save them. Figure 3-4 , a json file containing the coordinates of the labeled space points will be generated in the set folder path. Use the python script to convert the json file into a png label image. The label image obtained here is a grayscale image with values ​​of 0, 1, and 2. In order to better see the labeling effect, it can be visualized with the RGB color correspondence.

[0065] a3: Data enhancement: Flip, rotate, scale, crop, and shift the processed video clips and the corresponding annotated images. The final video semantic segmentation dataset of this article has 1,200 images, including 40 videos, with 30 images from each video. 900 of them are used for training, 300 for verification, and the rest are supplemented by image datasets.

[0066] b: Build and train the video semantic segmentation network: Establish a network model including the video semantic segmentation network branch, input layer, N convolutional layers, optical flow propagation module and Softmax function, input the above training set into the network model, complete the training to obtain the optimal model, and then input the test set into the optimal model to obtain the recognition result, such as Figure 3 As shown:

[0067] b1: Define the model structure: Construct a network model including a video semantic segmentation network branch, an input layer, the first convolution layer-the Nth convolution layer, a fully connected layer, an optical flow propagation module and a Softmax function. When M is 2, the model includes two identical feature extraction branch networks. The first convolution layer contains a 3*3 convolution and a 3*3 residual structure (bottleneck), and the number of output channels is 16. The second convolution layer has two bottlenecks with a convolution kernel size of 3*3 and a bottleneck with a convolution kernel size of 5*5, and the number of output channels is 40. The third convolution layer consists of five bottlenecks with a convolution kernel size of 5*5, and the number of output channels is 96. The fourth convolution kernel size is 2 bottlenecks with a convolution kernel size of 5*5, and the number of output channels is 96. The fifth layer is a 3*3 pooling layer.

[0068] b2: Establish a feature map optical flow network based on the above network model: connect the PPM network to collect context information after the fifth layer of the network, perform FAM (optical flow alignment) on the output of the third layer of the network and the PPM module and upsample to obtain the sixth layer. Similarly, perform FAM processing and upsample the output of the second layer of the network and the sixth layer to obtain the seventh layer. Perform FAM upsampling on the output of the first layer of the network and the seventh layer of the network and fuse the FAM of the sixth and seventh layers to obtain the eighth layer, and output the result of the branch network. Fuse another branch feature map to obtain the output result. Figure 3 shown.

[0069] b3: Use the back propagation algorithm and gradient descent method to train the above network model. Adjust the weight value of each layer of the network by calculating the IoU value of each category of drones. Repeat the iteration until the maximum mIoU value is obtained to save the network weight. Input the test set into the optimal network model to obtain the recognition result.

[0070] c: For the semantic segmentation results of the aircraft, the key components need to be further processed to obtain the center point coordinates for application, such as Figure 4 As shown;

[0071] c1: Filter the key components that need to be processed by using pixel replacement based on the recognition results;

[0072] c2: Based on the recognition results, the maximum connected domain algorithm is used for a single object to filter out the largest target and prevent the recognition results from affecting the center point calculation;

[0073] c3: Binarize the recognition result to get the target contour;

[0074] c4: Based on the above target contour, the central moment is calculated to obtain the coordinates of the center points of key components and visualize them.

[0075] This embodiment compares the parameter quantities and calculation amounts corresponding to different backbones, as shown in Table 1, the parameter quantities and total calculation amounts after SFNet replaces the backbone, as shown in Table 2, the results after training the network, as shown in Table 3, and the network training results after adding the video semantic segmentation framework, as shown in Table 4:

[0076] Table 1 Parameters and calculations corresponding to different backbones

[0077]

[0078] Table 2 Comparison of backbone parameter replacement and total calculation

[0079]

[0080] Table 3 Comparison of results after training the network

[0081]

[0082] Table 4 Network training results after adding video semantic segmentation framework

[0083]

[0084]

[0085] According to Table 4, it can be concluded that the recognition accuracy of the recognition processing in the embodiment of the present invention is higher and the recognition rate is faster; Figure 5 As shown, Figure 5 Group (a) is a multi-scale target test of the Td2-SFNet network. It can be seen that during the high-altitude flight of the target, the target gradually becomes smaller, but the segmentation accuracy does not decrease; Group (b) is a drone test with a glacier as the background. The visualization results show that drones with similar colors to the background are still accurately segmented; Group (c) is an occlusion test. The hand occludes part of the drone body, which does not affect the segmentation effect. In summary, the video semantic segmentation and recognition algorithm provided by the present invention realizes automatic and rapid recognition of drones, gets rid of tedious, time-consuming and labor-intensive manual recognition, and uses the video semantic segmentation framework and lightweight backbone to reduce the amount of calculation; at the same time, by using the flow alignment module of optical flow transmission, the recognition accuracy of edge information is improved, and the recognition accuracy rate is greatly improved.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0087] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A real-time video semantic segmentation method for key components of drones based on deep learning, characterized in that: The steps include: Step 1, create a data set: Based on network resources, all scene images where drones appear are collected, and after image annotation and data enhancement of target components, several video semantic segmentation datasets are obtained, with images of a specified number of consecutive frames as a scene segment. Step 2: Build and train the video semantic segmentation network: The video semantic segmentation network includes M feature extraction network branches, each of which includes multiple convolutional layers and an optical flow propagation module for optical flow alignment and upsampling. The output feature maps of the M feature extraction branches are spliced ​​and fused, and then input into a fully connected layer with a Softmax function after passing through an encoding network and an attention module to output probability information of each target component. The video semantic segmentation network is trained based on the video semantic segmentation dataset, and the optimal model is obtained after the training is completed; Step 3, inputting the video sequence to be identified into the optimal model obtained in step 2, and obtaining the recognition result based on its output, wherein the length of the video sequence to be identified is the same as the number of frames of the scene fragment of the video semantic segmentation dataset; Perform key point processing on the recognition results: replace pixels on the recognition results, retain the target component, perform binarization processing, and then obtain the contour of the target component based on the contour acquisition process; calculate the image center moment based on the contour of the target component to obtain the center point of the target component and output it visually; Among them, each feature extraction network branch is specifically set as: The first convolutional layer contains a 3*3 convolution and a 3*3 residual structure, and the number of output channels is 16; The second convolution layer has two residual structures with a convolution kernel size of 3*3 and one residual structure with a convolution kernel of 5*5, and the number of output channels is 40; The third convolutional layer consists of five residual structures with a convolution kernel size of 5*5 and 96 output channels; The size of the fourth convolution kernel is 5*5 with 2 residual structures, and the number of output channels is 96; The fifth layer is a 3*3 pooling layer; The fifth layer of the network is connected to the pyramid pooling network, and the output of the third layer of the network and the pyramid pooling network are optically aligned and upsampled to obtain the sixth layer, and the second layer of the network and the output of the sixth layer are optically aligned and upsampled to obtain the seventh layer, and then the first layer of the network and the output of the seventh layer of the network are optically aligned and upsampled, and then the optical flow alignment results of the sixth and seventh layers are fused to obtain the output feature map of the feature extraction network branch.

2. The method according to claim 1, characterized in that In step 1, the image annotation of the target component is specifically as follows: the target component of the drone is framed with continuous points, and different pixel values ​​are used for target components of different categories.

3. The method according to claim 1, characterized in that In step 1, data enhancement includes: image flipping, rotation, scaling, cropping and shifting.

4. The method according to any one of claims 1 to 3, characterized in that In step 2, when training the video semantic segmentation network, the back propagation algorithm and the gradient descent method are used for training, and the weight value of each layer of the network is adjusted by calculating the intersection-and-union ratio of each category of drones, and it is iterated repeatedly until the maximum intersection-and-union ratio is obtained.

5. The method according to claim 1, characterized in that In step 3, performing key point processing on the recognition result specifically includes: Filter the target parts by using pixel replacement based on the recognition results; Based on the recognition results, the maximum connected domain algorithm is used for a single object to filter out the largest target and prevent the recognition results from affecting the center point calculation. Binarization is performed based on the recognition result to obtain the outline of the target component; The image center moment is calculated based on the contour of the target part to obtain the center point of the target part and visualize the output.

Images