Optical detection method and system for small motion unmanned aerial vehicle based on deep neural network

Abstract

The application provides a small motion unmanned aerial vehicle optical detection method and system based on a deep neural network, first, a video is acquired, and frame extraction is performed on the video; then, a Vibe algorithm is used for dynamic target detection on the video image obtained after frame extraction, so that the coordinates of the dynamic target in the image are obtained; the position of the dynamic target in the video is image cut according to the coordinates of the dynamic target, so that a small-size image containing the dynamic target is obtained; finally, the small-size image is input into an improved YOLOv7 model, the position information of the dynamic target is acquired, the position information of the dynamic target is regressed and mapped to the video, and the detection of the unmanned aerial vehicle target is completed. The beneficial effect of the application is that the detection accuracy of the "low and small" unmanned aerial vehicle target is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of optical inspection methods for small mobile drones, and more particularly to an optical inspection method and system for small mobile drones based on deep neural networks. Background Technology

[0002] In recent years, the rapid development of the drone market and the increasing number of drones have brought about numerous security and privacy issues. As the government strengthens its governance and control of unauthorized drone flights, drone countermeasures technology plays a crucial role in security. Real-time detection and accurate identification are the prerequisites and foundation for effective countermeasures. Currently, mainstream drone detection and identification technologies include active radar detection, passive radio frequency sensing, sound recognition, and photoelectric vision detection. Among these, photoelectric vision detection technology has advantages such as low power consumption, no radiation, and low cost, and can effectively distinguish between non-drone targets such as birds and radiation sources. Therefore, machine vision-based drone target detection technology has become an important component and development direction of low-altitude security.

[0003] Traditional target detection algorithms suffer from overly complex detection processes, severely impacting efficiency and leading to their gradual withdrawal from mainstream adoption. Deep learning-based target detection algorithms are increasingly demonstrating their advantages, with the YOLO series serving as a prime example of one-stage algorithms. The newer YOLOv7 algorithm exhibits superior performance in both accuracy and speed. However, despite its excellent speed and accuracy, the YOLOv7 algorithm suffers from relatively low accuracy for small targets due to its deep downsampling layers, including numerous pooling and strided convolutional layers. Furthermore, the small size and high maneuverability of low-altitude UAVs present a technical bottleneck, resulting in a high false detection rate for these small, low-altitude targets, hindering practical application requirements. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method and system for optical detection of small mobile drones based on deep neural networks, aimed at improving the detection accuracy of "unauthorized" drones. The method for optical detection of small mobile drones based on deep neural networks mainly includes: S1: Acquire video and extract frames from the video; S2: The Vibe algorithm is used to detect moving targets in the video images obtained after frame extraction, and the coordinates of moving targets in the image are obtained. S3: Based on the coordinates of the moving target, perform image segmentation on the location of the moving target in the video of step S1 to obtain a small-sized image containing the moving target; S4: Input the small-sized image into the improved YOLOv7 model to obtain the position information of the moving target, and regress the position information of the moving target onto the video in step S1 to complete the detection of the UAV target; The improved YOLOv7 model includes the original CBS, MP, E-ELAN, SPPCSPC, and UPSample modules, as well as the improved ELAN-CA module in the backbone network and the MP-SPD module in the neck network. ELAN-CA is a module that incorporates a CA coordinate attention mechanism into the MP module; MP-SPD is a module generated by incorporating an SPDConv convolutional block into the MP module. After the small-sized image enters the backbone network, it first undergoes several layers of CBS convolution, then the MP and ELAN-CA modules are used to extract features from moving targets; then, the features are fused in the neck network, and finally fed into the head network for detection.

[0005] Furthermore, the process of moving target detection is as follows: Using the video sequence as input, the video sequence is first decoded to obtain image frames. A background model is established for the first frame image. Then, the next frame image is grayscaled. The ViBe algorithm is then used to calculate the background model of the current frame and the grayscale image of the next frame to obtain an image containing a foreground target region. The image containing the foreground target region is then subjected to morphological processing. After binarization denoising, the foreground target is selected by bounding box. Finally, the coordinates of the selected moving target are mapped to the video output in step S1.

[0006] Furthermore, the coordinate attention mechanism in the ELAN-CA module includes two processes: coordinate information embedding and coordinate attention generation; Coordinate information embedding: by decomposing the global pooling operation into a one-to-one one-dimensional feature encoding operation; utilizing two spatially expansive pooling kernels ( H ,1) and (1, W ) for feature input Each channel is encoded and aggregated along two directions to form a pair of position-aware feature maps; Coordinate attention generation: Coordinate information is embedded into the extracted feature map concatenation, and then the spatial information in the horizontal and vertical directions is encoded through calculation to obtain intermediate feature maps. k Then The input is split into two independent tensors along the spatial dimension, and then subjected to convolution and activation functions to make the two independent tensors correlate with the input. With the same number of channels, the attention weights are finally obtained.

[0007] Furthermore, intermediate feature mapping kThe calculation formula is as follows:

[0008] In the formula The meaning is to output along the horizontal direction. and Assembled along the spatial passage. It is a non-linear activation function. It is a 1×1 convolution transformation function.

[0009] Furthermore, the MP-SPD module adds an SPDConv convolutional block consisting of a spatial-to-depth layer and a non-staggered convolutional layer starting from the second branch of the MP module; the MP-SPD consists of two branch structures, one of which first goes through MaxPool and then a CBS convolutional block; the other branch first goes through an SPDConv convolutional module and then through two CBS convolutional blocks; finally, the two branches are concatenated together by Concat.

[0010] An optical inspection system for a small mobile drone based on a deep neural network includes: The video acquisition module is used to acquire video and extract frames from the video. The moving target coordinate acquisition module is used to perform moving target detection on the video image obtained after frame extraction using the Vibe algorithm, and obtain the coordinates of the moving target in the image. The small-size image acquisition module is used to perform image segmentation on the location of the moving target in the video according to the coordinates of the moving target, so as to obtain a small-size image containing the moving target. The drone target detection module is used to input the small-sized image into the improved YOLOv7 model to obtain the position information of the moving target, and regress the position information of the moving target onto the video in step S1 to complete the detection of the drone target; The improved YOLOv7 model includes the original CBS, MP, E-ELAN, SPPCSPC, and UPSample modules, as well as the improved ELAN-CA module in the backbone network and the MP-SPD module in the neck network. ELAN-CA is a module that incorporates a CA coordinate attention mechanism into the MP module; MP-SPD is a module generated by incorporating an SPDConv convolutional block into the MP module. After the small-sized image enters the backbone network, it first undergoes several layers of CBS convolution, then the MP and ELAN-CA modules are used to extract features from moving targets; then, the features are fused in the neck network, and finally fed into the head network for detection.

[0011] Furthermore, the moving target detection process is as follows: taking the video sequence as input, the video sequence is first decoded to obtain image frames. A background model is established for the first frame image. Then, the next frame image is grayscaled. The ViBe algorithm is then used to calculate the background model of the current frame and the grayscale image of the next frame to obtain an image containing a foreground target region. The image containing the foreground target region is then subjected to morphological processing. After binarization denoising, the foreground target is selected by bounding box. Finally, the coordinates of the selected moving target are mapped to the video output.

[0012] Furthermore, the coordinate attention mechanism in the ELAN-CA module includes two processes: coordinate information embedding and coordinate attention generation; Coordinate information embedding: by decomposing the global pooling operation into a one-to-one one-dimensional feature encoding operation; utilizing two spatially expansive pooling kernels ( H ,1) and (1, W ) for feature input Each channel is encoded and aggregated along two directions to form a pair of position-aware feature maps; Coordinate attention generation: Coordinate information is embedded into the extracted feature map concatenation, and then the spatial information in the horizontal and vertical directions is encoded through calculation to obtain intermediate feature maps. k Then The input is split into two independent tensors along the spatial dimension, and then subjected to convolution and activation functions to make the two independent tensors correlate with the input. With the same number of channels, the attention weights are finally obtained.

[0013] Furthermore, intermediate feature mapping k The calculation formula is as follows:

[0014] In the formula The meaning is to output along the horizontal direction. and Assembled along the spatial passage. It is a non-linear activation function. It is a 1×1 convolution transformation function.

[0015] Furthermore, the MP-SPD module adds an SPDConv convolutional block consisting of a spatial-to-depth layer and a non-staggered convolutional layer starting from the second branch of the MP module; the MP-SPD consists of two branch structures, one of which first goes through MaxPool and then a CBS convolutional block; the other branch first goes through an SPDConv convolutional module and then through two CBS convolutional blocks; finally, the two branches are concatenated together by Concat.

[0016] The beneficial effects of the technical solution provided by this invention are as follows: This invention acquires video and extracts frames from the video; it can combine the features of the UAV in the time domain, spatial domain, and frequency domain to eliminate false UAVs caused by electromagnetic environment influences, and mine real UAV target information, thereby improving the detection accuracy of UAVs and reducing the false alarm rate. Then, the Vibe algorithm is used to detect moving targets in the video images obtained after frame extraction, and the coordinates of the moving targets in the images are obtained; based on the coordinates of the moving targets, the image of the moving target in the video is segmented to obtain a small-sized image containing the moving targets; finally, the small-sized image is input into the improved YOLOv7 model to obtain the position information of the moving targets, and the position information of the moving targets is regressed and mapped back to the video to complete the detection of UAV targets; this invention incorporates the motion features, color features, and shape features of UAVs into the algorithm calculation, increases the utilization rate of UAV feature information, reduces the impact of background noise caused by complex environments, improves the signal-to-noise ratio of UAV targets in the image, and ultimately improves the detection and recognition accuracy of UAVs and reduces the false detection rate. Finally, the detection accuracy of "low-motion" UAV targets is significantly improved. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of an optical detection method for a small mobile drone based on a deep neural network, as described in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the construction of the moving target detection model in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of coordinate information embedding in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the SPD-Conv layer with a scaling factor of 2 in an embodiment of the present invention.

[0021] Figure 5 This is a diagram of the improved YOLOv7 network structure in an embodiment of the present invention. Detailed Implementation

[0022] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] This invention provides an optical detection method for small mobile unmanned aerial vehicles (UAVs) based on deep neural networks, implemented by fusing Vibe and an improved YOLOv7 anti-low-altitude UAV target detection method. First, moving target preprocessing is performed, using the Vibe algorithm to extract moving targets and then segmenting the moving target image. Second, the SPD-Conv module and coordinate attention mechanism are introduced to improve the YOLOv7 network structure, and a two-level detection architecture fusing Vibe and the improved YOLOv7 is proposed. The improved YOLOv7 is used as the network model to detect the image processed by Vibe. Based on the positional and size relationship between the original image and the processed image, the coordinates of the detected targets are mapped back to the original image, thus completing the target detection and extraction.

[0024] Please refer to Figure 1 , Figure 1 This is a flowchart of an optical detection method for a small mobile drone based on a deep neural network, as described in an embodiment of the present invention, specifically including: S1: Use a camera to capture video and extract frames from the video. S2: The Vibe algorithm is used to detect moving targets in the video images obtained after frame extraction, and the coordinates of moving targets in the images are obtained. S3: Based on the coordinates of the moving target, perform image segmentation on the location of the moving target in the video of step S1 to obtain a small-sized image containing the moving target; S4: Input the small-sized image into the improved YOLOv7 model to obtain the position information of the moving target, and regress the position information of the moving target onto the video in step S1 to complete the detection of the UAV target; The improved YOLOv7 model includes the original CBS, MP, E-ELAN, SPPCSPC, and UPSample modules, as well as the improved ELAN-CA module for the backbone network and the MP-SPD module for the neck network. ELAN-CA is a module that incorporates a CA coordinate attention mechanism into the MP module. MP-SPD is a module generated by incorporating an SPDConv convolutional block into the MP module. When a small image enters the backbone network, it first undergoes four layers of CBS convolution, then the MP and ELAN-CA modules are used to extract features from moving targets. Next, the image enters the neck network for feature fusion, and finally, it is fed into the head network for detection. The MP-SPD module incorporates an SPDConv convolutional block consisting of a spatial-to-depth layer and a non-staggered convolutional layer starting from the second branch of the MP module. MP-SPD consists of two branches: one branch first undergoes max pooling followed by a CBS convolutional block; the other branch first undergoes an SPDConv convolutional module followed by two CBS convolutional blocks; finally, the two branches are concatenated using Concat.

[0025] The flowchart for moving target detection is as follows: Figure 2 As shown, the video sequence is used as input. First, the video sequence is decoded to obtain image frames. The background model is established for the first frame image. Then, the next frame image is grayscaled. The ViBe algorithm is then used to calculate the background model of the current frame and the grayscale image of the next frame to obtain the moving target detection result. The image containing the detected foreground target region is subjected to morphological processing. After binarization and denoising, the foreground target is selected by bounding box. Finally, the bounding box coordinates are mapped to the original RGB image for output.

[0026] ELAN structure with coordinate attention mechanism: While the original YOLOv7 algorithm boasts advantages in detection speed and accuracy, it also suffers from drawbacks such as susceptibility to environmental factors like brightness and lighting conditions, and poor performance with small targets and complex backgrounds. Therefore, to enhance the model's ability to detect low-altitude small UAVs under varying weather conditions and backgrounds, and to improve the network's attention to important features, this invention incorporates a coordinate attention mechanism into the ELAN module, proposing an efficient and lightweight ELAN-CA module.

[0027] Coordinate attention (CA) is a lightweight and efficient attention mechanism that incorporates positional information into channel attention and fuses inter-channel relationships and positional information, thereby enabling the model to more accurately determine the target region. It consists of two processes: coordinate information embedding and coordinate attention generation.

[0028] Coordinate information embedding primarily addresses the problem of capturing spatial structure and preserving positional information in visual tasks. For example... Figure 3 As shown, H, W, CThese represent height, width, and number of channels, respectively. Input represents the input, and Output represents the output. Residual refers to residual connections, a technique introduced in neural networks. In the implementation of the CA coordinate attention mechanism, residual connections are used to pass information across different layers to help the network better capture the spatial coordinate relationships in the input data. X AvgPool and Y AvgPool represent average pooling operations on the X-axis and Y-axis of the feature maps, respectively. Concat+Conv2d represents concatenating two feature maps and performing the operation through two-dimensional convolution (Conv2d). BatchNorm+Non-linear represents the combination of batch normalization and non-linear transformation. Sigmoid is a type of activation function. Re-weight refers to the operation of re-weighting features. CA, by decomposing the global pooling operation into a one-to-one one-dimensional feature encoding operation, greatly improves the problem that traditional attention mechanisms are unable to retain positional information, enabling accurate capture of spatial information and thus achieving precise spatial capture of remote interaction.

[0029] The principle of CA is as follows: by utilizing two spatially defined pooling kernels ( H ,1) and (1, W ) for feature input Each channel is encoded and then aggregated along two directions to form a pair of position-aware feature maps. H This represents the height, which is the number of rows in the spatial dimension of the feature input; W This represents the width, i.e., the number of columns in the spatial dimension of the feature input. c The channel's output in height and width is shown below:

[0030] In the formula, For channel No. Line 1 The values ​​in the column, For channel No. Liede The value of the row. For vertical output, This is the output along the horizontal direction.

[0031] Coordinate attention generation first embeds coordinate information into the extracted feature map concatenation, then encodes spatial information in the horizontal and vertical directions through calculation, and obtains intermediate feature maps. k The specific calculation formula is as follows:

[0032] In the formula The meaning is to output along the horizontal direction. and Assembled along the spatial passage. It is a non-linear activation function. It is a 1×1 convolution transformation function.

[0033] Then Split into two independent tensors along the spatial dimension and , C This represents the number of channels in the input feature. r This represents the scaling factor, used to control the reduction ratio of the feature map dimensions. It then undergoes convolutional transformation and activation functions to... and With input With the same number of channels, the final result is and The specific formula for the attention weight is as follows:

[0034] In the formula and It is a 1×1 convolution transformation. It is the sigmoid activation function.

[0035] Integrating CA into the ELAN structure not only yields rich gradient information but also allows for more accurate localization of important features. Compared to the original ELAN structure, it reduces the number of parameters, making feature extraction more efficient, while also mitigating the impact of different lighting and background conditions on low-altitude UAV detection performance, thus increasing the model's detection capabilities in various environments.

[0036] Introducing the SPDConv module to improve the MPConv layer: The overall network architecture of YOLOv7 contains numerous strided convolutional and pooling layers in the CBS layer of the backbone network and the MPConv layer of the head network. This can easily lead to the loss of fine-grained information during feature learning, resulting in poor performance in tasks involving low-resolution images and small targets. Low-altitude small drones are often considered small targets in detection tasks, and the cropped images processed by Vibe are low-resolution images. Therefore, the SPDConv module is introduced to improve the MPConv layer of the YOLOv7 head network. The schematic diagram of the SPDConv module is shown below. Figure 4 As shown, ( i,j () represents the spatial coordinates of the input feature map, that is, the row and column of the feature map. Given a feature map X, such as Figure 4As shown in (a), the size is S×S×C 1, S The dimension of the feature map. C 1 represents the number of input channels. The formula for dividing the feature map into sub-feature maps is as follows:

[0037] Given any feature map X, the subgraph f x，y Depend on i + x and j + y All entries divisible by a fixed ratio X ( i,j Therefore, each subgraph is downsampled by a scaling factor for X. Figure 4 Examples (a), (b), and (c) are given when scale=2, resulting in 4 sub-feature maps. f 0，0 , f 0，1 , f 1，0 , f 0，scale-1 Each has a shape Then, X is downsampled by a factor of 2. Next, these sub-feature maps are concatenated along the channel dimension to obtain a feature map X', as shown below. Figure 4 As shown in (d), its spatial dimension is reduced by a scaling factor, while its channel dimension is increased by a scaling factor of 2. In other words, SPD reduces the feature map... X(S,S,C 1 ) Transform into intermediate feature map . Figure 4 (e) is an addition with a feature transformation after the feature transformation. C A non-staggered convolutional layer with a stride of 1 and 2 filters, wherein and further Figure 4 (d) Transform into Figure 4 (e) .

[0038] The improved YOLOv7 network structure diagram is as follows: Figure 5 As shown in the figure, 1 represents the input, 2 represents the backbone network, 3 represents the feature fusion network Neck, and 4 represents the detection head.

[0039] The CBS module consists of a convolutional layer (Conv), a batch normalization layer (BN), and a SiLU activation function. Its main purpose is to enhance the network's feature representation capabilities. The CBS modules of different colors in the diagram differ in their kernel size (k) and stride (s). A convolution with k = 1 and s = 1 is mainly used to change the number of channels; a convolution with k = 3 and s = 1 is mainly used for feature extraction; and a convolution with k = 3 and s = 2 is used for downsampling.

[0040] SPPCSPC (Spatial Pyramid Pooling, Cross Stage Partial Channel) comprises two parts: Spatial Pyramid Pooling (SPP) and Cross Stage Partial Channel (CSPC). The SPP module divides the feature map into sub-regions of different scales for pooling, thus creating a fixed-size feature map. The CSPC module divides the feature map into two parts: one part undergoes convolution, and the other part is output directly. Through these two sub-modules, SPPCSPC can reduce computation by half while maintaining high accuracy.

[0041] MaxPool represents max pooling.

[0042] Cat (Concat) represents the concatenation operation, which is used to concatenate different feature maps along the channel dimension.

[0043] The `add` operation is used to add two feature maps or tensors of the same shape element by element, that is, to add elements at corresponding positions.

[0044] ELAN-CA is a computational module improved from the Extended Efficient Layer Aggregation Network (ELAN), which introduces a CA (Coordinate Attention) mechanism into the original ELAN module.

[0045] E-ELAN (Extened-Efficient Layer Aggregation Network) is an extended efficient layer aggregation network, which is a computational block derived from the ELAN structure.

[0046] REP stands for RepConv structure. Its basic idea is to use a multi-branch residual structure during training and a single-path structure during inference.

[0047] UPSample refers to the upsampling operation.

[0048] MP module: A module with two branches, one of which goes through a max pooling operation followed by a CBS convolutional block, and the other of which goes through two CBS convolutional blocks with different kernel sizes and strides. Finally, the two branches are connected by a concat operation. This structure is called an MP module.

[0049] MP-SPD is a module created by adding a Spatial Pyramid Dilated Convolution (SPDConv) to the second branch of the original MP module.

[0050] The SPDConv module consists of spatial-to-depth (SPD) layers and non-staggered convolutional (NS-Conv) layers, and it performs better than traditional deep learning models in handling tasks with low-resolution images and small objects. First, the SPD layer divides the image into layers of size... intermediate feature input Cut out a series of sub-feature maps , The dimensions are all ( S / scale, S / scale, C 1), is the original feature input. Downsampling is then performed. Then, all sub-feature maps are concatenated along the channel dimension to obtain the feature map. ( S / scale, S / scale, scale2 C 1) Feature Map Sent to C Output feature map after two non-stepping convolutional layers (NS-Conv) with two filters ( S / scale, S / scale, C 2) Non-staggered convolutional layers retain as much discriminative information as possible, among which, S This represents the spatial dimensions of the input feature map, specifically its height and width. C 1 represents the number of channels in the input feature map; Scale represents the scaling factor, used to control the size of the sub-feature map and the degree of downsampling. S / scale indicates that the size of each sub-feature map is the size of the original feature map divided by the scaling factor.

[0051] An optical inspection system for a small mobile drone based on a deep neural network includes: The video acquisition module is used to acquire video and extract frames from the video. The moving target coordinate acquisition module is used to perform moving target detection on the video image obtained after frame extraction using the Vibe algorithm, and obtain the coordinates of the moving target in the image. The small-size image acquisition module is used to perform image segmentation on the location of the moving target in the video according to the coordinates of the moving target, so as to obtain a small-size image containing the moving target. The drone target detection module is used to input the small-sized image into the improved YOLOv7 model to obtain the position information of the moving target, and then regress and map the position information of the moving target into the video in step S1 to complete the detection of the drone target.

[0052] The beneficial effects of this invention are as follows: This invention acquires video and extracts frames from the video; it can combine the features of the UAV in the time domain, spatial domain, and frequency domain to eliminate false UAVs caused by electromagnetic environment influences, and mine real UAV target information, thereby improving the detection accuracy of UAVs and reducing the false alarm rate. Then, the Vibe algorithm is used to detect moving targets in the video images obtained after frame extraction, and the coordinates of the moving targets in the images are obtained; based on the coordinates of the moving targets, the image of the moving target in the video is segmented to obtain a small-sized image containing the moving targets; finally, the small-sized image is input into the improved YOLOv7 model to obtain the position information of the moving targets, and the position information of the moving targets is regressed and mapped back to the video to complete the detection of UAV targets; this invention incorporates the motion features, color features, and shape features of UAVs into the algorithm calculation, increases the utilization rate of UAV feature information, reduces the impact of background noise caused by complex environments, improves the signal-to-noise ratio of UAV targets in the image, and ultimately improves the detection and recognition accuracy of UAVs and reduces the false detection rate. Finally, the detection accuracy of "low-motion" UAV targets is significantly improved.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An optical detection method for a small mobile unmanned aerial vehicle (UAV) based on deep neural networks, characterized in that: include: S1: Acquire video and extract frames from the video; S2: The Vibe algorithm is used to detect moving targets in the video images obtained after frame extraction, and the coordinates of moving targets in the image are obtained. S3: Based on the coordinates of the moving target, perform image segmentation on the location of the moving target in the video of step S1 to obtain a small-sized image containing the moving target; S4: Input the small-sized image into the improved YOLOv7 model to obtain the position information of the moving target, and regress the position information of the moving target onto the video in step S1 to complete the detection of the UAV target; The original modules of the improved YOLOv7 model include CBS, MP, E-ELAN, SPPCSPC and UPSample modules, as well as the improved ELAN-CA module for the backbone network and the MP-SPD module for the neck network; the ELAN-CA module is an E-ELAN module with CA coordinate attention mechanism added. The MP-SPD module is generated by adding an SPDConv convolutional block to the MP module. The MP-SPD module adds an SPDConv convolutional block consisting of a spatial-to-depth layer and a non-staggered convolutional layer starting from the second branch of the MP module. The MP-SPD consists of two branches: one branch first passes through MaxPool followed by a CBS convolutional block; the other branch first passes through an SPDConv convolutional module followed by two CBS convolutional blocks; finally, the two branches are concatenated together using Concat. After the small-sized image enters the backbone network, it first goes through several layers of CBS convolution, and then uses MP and ELAN-CA modules to extract features from the moving target. Then, the features are fused into the neck network and finally fed into the head network for detection. The coordinate attention mechanism in the ELAN-CA module includes two processes: coordinate information embedding and coordinate attention generation. Coordinate information embedding: by decomposing the global pooling operation into a one-to-one one-dimensional feature encoding operation; Each channel of the feature input is encoded using two spatially defined pooling kernels (H, 1) and (1, W), and then aggregated along two directions to form a pair of position-aware feature maps. Coordinate attention generation: The coordinate information is embedded into the extracted feature map and connected, and then the spatial information in the horizontal and vertical directions is encoded by calculation to obtain the intermediate feature map k; The intermediate feature map k is split into two independent tensors along the spatial dimension, and then subjected to convolution and activation functions to ensure that the two independent tensors have the same number of channels as the input. Finally, the attention weights are obtained, as follows: Mapping intermediate features Split into two independent tensors along the spatial dimension and C represents the number of channels in the input feature, and r represents the scaling factor; after convolution and activation functions, ... and With input With the same number of channels, we get and The specific formula for the attention weight is as follows: In the formula and It is a 1×1 convolution transformation. It is the sigmoid activation function.

2. The optical detection method for a small mobile unmanned aerial vehicle based on a deep neural network as described in claim 1, characterized in that: The process of moving target detection is as follows: Using the video sequence as input, the video sequence is first decoded to obtain image frames. A background model is established for the first frame image. Then, the next frame image is grayscaled. The Vibe algorithm is then used to calculate the background model of the current frame and the grayscale image of the next frame to obtain an image containing a foreground target region. The image containing the foreground target region is then subjected to morphological processing. After binarization denoising, the foreground target is selected by bounding box. Finally, the coordinates of the selected moving target are mapped to the video output in step S1.

3. The optical detection method for a small mobile unmanned aerial vehicle based on a deep neural network as described in claim 1, characterized in that: The formula for calculating the intermediate feature map k is as follows: In the formula The meaning is to output along the horizontal direction. and Assembled along the spatial passage. It is a non-linear activation function. It is a 1×1 convolution transformation function.

4. An optical inspection system for a small mobile unmanned aerial vehicle (UAV) based on a deep neural network, characterized in that: include: The video acquisition module is used to acquire videos and extract frames from them. The moving target coordinate acquisition module is used to perform moving target detection on the video image obtained after frame extraction using the Vibe algorithm, and obtain the coordinates of the moving target in the image. The small-size image acquisition module is used to perform image segmentation on the location of the moving target in the video according to the coordinates of the moving target, so as to obtain a small-size image containing the moving target. The drone target detection module is used to input the small-sized image into the improved YOLOv7 model to obtain the position information of the moving target, and then regress and map the position information of the moving target into the video of the video acquisition module to complete the detection of the drone target. The original modules of the improved YOLOv7 model include CBS, MP, E-ELAN, SPPCSPC, and UPSample modules, as well as the improved ELAN-CA module for the backbone network and the MP-SPD module for the neck network. The ELAN-CA module is an E-ELAN module with the addition of a CA coordinate attention mechanism. The MP-SPD module is a module generated by adding an SPDConv convolutional block to the MP module. The MP-SPD module is generated by adding an SPDConv convolutional block to the MP module. The MP-SPD module adds an SPDConv convolutional block consisting of a spatial-to-depth layer and a non-staggered convolutional layer starting from the second branch of the MP module. The MP-SPD consists of two branches: one branch first passes through MaxPool followed by a CBS convolutional block; the other branch first passes through an SPDConv convolutional module followed by two CBS convolutional blocks; finally, the two branches are concatenated together using Concat. After the small-sized image enters the backbone network, it first goes through several layers of CBS convolution, and then uses MP and ELAN-CA modules to extract features from the moving target. Then, the features are fused into the neck network and finally fed into the head network for detection. The coordinate attention mechanism in the ELAN-CA module includes two processes: coordinate information embedding and coordinate attention generation. Coordinate information embedding: by decomposing the global pooling operation into a one-to-one one-dimensional feature encoding operation; Each channel of the feature input is encoded using two spatially defined pooling kernels (H, 1) and (1, W), and then aggregated along two directions to form a pair of position-aware feature maps. Coordinate attention generation: The coordinate information is embedded into the extracted feature map and connected, and then the spatial information in the horizontal and vertical directions is encoded by calculation to obtain the intermediate feature map k; The intermediate feature map k is split into two independent tensors along the spatial dimension, and then subjected to convolution and activation functions to ensure that the two independent tensors have the same number of channels as the input. Finally, the attention weights are obtained, as follows: Mapping intermediate features Split into two independent tensors along the spatial dimension and C represents the number of channels in the input feature, and r represents the scaling factor; after convolution and activation functions, ... and With input With the same number of channels, we get and The specific formula for the attention weight is as follows: In the formula and It is a 1×1 convolution transformation. It is the sigmoid activation function.

5. The optical inspection system for a small mobile unmanned aerial vehicle based on a deep neural network as described in claim 4, characterized in that: The moving target detection process is as follows: taking the video sequence as input, the video sequence is first decoded to obtain image frames. A background model is established for the first frame image. Then, the next frame image is grayscaled. The Vibe algorithm is then used to calculate the background model of the current frame and the grayscale image of the next frame to obtain the image containing the detected foreground target region. The image containing the detected foreground target region is then subjected to morphological processing. After binarization denoising, the foreground target is selected by bounding box. Finally, the coordinates of the selected moving target are mapped to the video output.

6. The optical inspection system for a small mobile unmanned aerial vehicle based on a deep neural network as described in claim 4, characterized in that: The formula for calculating the intermediate feature map k is as follows: In the formula The meaning is to output along the horizontal direction. and Assembled along the spatial passage. It is a non-linear activation function. It is a 1×1 convolution transformation function.

Images