A disguised person target detection method based on improved Scaled-YOLOv4

Abstract

The present invention discloses a method for detecting a disguised person based on an improved Scaled-YOLOv4. The method pre-processes the pre-acquired active images of the disguised person and divides them into a training set and a test set. The method constructs an extended cross-stage network Scaled-YOLOv4 model including a backbone network, a path aggregation network and a detection head. The method designs a top-down path feature enhancement network structure including two receptive field blocks and two effective channel attention modules in the Scaled-YOLOv4 model to replace the original path aggregation network to form an improved Scaled-YOLOv4 model. The method detects the target of the disguised person based on the improved Scaled-YOLOv4 model. The optimization of the network structure proposed by the present invention can enhance the feature map of the disguised target and obtain the position information of the disguised target in the original image to reduce the missed detection rate and the false detection rate of the network. The network structure is stable and has strong practicality, which further improves the detection accuracy of multi-scale targets in the camouflaged target image in the woods.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology and machine vision, and in particular relates to a disguised person target detection method based on improved Scaled-YOLOv4. Background Art

[0002] In today's military battlefield, the more battlefield information is obtained, the easier it is to win the battle. Accurate detection of camouflaged targets on the battlefield is an important basis for obtaining battlefield information and making important decisions.

[0003] In recent years, drone technology has been widely used in various fields, especially in the military field. With its small size and strong flexibility, drones are widely used in military reconnaissance as a new type of auxiliary combat equipment on the military battlefield, playing a vital role in obtaining information about camouflaged personnel targets. In addition, camouflage technology has also been greatly developed. The camouflage clothing of military personnel helps them to blend in with the background, making it impossible to distinguish between objects in the military background and camouflaged personnel targets by the naked eye, which brings great challenges to the detection of camouflaged personnel targets.

[0004] Currently, there is limited research on the detection of camouflaged targets, and most camouflaged target detection tasks treat it as a target segmentation task, which uses a convolutional neural network to search for edge detail information of the camouflaged target to determine the camouflaged target. However, for the camouflaged personnel target image captured by the drone on the military battlefield, it is not necessary to determine all the edges of the camouflaged target so finely. It is only necessary to determine the approximate hiding position of the camouflaged personnel target and then accurately hit the position. Therefore, in this patent, the detection of camouflaged personnel targets is regarded as a target detection task to complete the detection and selection of camouflaged targets.

[0005] The current mainstream target detection methods based on deep learning are divided into two categories: one-stage target detection models (YOLO, SSD series) and two-stage target detection models (RCNN series). Compared with the two-stage target detection network model, the one-stage target detection model is slightly inferior in accuracy, but it is far ahead in detection speed. For the detection of camouflaged targets in images collected by drones on military battlefields, it is obvious that the one-stage target detection network is more suitable. The above-mentioned convolutional neural network algorithms are mostly Anchor-based algorithms. At present, Anchor-free algorithms with comparable detection performance have emerged one after another. However, the color and background of the camouflaged target are highly integrated, which is not conducive to the Anchor-free algorithm to determine the key points of the camouflaged personnel target. Therefore, the Anchor-based algorithm is still used.

[0006] The present invention aims to improve the original YOLOv4 algorithm by combining the characteristic information of the camouflaged target while ensuring the detection speed, so that the algorithm can accurately detect camouflaged personnel targets and meet the application requirements of military scenarios. Summary of the invention

[0007] Purpose of the invention: The present invention proposes a camouflaged person target detection method based on improved Scaled-YOLOv4, which overcomes the shortcomings of the prior art, improves the network structure framework, and improves the accuracy of camouflaged target image detection obtained by drone acquisition.

[0008] Technical solution: The present invention discloses a method for detecting a disguised person based on an improved Scaled-YOLOv4, comprising the following steps:

[0009] (1) Preprocessing the pre-acquired active images of disguised personnel and dividing them into training sets and test sets;

[0010] (2) Construct an extended cross-stage network Scaled-YOLOv4 model including a backbone network, a path aggregation network, and a detection head;

[0011] (3) In the Scaled-YOLOv4 model, a top-down path feature enhancement network structure including two receptive field blocks and two effective channel attention modules is designed to replace the original path aggregation network to form an improved Scaled-YOLOv4 model; the receptive field block obtains richer target feature information; the effective channel attention module improves the attention to the target feature information by calculating different weights for different features;

[0012] (4) Based on the improved Scaled-YOLOv4 model, target detection of disguised persons is performed.

[0013] Furthermore, the implementation process of step (1) is as follows:

[0014] A drone equipped with visible light and infrared cameras is used to collect images of camouflaged personnel activities in the woods, and N images are selected to produce a data set, which includes M background environment images; the camouflaged personnel targets are annotated in the N selected images, and the sizes of all images are normalized; the data set images are divided into a training set and a test set in a ratio of 6:4.

[0015] Furthermore, the top-down path feature enhancement network structure described in step (3) includes:

[0016] The first feature layer passes through a downsampling module and is fused with the second feature layer through channel cascade fusion to obtain a new output feature layer P / 4_1 after 4 times downsampling; the P / 4_1 feature layer passes through a newly designed receptive field block, an effective channel attention module and a downsampling module, and is fused with the third feature layer through channel cascade fusion to obtain a new output feature layer P / 8_1 after 8 times downsampling; the P / 8_1 feature layer passes through a newly designed receptive field block and a downsampling module, and is fused with the fourth feature layer through channel cascade fusion to obtain a new output feature layer P / 16_1 after 16 times downsampling; the P / 16_1 feature layer passes through a downsampling module, and is fused with the fifth feature layer through channel cascade fusion to obtain After 32 times downsampling, the new output feature layer P / 32_1 is obtained; after the P / 32_1 feature layer passes through an effective channel attention module and a convolution module, the first detection head P / 32_3 is obtained; after the P / 32_1 feature layer passes through an effective channel attention module, the P / 32_2 feature layer passes through an upsampling module and is channel-concatenated with the P / 16_1 feature layer to obtain the P / 16_2 feature layer, and the feature layer passes through a convolution module to obtain the second detection head P / 16_3; the P / 16_2 feature layer passes through an upsampling module and is channel-concatenated with the P / 8_1 feature layer to obtain the P / 8_2 feature layer, and the feature layer passes through a convolution module to obtain the third detection head P / 8_3.

[0017] Furthermore, the effective channel attention module in step (3) calculates different weights for different features as follows:

[0018] Adopt channel-level global average pooling without reducing the dimension; calculate each channel and its k neighbors to capture local cross-channel interaction information, use adaptive method to determine k, and calculate different weight values ​​for each channel; the calculation process of its k value is as follows:

[0019] C=φ(k)=2 (2×k-1)

[0020]

[0021] Where |x| odd It is to find the odd value closest to the variable x.

[0022] Furthermore, the receptive field block structure in step (3) is as follows:

[0023] The first branch consists of a 1×1 convolutional layer and a 3×3 convolutional layer with a dilation rate of 1; the second branch consists of a 1×1 convolutional layer, a 3×3 convolutional layer, and a 3×3 convolutional layer with a dilation rate of 3; the third branch consists of a 1×1 convolutional layer, a 1×3 convolutional layer, a 3×1 convolutional layer, and a 3×3 convolutional layer with a dilation rate of 5; the above three branches are concatenated and followed by a 1×1 convolutional layer, and then channel feature cascaded with the output of the original feature input layer after a 1×1 convolutional layer.

[0024] Furthermore, the detection head loss function is defined as follows:

[0025]

[0026] Among them, L box is the bounding box loss function, L obj is the target object loss function, L cls is the classification loss function, N is the number of detection layers, B is the number of targets whose labels are assigned to the prior box, S×S is the number of grids into which the scale is divided, and λ1, λ2, and λ3 are the weights of the three parts of the loss function;

[0027] Bounding box loss function L box Select CIoU loss function L CIoU ,L CIoU Defined as:

[0028]

[0029]

[0030]

[0031] Among them, b, b gt are the predicted box and the true label box, w, h, w gt 、h gt are the width and height of the predicted box and the real label box, respectively, ρ represents the center point distance between the predicted box and the real label box, α represents the weight coefficient, and c represents the distance between the predicted box b and the real label box b gt The diagonal distance of the minimum outer rectangle of the two;

[0032] Target object loss function L obj And the classification loss function L cls Both use the binary cross entropy loss function BCEWithLogitsLoss, which is defined as:

[0033]

[0034]

[0035] Among them, σ(x) is the Sigmoid function, which maps the variable x to between 0 and 1; x represents the probability value of the existence of the target in the prediction box or the probability value of belonging to a certain class, y represents the true label value, and n represents the number of prediction box labels in each batch.

[0036] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: According to the characteristics of the images of camouflaged personnel activities collected by drones, the present invention improves the Scaled-YOLOv4 network model, designs a new path aggregation network structure, and achieves improved detection accuracy of camouflaged target images collected by drones; the new receptive field block module proposed in the present invention greatly reduces the false detection rate of the network, and the introduced attention mechanism module greatly reduces the missed detection rate of the network, which can effectively solve the problem of low detection accuracy of the original network, and provides a new idea for realizing real-time reconnaissance target detection by drones in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the improved Scaled-YOLOv4 network structure proposed in the present invention;

[0038] Figure 2 This is a network structure diagram of feature layers 1-5 in the improved Scaled-YOLOv4 network proposed in the present invention;

[0039] Figure 3 This is a schematic diagram of the attention module structure in the improved Scaled-YOLOv4 network proposed in the present invention;

[0040] Figure 4 This is a schematic diagram of the receptive field block structure in the improved Scaled-YOLOv4 network proposed in the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below with reference to the accompanying drawings.

[0042] The present invention provides a disguised person target detection method based on improved Scaled-YOLOv4, which specifically includes the following steps:

[0043] Step 1: Preprocess the pre-acquired active images of disguised personnel and divide them into training set and test set.

[0044] Prepare network input data, use drones equipped with infrared cameras and visible light cameras to collect images of camouflaged personnel in the woods, and select 1120 1920pixel×1080pixel images to create a data set, including 63 background environment images. The camouflaged personnel in the images are in three postures: standing, lying down, and squatting. After that, select the image annotation tool LabelME to complete the annotation of the camouflaged personnel targets in the 1120 images in the data set, and divide the data set images into training sets and test sets in a ratio of 6 to 4. Finally, all images are normalized to 416pixel×416pixel and input into the set detection network.

[0045] Step 2: Build an extended cross-stage network Scaled-YOLOv4 model including the backbone network CSPDarkNet53, the path aggregation fusion network PANet and the detection head YoloHead.

[0046] The backbone network is mainly used to extract features from the input image to obtain more spatial position and semantic information of the target; the path aggregation network is mainly used to aggregate features between different feature extraction layers so that the spatial information and semantic information of the target can be highly integrated, laying a deeper foundation for target detection; the detection head can ultimately determine the final position of the target in the image.

[0047] Step 3: A top-down path feature enhancement network structure is designed in the Scaled-YOLOv4 model to replace the PANet part in the original network structure to form an improved YOLOv4 model. The top-down path feature enhancement network structure mainly includes two receptive field blocks and two effective channel attention modules; the improved Scaled-YOLOv4 model consists of three parts, the feature extraction network CSPDarkNet53, the feature fusion network, i.e. the top-down path feature enhancement network and the detection head Yolo Head; the specific structure is as follows Figure 1 shown.

[0048] The scale of the input image is calibrated to P, then P / 2 means that the width and height of the image are 1 / 2 of the original scale, and so on. Figure 1 and Table 1. The '_1', '_2', and '_3' in P / 8_1, P / 8_2, and P / 8_3 represent different aggregation enhancements of features at the feature scale, respectively; the same is true for other scales.

[0049] Table 1 shows the network parameters of the improved Scaled-YOLOv4 model

[0050] P(input image) 416×416×3 (feature: width×height×number of channels) P / 2 208×208×64 P / 4 104×104×128 P / 8 52×52×256 P / 16 26×26×512 P / 32 13×13×1024 P / 4_1 104×104×128 P / 8_1 52×52×256 P / 16_1 26×26×512 P / 32_1 13×13×1024 P / 8_2 52×52×128 P / 16_2 26×26×256 P / 32_2 13×13×512 P / 8_3 52×52×256 P / 16_3 26×26×512 P / 32_3 13×13×1024

[0051] The size of the feature map of the image obtained by the improved Scaled-YOLOv4 network in three dimensions is called size, which is set to W×H×C, where W, H, and C represent the width, height, and number of channels of the feature map, respectively. In the present invention, the scale of the original input image is set to 416×416×3, indicating that the image size of the input network is a 416×416 pixel image with 3 channels.

[0052] like Figure 1 As shown, the specific structure of the feature extraction network in the improved Scaled-YOLOv4 network is as follows:

[0053] For a 3-channel input image, after a 3×3 standard convolution layer, the standard convolution layer includes a convolution operation, a normalization operation, and a Mish activation function. The output scale remains unchanged and the number of channels increases to 32.

[0054] After a 3×3 standard convolution layer with a step size s of 2, it is actually a convolution downsampling operation. At this time, the feature map scale is P / 2, and then it passes through a residual structure block A, that is, the first branch passes through a 1×1 and a 3×3 standard convolution layer. The 1×1 standard convolution layer will half the number of channels of the input, and the 3×3 standard convolution layer will increase the number of channels to the number of input channels. The second branch does not perform any operation, and then the features of the two branches are added. Specifically, Figure 2 As shown in Figure 1, the output feature map scale is P / 2 and the number of channels increases to 64.

[0055] After a 3×3 standard convolution layer with a step size s of 2, it is actually a convolution downsampling operation. At this time, the feature map scale is P / 4, and then it passes through a residual structure block B. That is, the first branch passes through a 1×1 standard convolution layer and 2 residual structure blocks A, and the second branch does not perform any operation. Then the features of the two branches are added together, as shown in the following figure. Figure 2 As shown in the figure, the output feature map scale is P / 4 and the number of channels increases to 128.

[0056] After a 3×3 standard convolution layer with a step size s of 2, it is actually a convolution downsampling operation. At this time, the feature map scale is P / 8, and then it passes through a residual structure block B. That is, the first branch passes through a 1×1 standard convolution layer and 3 residual structure blocks A, and the second branch does not perform any operation. Then the features of the two branches are added together, as shown in the following figure. Figure 2 As shown in the figure, the output feature map scale is P / 8 and the number of channels is increased to 256.

[0057] After a 3×3 standard convolution layer with a step size s of 2, it is actually a convolution downsampling operation. At this time, the feature map scale is P / 16, and then it passes through a residual structure block B. That is, the first branch passes through a 1×1 standard convolution layer and 5 residual structure blocks A, and the second branch does not perform any operation. Then the features of the two branches are added together, as shown in the following figure. Figure 2 As shown in the figure, the output feature map scale is P / 16 and the number of channels increases to 512.

[0058] After a 3×3 standard convolution layer with a step size s of 2, it is actually a convolution downsampling operation. At this time, the feature map scale is P / 32, and then it passes through a residual structure block B. That is, the first branch passes through a 1×1 standard convolution layer and 3 residual structure blocks A, and the second branch does not perform any operation. Then the features of the two branches are added. Specifically, Figure 2 As shown in the figure, the output feature map scale is P / 32 and the number of channels increases to 1024.

[0059] The above standard convolutional layers without specifying the step size s all have a step size of 1.

[0060] like Figure 1 As shown, the specific structure of the feature fusion network in the improved Scaled-YOLOv4 network is as follows:

[0061] (1) The feature structure map of the P / 2 feature layer passes through a downsampling module, which is essentially a 3×3 standard convolution operation with a step size s of 2, and then performs a feature channel cascade with the feature structure map of the P / 4 feature layer to obtain the P / 4_1 feature layer, whose number of channels is 128. The P / 4_1 feature layer then passes through a self-designed receptive field module, an effective channel attention mechanism module, and a 3×3 standard convolution downsampling module with a step size s of 2, and then performs a channel feature cascade with the feature structure map of the P / 8 feature layer to obtain the P / 8_1 feature layer, whose number of channels is 256.

[0062] (2) The feature structure graph of the output P / 8_1 feature layer in (1) is passed through a self-designed receptive field block module and a 3×3 standard convolution downsampling module with a step size s of 2, and then a channel feature cascade is performed with the feature structure graph of the P / 16 feature layer to obtain a P / 16_1 feature layer with 512 channels.

[0063] (3) The feature structure graph of the output P / 16_1 feature layer in (2) is passed through a 3×3 standard convolution downsampling module with a step size s of 2, and then a channel feature cascade is performed with the feature structure graph of the P / 32 feature layer to obtain a P / 32_1 feature layer with 1024 channels.

[0064] (4) The feature structure diagram of the output P / 32_1 feature layer in (3) is passed through a channel-effective attention module to obtain the P / 32_2 feature layer, where the number of channels is 512. The P / 32_2 feature layer is then passed through a 3×3 standard convolution module with s=1 to obtain the P / 32_3 feature layer, which becomes one of the detection heads.

[0065] (5) The feature structure diagram of the P / 32_2 feature layer in (4) is passed through a linear interpolation upsampling module, and the channel features are cascaded with the P / 16_1 feature layer to obtain the P / 16_2 feature layer. At this time, the number of channels is 256. The P / 16_2 feature layer is then passed through a 3×3 standard convolution module with s=1 to obtain the P / 16_3 feature layer, which becomes one of the detection heads.

[0066] (6) The feature structure diagram of the P / 16_2 feature layer in (5) is passed through a linear interpolation upsampling module, and the channel features are cascaded with the P / 8_1 feature layer to obtain the P / 8_2 feature layer. At this time, the number of channels is 128. The P / 8_2 feature layer is then passed through a 3×3 standard convolution module with s=1 to obtain the P / 8_3 feature layer, which becomes one of the detection heads.

[0067] The effective channel attention module is as follows Figure 3 As shown in the figure, specifically, in the structure of the effective channel attention mechanism, for the feature map of input size W×H×C, the global average pooling layer processes the output tensor of size 1×1×C, and then successively passes through the adaptive k value selection and Sigmoid activation layer to output a 1×1×C tensor, which is equivalent to recalculating the weights of each channel. Finally, the 1×1×C tensor is multiplied by the corresponding channel of the feature map of the original input size W×H×C to obtain a new output feature map of size W×H×C. The calculation process of the adaptive k value is:

[0068] C=φ(k)=2 (2×k-1)

[0069]

[0070] Where |x| odd is the odd number closest to the variable x. By calculating the different weights between different channels of the feature, the feature can better focus on more effective information of the disguised person target when it is transmitted, which is conducive to improving the accuracy of disguised person target detection.

[0071] like Figure 4 As shown in Figure 1, the receptive field block structure is as follows:

[0072] The first branch consists of a 1×1 standard convolution layer (the standard convolution layer includes a convolution operation, a normalization operation and a ReLU activation function) and a 3×3 standard convolution layer with a dilation rate of 1; the second branch consists of a 1×1 standard convolution layer, a 3×3 standard convolution layer and a 3×3 standard convolution layer with a dilation rate of 3; the third branch consists of a 1×1 standard convolution layer, a 1×3 standard convolution layer, a 3×1 standard convolution layer and a 3×3 standard convolution layer with a dilation rate of 5; the nonlinear convolution layers with kernels of 1×3 and 3×1 can enhance the irregular feature extraction of the camouflaged target, and then obtain the feature enhancement map of the camouflaged target; the above three branches are concatenated and followed by a 1×1 standard convolution layer, and then the features are added to the output of the original feature input layer after a 1×1 standard convolution layer.

[0073] The invention is implemented in Python 3.6 based on the Pytorch1.7.1 toolbox, accelerated by a NAVIDIA RTX 2080 GPU device with 11GB GPU memory. All input images are resized to 416×416 pixels. The model is trained with a stochastic gradient descent (SGD) optimizer, with the momentum size set to 0.937 and the initial learning rate set to 1e-2. The training progress takes about 100 minutes, with a total of 200 batches trained and the batch size set to 4.

[0074] The detection head divides the feature map into S×S grids. The grid in which the center point of the target to be detected is located is responsible for detecting the target. Each grid predicts B prediction boxes, and each prediction box predicts 6 values. These 6 values ​​are used to describe the position and size of the prediction box and determine whether there is a target in the box and what category the target belongs to. They are: the distance t between the center coordinate of the prediction box and the coordinate of the upper left corner of the grid x and t y , and the ratio of the width and height of the real target t w ,t h And the confidence of judging whether there is a target in the prediction box, and the probability value of belonging to a certain category. The conversion formula between the prediction value and the output target position is as follows:

[0075] b x =σ(t x )+c x

[0076] b y =σ(t y )+c y

[0077]

[0078]

[0079] Among them, pw and p h Indicates the width and height of the prediction box, c x and c y Represents the coordinates of the upper left corner of the grid where the prediction box is located, σ represents the Sigmoid activation function, and e represents the natural logarithm.

[0080] The overall loss function in the detection head is defined as follows:

[0081]

[0082] Among them, L box is the bounding box loss function, L obj is the target object loss function, L cls is the classification loss function, N is the number of detection layers, B is the number of targets whose labels are assigned to the prior boxes, S×S is the number of grids into which the scale is divided, and λ1, λ2, and λ3 are the weights of the three parts of the loss function.

[0083] Bounding box loss function L box Select CIoU loss function L CIoU , which is defined as:

[0084]

[0085]

[0086]

[0087] Among them, b, b gt are the predicted box and the true label box, w, h, w gt 、h gt are the width and height of the predicted box and the real label box, respectively, ρ represents the center point distance between the predicted box and the real label box, α represents the weight coefficient, and c represents the distance between the predicted box b and the real label box b gt The diagonal distance of the minimum outer rectangle of the two.

[0088] Target object loss function L obj And the classification loss function L cls Both use the binary cross entropy loss function BCEWithLogitsLoss, which is defined as:

[0089]

[0090]

[0091] Among them, σ(x) is the Sigmoid function, which maps the variable x to between 0 and 1. x represents the probability value of the existence of the target in the prediction box or the probability value of belonging to a certain class, y represents the true label value, and n represents the number of prediction box labels in each batch.

[0092] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method 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 in the scope of protection of the present invention.

Claims

1. A disguised person target detection method based on improved Scaled-YOLOv4, characterized in that: The following steps are involved: (1) Preprocessing the pre-acquired active images of disguised personnel and dividing them into training sets and test sets; (2) Construct an extended cross-stage network Scaled-YOLOv4 model including a backbone network, a path aggregation network, and a detection head; (3) In the Scaled-YOLOv4 model, a top-down path feature enhancement network structure including two receptive field blocks and two effective channel attention modules is designed to replace the original path aggregation network to form an improved Scaled-YOLOv4 model; the receptive field block obtains richer target feature information; The effective channel attention module improves the attention to the target feature information by calculating different weights for different features; (4) Detect disguised persons based on the improved Scaled-YOLOv4 model; The top-down path feature enhancement network structure described in step (3) includes: The first feature layer passes through a downsampling module and is fused with the second feature layer through channel cascade fusion to obtain a new output feature layer P / 4_1 after 4 times downsampling; the P / 4_1 feature layer passes through a newly designed receptive field block, an effective channel attention module and a downsampling module, and is fused with the third feature layer through channel cascade fusion to obtain a new output feature layer P / 8_1 after 8 times downsampling; the P / 8_1 feature layer passes through a newly designed receptive field block and a downsampling module, and is fused with the fourth feature layer through channel cascade fusion to obtain a new output feature layer P / 16_1 after 16 times downsampling; the P / 16_1 feature layer passes through a downsampling module, and is fused with the fifth feature layer through channel cascade fusion to obtain After 32 times downsampling, the new output feature layer P / 32_1 is obtained; the P / 32_1 feature layer passes through an effective channel attention module and a convolution module to obtain the first detection head P / 32_3; the P / 32_2 feature layer after the P / 32_1 feature layer passes through an effective channel attention module is channel-concatenated with the P / 16_1 feature layer through an upsampling module to obtain the P / 16_2 feature layer, and the feature layer passes through a convolution module to obtain the second detection head P / 16_3; the P / 16_2 feature layer passes through an upsampling module and is channel-concatenated with the P / 8_1 feature layer to obtain the P / 8_2 feature layer, and the feature layer passes through a convolution module to obtain the third detection head P / 8_3; The effective channel attention module in step (3) calculates different weights for different features as follows: Adopt channel-level global average pooling without reducing the dimension; calculate each channel and its k neighbors to capture local cross-channel interaction information, use adaptive method to determine k, and calculate different weight values ​​for each channel; the calculation process of its k value is as follows: C=φ(k)=2 (2×k-1) Where |x| odd It is to find the odd value closest to the variable x; The receptive field block structure in step (3) is as follows: The first branch consists of a 1×1 convolutional layer and a 3×3 convolutional layer with a dilation rate of 1; the second branch consists of a 1×1 convolutional layer, a 3×3 convolutional layer, and a 3×3 convolutional layer with a dilation rate of 3; the third branch consists of a 1×1 convolutional layer, a 1×3 convolutional layer, a 3×1 convolutional layer, and a 3×3 convolutional layer with a dilation rate of 5; the above three branches are concatenated and followed by a 1×1 convolutional layer, and then channel feature cascaded with the output of the original feature input layer after a 1×1 convolutional layer.

2. According to claim 1, a disguised person target detection method based on improved Scaled-YOLOv4 is characterized in that: The implementation process of step (1) is as follows: A drone equipped with visible light and infrared cameras is used to collect images of camouflaged personnel activities in the woods, and N images are selected to produce a data set, which includes M background environment images; the camouflaged personnel targets are annotated in the N selected images, and the sizes of all images are normalized; the data set images are divided into a training set and a test set in a ratio of 6:

4.

3. According to claim 1, a disguised person target detection method based on improved Scaled-YOLOv4 is characterized in that: The detection head loss function is defined as follows: Among them, L box is the bounding box loss function, L obj is the target object loss function, L cls is the classification loss function, N is the number of detection layers, B is the number of targets whose labels are assigned to the prior boxes, and λ1, λ2, and λ3 are the weights of the three parts of the loss function; Bounding box loss function L box Choose CIoU loss function L CIoU ,L CIoU Defined as: Among them, b, b gt are the predicted box and the true label box, w, h, w gt 、h gt are the width and height of the predicted box and the real label box, respectively, ρ represents the center point distance between the predicted box and the real label box, α represents the weight coefficient, and c represents the distance between the predicted box b and the real label box b gt The diagonal distance of the minimum outer rectangle of the two; Target object loss function L obj And the classification loss function L cls Both use the binary cross entropy loss function BCEWithLogitsLoss, which is defined as: Among them, σ(x) is the Sigmoid function, which maps the variable x to between 0 and 1; x represents the probability value of the existence of the target in the prediction box or the probability value of belonging to a certain class, y represents the true label value, and n represents the number of prediction box labels in each batch.

Images