Fog-removed target detection method and system

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) inspection, and particularly to a target detection method and system for defogging: Feature extraction is performed on a foggy image to obtain original feature maps at N scales; a fog line transmittance map is obtained from the first-scale original feature map that matches the target to be detected; the first-scale original feature map belongs to one of the N scales of original feature maps; using the fog line transmittance map, the first-scale original feature map is corrected to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected; the second-scale feature map and all N-1 scales of original feature maps excluding the first-scale original feature map are fused to output N fused feature maps of different scales; target detection is performed using the N fused feature maps to obtain the target to be detected. This method accurately identifies red / yellow safety helmets / reflective clothing in foggy conditions while reducing computational overhead.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) inspection, and in particular to a defogging target detection method and a defogging target detection system. Background Technology

[0002] Drone aerial photography is widely used in construction safety inspections, but aerial images are often affected by weather conditions such as haze and fog, resulting in decreased image contrast and color distortion. In particular, high-frequency color targets such as safety helmets (usually red / yellow) and reflective vests (usually fluorescent yellow / orange) have a sharp drop in saturation in foggy conditions, making them easily confused with the background (such as loess and fallen leaves), resulting in missed detection. Existing dehazing techniques are mainly divided into two categories: (1) Image enhancement-based preprocessing methods: such as Dark Channel Prior (DCP), Histogram Equalization (HE), Retinex, etc. These methods perform pixel-level reconstruction in the image spatial domain, restore a clear image, and then input it into the detection network. For example, an aerial image dehazing enhancement method based on HTCS. Defects: Pixel-level reconstruction requires a large amount of computation (DCP processing of a single 1080P image takes about 200ms-500ms), which cannot meet the real-time requirements of UAV on-board edge computing (usually requiring <30ms); and the dehazing process is decoupled from the detection task, so the enhanced target may not be the target of the detection network. (2) Based on feature domain enhancement methods, in recent years, some studies have attempted to introduce attention mechanisms or multi-scale fusion into feature extraction networks to enhance fog features, such as end-to-end dehazing networks like DehazeNet and AOD-Net. Defects: These methods usually require a large amount of paired fog-free training data, and the network structure is complex with a large number of parameters; more importantly, they lack specificity for enhancing specific color targets (red / yellow). In the high-altitude view of UAVs, safety helmets / reflective vests only occupy tens of pixels, and general dehazing networks cannot effectively preserve their color features. The shortcomings of existing technologies are as follows: pixel-level dehazing has good effect but is slow, feature-level enhancement is fast but has poor targeting; the dehazing module is trained independently of the detection network, and the enhanced image features may not be optimally adapted to the detection task; the red / yellow of safety helmets / reflective vests has a unique attenuation pattern in foggy weather, and existing methods have not specifically modeled this. Summary of the Invention

[0003] This invention addresses the technical problem of detecting red / yellow safety helmets / reflective vests in foggy weather by providing a defogging target detection method and a defogging target detection system.

[0004] Firstly, a dehazing target detection method is provided, including:

[0005] Feature extraction is performed on the foggy image to obtain original feature maps at N scales; Obtain the fog line transmittance map from the first-scale original feature map that matches the target to be detected; the first-scale original feature map belongs to one of the N scale original feature maps; Using the fog line transmittance map, the original feature map at the first scale is corrected to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected. The second-scale feature map and all N-1 original feature maps (excluding the first-scale original feature map) are fused to output N fused feature maps of different scales. The target to be detected is obtained by using the N fused feature maps.

[0006] Preferably, the step of obtaining the fog line transmittance map in the first-scale original feature map that matches the target to be detected includes: The original feature map at the first scale is passed through a deep convolutional layer to obtain a fog line spatial feature map that is completely independent between channels with the same scale and number of channels. The spatial feature map of the fog line is input into the channel attention mechanism unit to obtain the sensitivity maps of different channels to the fog line. The sensitivity maps of different channels to fog lines are input into a pointwise convolutional layer, and a first activation function is introduced to obtain a single-channel fog line transmittance map representing the fog line intensity.

[0007] Preferably, after inputting the fog line spatial feature map into the channel attention mechanism unit, a second activation function is introduced to filter out the feature channels most sensitive to fog lines and suppress irrelevant noise channels.

[0008] Preferably, the step of using the fog transmittance map to correct the original feature map at the first scale to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected includes: The second-scale feature map can be obtained using the following formula: F' = F ⊙ (1 + α · M_color ⊙ (1 - t_map)) Where F: original feature map at the first scale; ⊙: element-wise multiplication; α: learnable enhancement coefficient, initialized to 1.0, jointly optimized with the network; M_color: color prior mask, with a dimension of 1×1×C, representing the sensitivity of each feature channel to red / yellow; t_map: fog transmittance map; F': feature map at the second scale.

[0009] Preferably, obtaining the color prior mask includes: The image of the specified color patch is input into the pre-trained backbone layer. The average response value of each channel of the backbone layer output is recorded. The channel with the higher response is assigned a higher first weight in the color prior mask, and the remaining channels are assigned a lower second weight.

[0010] Secondly, a defogging target detection system is also provided, including: Backbone layer: Used to extract features from foggy images to obtain original feature maps at N scales; Fog line estimation unit: used to acquire the fog line transmittance map in the first-scale original feature map that matches the target to be detected; the first-scale original feature map belongs to one of the N scale original feature maps; Feature correction unit: used to correct the original feature map of the first scale using the fog line transmittance map, so as to obtain a second scale feature map that suppresses fog interference and enhances the target to be detected; Feature fusion unit: used to fuse the second-scale feature map with all N-1 original feature maps at all scales excluding the first-scale original feature map, and output a fused feature map of N different scales after deep and shallow layer fusion; Detection unit: used to perform target detection using the N fused feature maps to obtain the target to be detected.

[0011] Preferably, the step of obtaining the fog line transmittance map in the first-scale original feature map that matches the target to be detected includes: Deep convolutional unit: used to pass the original feature map of the first scale through a deep convolutional layer to obtain a fog line spatial feature map that is completely independent between channels with the same scale and number of channels; Channel attention unit: used to input the fog line spatial feature map into the channel attention mechanism unit to obtain the sensitivity maps of different channels to fog lines; Pointwise convolutional unit: used to input the sensitivity maps of different channels to fog lines into the pointwise convolutional layer and introduce a first activation function to obtain a single-channel fog line transmittance map representing the fog line intensity.

[0012] Preferably, after inputting the fog line spatial feature map into the channel attention mechanism unit, a second activation function is introduced to filter out the feature channels most sensitive to fog lines and suppress irrelevant noise channels.

[0013] Preferably, the feature correction unit includes: obtaining a second-scale feature map using the following formula: F' = F ⊙ (1 + α · M_color ⊙ (1 - t_map)) Where F: original feature map at the first scale; ⊙: element-wise multiplication; α: learnable enhancement coefficient, initialized to 1.0, jointly optimized with the network; M_color: color prior mask, with a dimension of 1×1×C, representing the sensitivity of each feature channel to red / yellow; t_map: fog transmittance map; F': feature map at the second scale.

[0014] Preferably, obtaining the color prior mask includes: The image of the specified color patch is input into the pre-trained backbone layer. The average response value of each channel of the backbone layer output is recorded. The channel with the higher response is assigned a higher first weight in the color prior mask, and the remaining channels are assigned a lower second weight.

[0015] Beneficial effects: 1. Unlike the classic dark channel prior that estimates transmittance in RGB space, this application directly estimates fog lines in feature space, avoiding the high computational cost of pixel-level reconstruction; while reducing computational cost, it accurately identifies red / yellow safety helmets / reflective clothing in foggy conditions.

[0016] 2. Introduce a channel attention mechanism to enable the network to automatically learn which feature channels are sensitive to fog, thereby improving estimation accuracy; 3. The total number of side parameters is less than 10,000, and the computational load is less than 0.1G FLOPs, which fully meets the real-time requirements of the airborne terminal. Attached Figure Description

[0017] Figure 1 This is an embodiment diagram of a defogging target detection method; Figure 2 A first embodiment diagram for obtaining the fog line transmittance map in the first-scale original feature map that matches the target to be detected; Figure 3 A second embodiment diagram for obtaining the fog line transmittance map in the first-scale original feature map that matches the target to be detected; Figure 4 This is an embodiment diagram of a defogging target detection system; Figure 5 An example diagram illustrating the process of obtaining a fog line transmittance map from the first-scale original feature map that matches the target to be detected.

[0018] Explanation of reference numerals in the attached figures: 1. A defogging target detection system; 11. Backbone layer; 12. Fog line estimation unit; 13. Feature correction unit; 14. Feature fusion unit; 15. Detection unit; 121. Depth convolution unit; 122. Channel attention unit; 123. Pointwise convolution unit. Detailed Implementation

[0019] Terminology Explanation: YOLO: You Only Look Once; a single-stage, end-to-end, real-time object detection algorithm; applicable to object detection; consisting of: backbone layer, neck layer, and head layer.

[0020] Firstly, such as Figure 1 As shown, a target detection method for defogging is provided, including: S11: Feature extraction is performed on the foggy image to obtain N original feature maps at different scales. In this technical solution, the foggy image is an optical or digital image acquired through an imaging device. The foggy image is processed by the backbone layer to obtain N original feature maps of different scales. A foggy image only contains RGB pixels, making it impossible to distinguish between fog and fog lines, or between different levels of fog density, and therefore cannot be directly used for fog line estimation and removal. The backbone layer outputs a set of original feature maps (P3, P4, P5) from smallest to largest, with completely different information in each layer. Shallow original feature map P3: high resolution, rich details, suitable for: small targets, fog lines, contours, and fine structure estimation. Mid-level original feature map P4: balances details and semantics, suitable for: medium targets, local shapes, and regional structures. Deep original feature map P5: very low resolution, loss of details, suitable for: large targets and overall scene judgment. The original feature maps are processed semi-finished products, multi-scale, multi-channel, and layered storage of detailed / semantic features. In this technical solution, the shallow original feature map is obtained by downsampling the foggy image by 8 times; the middle original feature map is obtained by downsampling the foggy image by 16 times; and the deep original feature map is obtained by downsampling the foggy image by 32 times.

[0021] S12: Obtain the fog line transmittance map from the first-scale original feature map that matches the target to be detected; the first-scale original feature map belongs to one of the N scale original feature maps; any one or more of the shallow, medium, and deep original feature maps can be selected for subsequent processing. In this technical solution, the purpose is to detect small targets such as red / yellow safety helmets and reflective clothing, so the shallow original feature map is selected for processing. After processing by the fog line estimation unit, the fog line transmittance map is obtained. The fog line transmittance map refers to the map showing the dense fog area and the fog-free area.

[0022] S13: Using the fog transmittance map, the first-scale original feature map is corrected to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected. In this technical solution, the fog transmittance map is used to correct the shallow original feature map P3, and then a second-scale feature map that suppresses fog interference and enhances the red / yellow safety helmet and reflective clothing is obtained. The second-scale feature map has the same scale as the first-scale original feature map, but the second-scale feature map enhances the image of the target to be detected, that is, enhances the red / yellow safety helmet and reflective clothing.

[0023] S14: Perform feature fusion on the second-scale feature map and all N-1 original feature maps (excluding the first-scale original feature map), outputting N fused feature maps of different scales. In this technical solution, the Neck layer is used to perform feature fusion on the second-scale feature map and all N-1 original feature maps (excluding the first-scale original feature map). After the Neck layer performs bidirectional feature fusion on the input N scale feature maps (N=3 in this application, i.e., P3, P4, P5), it outputs 3 new feature maps with the same number and scale, but significantly enhanced feature quality. These 3 new feature maps are the 3 fused feature maps of different scales. S15: Target detection is performed using the N fused feature maps to obtain the target to be detected. In this technical solution, target detection is performed using the head layer. In this technical solution, the target to be detected is: small target: safety helmet / reflective clothing; medium target: worker / equipment; large target: vehicle / scene. Thus, this technical solution achieves the technical effect of accurately detecting red / yellow safety helmets and reflective clothing in foggy conditions.

[0024] Preferred, such as Figure 2 As shown, obtaining the fog line transmittance map in the first-scale original feature map that matches the target to be detected includes: S121: The original feature map at the first scale is passed through a deep convolutional layer to obtain a completely independent fog line spatial feature map with the same scale and number of channels. In this technical solution, the shallow original feature map P3 is passed through a deep convolutional layer. In the deep convolutional layer, each convolutional kernel is only responsible for the convolution operation of one channel. Assuming there is an input feature map with n channels, there is an independent convolutional kernel for each channel. These convolutional kernels slide in the spatial dimension, performing convolution operations on their respective channels to generate an intermediate feature map with the same number of input channels. The main function of the deep convolutional layer is to extract spatial features. Since each convolutional kernel only operates on one channel, the computational cost is greatly reduced.

[0025] S122: Input the fog line spatial feature map into the channel attention mechanism unit to obtain the sensitivity maps of different channels to fog lines; the core function of the channel attention mechanism unit is to enable the model to adaptively focus on useful feature channels and suppress irrelevant channels, thereby improving the feature expression ability and adding almost no additional computational burden.

[0026] S124: The sensitivity maps of different channels to fog lines are input into the pointwise convolutional layer, and a first activation function is introduced to obtain a single-channel fog line transmittance map representing the fog line intensity. The pointwise convolutional layer is immediately followed by the depthwise convolutional layer, and the kernel size of the pointwise convolutional layer is 1×1. The function of pointwise convolution is to linearly combine the intermediate feature maps obtained by depthwise convolution in the channel dimension to generate the final output feature map. Through multiplication and accumulation operations of the 1×1 convolutional kernel in the channel dimension, pointwise convolution can adjust the number of channels of the feature map to the desired number of output channels M. This method achieves the fusion and transformation of features from different channels without increasing the computational cost in the spatial dimension. In this technical solution, the first activation function is ReLU; the purpose of introducing the first activation function here is to introduce nonlinearity. Because the pointwise convolutional layer is a linear fusion, it cannot learn complex fog shapes; adding the first activation function to introduce nonlinearity is to enable the defogging target detection system in this application to truly learn to recognize and fit irregular and non-uniform fog lines and fog concentration distributions in the real world.

[0027] Preferred, such as Figure 3 As shown, after inputting the fog line spatial feature map into the channel attention mechanism unit, S123: a second activation function is introduced to filter out the feature channels most sensitive to fog lines and suppress irrelevant noise channels. In this technical solution, the second activation function is Sigmoid, which is used to constrain the output range.

[0028] Preferably, the step of using the fog transmittance map to correct the original feature map at the first scale to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected includes: The second-scale feature map can be obtained using the following formula: F' = F ⊙ (1 + α · M_color ⊙ (1 - t_map)) Where F: the original feature map at the first scale; ⊙: element-wise multiplication; α: a learnable enhancement coefficient, initialized to 1.0 and jointly optimized with the network; M_color: a color prior mask with dimensions 1×1×C, representing the sensitivity of each feature channel to red / yellow; t_map: fog line transmittance map; F': the feature map at the second scale. According to the atmospheric scattering model, light of different wavelengths attenuates at different rates in fog. Red / yellow light (long wavelength) has stronger penetrating power than blue-violet light (short wavelength), but its response attenuation in image sensors is not linear. Moreover, at high altitudes, due to the influence of atmospheric path length radiation, the saturation of red / yellow targets drops sharply, and the color difference with the background (such as loess and green plants) is significantly reduced. This unique attenuation law of 'wavelength correlation - sharp drop in saturation' is the physical root cause of the failure to detect safety helmets / reflective vests.

[0029] First, a mathematical model was constructed to investigate the attenuation of red / yellow colors in foggy conditions. Based on atmospheric scattering models and color constancy theory, the attenuation of red / yellow targets is mainly reflected in two aspects: contrast attenuation: the color difference between the target and the background decreases exponentially with fog concentration (transmittance t); saturation attenuation: the response values ​​of the red / yellow channels decrease nonlinearly with increasing fog concentration.

[0030] Preferably, obtaining the color prior mask includes: The specified color patch image is input into the pre-trained backbone layer. The average response value of each channel output by the backbone layer is recorded. Channels with high responses are assigned higher first weights in the color prior mask, while the remaining channels are assigned lower second weights. The color prior mask M_color is generated as follows: Before training, the activated feature channel indices corresponding to red / yellow targets in ImageNet or a custom dataset are statistically analyzed. Specifically, red / yellow patch images are input into the pre-trained backbone, and the average response value of each channel is recorded. Channels with high responses are assigned higher first weights (0.8-1.0) in M_color, while other channels are assigned lower second weights (0.1-0.3). M_color is fixed during training and does not participate in gradient updates, ensuring that the correction always focuses on the red / yellow related channels. Correction logic: When t_map is close to 1 (dense fog region), (1 - t_map) is close to 0, resulting in a small correction amplitude to avoid over-enhancing noise; when t_map is close to 0 (no fog region), (1 - t_map) is close to 1, the high-weight channels in M_color are significantly enhanced, while the low-weight channels remain essentially unchanged. The above technical solution enables selective enhancement only in foggy weather and for the red / yellow channels, avoiding false enhancement in fog-free weather.

[0031] The entire system is trained end-to-end, and the loss function is: L_total = L_det + λ · L_t in: L_det: YOLOv11 original detection loss (classification loss + regression loss + confidence loss) L_t: Auxiliary loss from the fog line estimation side branch, measured by mean squared error (MSE) to determine the difference between the estimated transmittance map t_map and the true transmittance map. λ: Balance coefficient, set to 0.1 Acquisition of the true transmittance map: During the training phase, for foggy images, the classic dark channel prior algorithm is used to quickly generate transmittance maps as pseudo-labels (only used for supervised side branch training, not needed during inference). Since the fog line estimation unit itself is extremely lightweight, this method avoids the expensive cost of data acquisition.

[0032] During inference, the fog line estimation side branch directly outputs the transmittance map without any external input. It is fully integrated with the detection network to achieve single-stage, end-to-end fog detection.

[0033] Secondly, such as Figure 4 As shown, a defogging target detection system 1 is also provided, comprising: Backbone layer 11: Used to extract features from foggy images to obtain original feature maps at N scales; Fog line estimation unit 12: used to acquire the fog line transmittance map in the first-scale original feature map that matches the target to be detected; the first-scale original feature map belongs to one of the N scale original feature maps; Feature correction unit 13: Used to correct the original feature map of the first scale using the fog line transmittance map to obtain a second scale feature map that suppresses fog interference and enhances the target to be detected; Feature fusion unit 14: used to fuse the second scale feature map with all N-1 scale original feature maps excluding the first scale original feature map, and output the fused feature map after deep and shallow layer fusion at N different scales; Detection unit 15: Used to perform target detection using the N fused feature maps to obtain the target to be detected.

[0034] Preferred, such as Figure 5 As shown, obtaining the fog line transmittance map in the first-scale original feature map that matches the target to be detected includes: Deep convolutional unit 121: used to pass the original feature map of the first scale through a deep convolutional layer to obtain a fog line spatial feature map that is completely independent between channels with the same scale and number of channels. Channel attention unit 122: used to input the fog line spatial feature map into the channel attention mechanism unit to obtain the sensitivity maps of different channels to fog lines; Pointwise convolutional unit 123: used to input the sensitivity maps of different channels to fog lines into the pointwise convolutional layer and introduce a first activation function to obtain a single-channel fog line transmittance map representing the fog line intensity.

[0035] Preferably, after inputting the fog line spatial feature map into the channel attention mechanism unit, a second activation function is introduced to filter out the feature channels most sensitive to fog lines and suppress irrelevant noise channels.

[0036] Preferably, the feature correction unit includes: obtaining a second-scale feature map using the following formula: F' = F ⊙ (1 + α · M_color ⊙ (1 - t_map)) Where F: original feature map at the first scale; ⊙: element-wise multiplication; α: learnable enhancement coefficient, initialized to 1.0, jointly optimized with the network; M_color: color prior mask, with a dimension of 1×1×C, representing the sensitivity of each feature channel to red / yellow; t_map: fog transmittance map; F': feature map at the second scale.

[0037] Preferably, obtaining the color prior mask includes: The image of the specified color patch is input into the pre-trained backbone layer. The average response value of each channel of the backbone layer output is recorded. The channel with the higher response is assigned a higher first weight in the color prior mask, and the remaining channels are assigned a lower second weight.

[0038] Beneficial effects: 1. Unlike the classic dark channel prior that estimates transmittance in RGB space, this application directly estimates fog lines in feature space, avoiding the high computational cost of pixel-level reconstruction; while reducing computational cost, it accurately identifies red / yellow safety helmets / reflective clothing in foggy conditions.

[0039] 2. Introduce a channel attention mechanism to enable the network to automatically learn which feature channels are sensitive to fog, thereby improving estimation accuracy; 3. The total number of side parameters is less than 10,000, and the computational load is less than 0.1G FLOPs, which fully meets the real-time requirements of the airborne terminal.

[0040] Finally, it should be noted that any modification or equivalent substitution of some or all of the technical features based on the device structure and the technical solutions of the embodiments of the present invention, without departing from the corresponding technical solutions of the present invention, shall fall within the patent scope of the device structure and the embodiments of the present invention.

Claims

1. A target detection method for defogging, characterized in that, include: Feature extraction is performed on the foggy image to obtain original feature maps at N scales; Obtain the fog line transmittance map from the first-scale original feature map that matches the target to be detected; The first-scale original feature map belongs to one of the N scales of original feature maps; Using the fog line transmittance map, the original feature map at the first scale is corrected to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected. The second-scale feature map and all N-1 original feature maps (excluding the first-scale original feature map) are fused to output N fused feature maps of different scales. The target to be detected is obtained by using the N fused feature maps.

2. The target detection method for defogging according to claim 1, characterized in that, The process of obtaining the fog line transmittance map from the first-scale original feature map that matches the target to be detected includes: The original feature map at the first scale is passed through a deep convolutional layer to obtain a fog line spatial feature map that is completely independent between channels with the same scale and number of channels. The spatial feature map of the fog line is input into the channel attention mechanism unit to obtain the sensitivity maps of different channels to the fog line. The sensitivity maps of different channels to fog lines are input into a pointwise convolutional layer, and a first activation function is introduced to obtain a single-channel fog line transmittance map representing the fog line intensity.

3. The target detection method for defogging according to claim 2, characterized in that, After inputting the fog line spatial feature map into the channel attention mechanism unit, a second activation function is introduced to filter out the feature channels most sensitive to fog lines and suppress irrelevant noise channels.

4. The target detection method for defogging according to claim 3, characterized in that, The step of using the fog transmittance map to correct the original feature map at the first scale to obtain a second-scale feature map that suppresses fog interference and enhances the target to be detected includes: The second-scale feature map can be obtained using the following formula: F' = F ⊙ (1 + α · M_color ⊙ (1 - t_map)) Where F: original feature map at the first scale; ⊙: element-wise multiplication; α: learnable enhancement coefficient, initialized to 1.0, jointly optimized with the network; M_color: color prior mask, with a dimension of 1×1×C, representing the sensitivity of each feature channel to red / yellow; t_map: fog transmittance map; F': feature map at the second scale.

5. The target detection method for defogging according to claim 4, characterized in that, Obtaining the color prior mask includes: The image of the specified color patch is input into the pre-trained backbone layer. The average response value of each channel of the backbone layer output is recorded. The channel with the higher response is assigned a higher first weight in the color prior mask, and the remaining channels are assigned a lower second weight.

6. A defogging target detection system, characterized in that, include: Backbone layer: Used to extract features from foggy images to obtain original feature maps at N scales; Fog line estimation unit: used to acquire the fog line transmittance map in the first-scale original feature map that matches the target to be detected; the first-scale original feature map belongs to one of the N scale original feature maps; Feature correction unit: used to correct the original feature map of the first scale using the fog line transmittance map, so as to obtain a second scale feature map that suppresses fog interference and enhances the target to be detected; Feature fusion unit: used to fuse the second-scale feature map with all N-1 original feature maps at all scales excluding the first-scale original feature map, and output a fused feature map of N different scales after deep and shallow layer fusion; Detection unit: used to perform target detection using the N fused feature maps to obtain the target to be detected.

7. The target detection system for defogging according to claim 6, characterized in that, The process of obtaining the fog line transmittance map from the first-scale original feature map that matches the target to be detected includes: Deep convolutional unit: used to pass the original feature map of the first scale through a deep convolutional layer to obtain a fog line spatial feature map that is completely independent between channels with the same scale and number of channels; Channel attention unit: used to input the fog line spatial feature map into the channel attention mechanism unit to obtain the sensitivity maps of different channels to fog lines; Pointwise convolutional unit: used to input the sensitivity maps of different channels to fog lines into the pointwise convolutional layer and introduce a first activation function to obtain a single-channel fog line transmittance map representing the fog line intensity.

8. The target detection system for defogging according to claim 7, characterized in that, After inputting the fog line spatial feature map into the channel attention mechanism unit, a second activation function is introduced to filter out the feature channels most sensitive to fog lines and suppress irrelevant noise channels.

9. The target detection system for defogging according to claim 8, characterized in that, The feature correction unit includes: obtaining a second-scale feature map using the following formula: F' = F ⊙ (1 + α · M_color ⊙ (1 - t_map)) Where F: original feature map at the first scale; ⊙: element-wise multiplication; α: learnable enhancement coefficient, initialized to 1.0, jointly optimized with the network; M_color: color prior mask, with a dimension of 1×1×C, representing the sensitivity of each feature channel to red / yellow; t_map: fog transmittance map; F': feature map at the second scale.

10. The defogging target detection system according to claim 9, characterized in that, Obtaining the color prior mask includes: The image of the specified color patch is input into the pre-trained backbone layer. The average response value of each channel of the backbone layer output is recorded. The channel with the higher response is assigned a higher first weight in the color prior mask, and the remaining channels are assigned a lower second weight.

Images