A channel recalibration fog detection method based on detail enhancement

Abstract

The present application relates to a kind of channel recalibration fog detection methods based on detail enhancement, comprising: the image dataset of fog scene is constructed;The image dataset is input into improved YOLOv11 neural network model, model training is carried out, and fog detection model is acquired;Wherein, improved YOLOv11 neural network model includes: in main network, with detail enhancement convolution module instead of main network Backbone first layer standard convolution, in neck network, introduce iterative attention feature fusion mechanism, before embedding extrusion excitation module in each detection head Detect of head network;Unmanned vehicle is based on the fog detection model, detection is carried out in foggy day.The present application aims at making up the defects of present stage dehazing algorithm, further improve the precision and speed of unmanned vehicle in fog detection, promote the development of unmanned technology.

Description

Technical Field

[0001] This invention relates to the field of target detection technology for unmanned vehicles, and in particular to a fog detection method based on channel recalibration using enhanced detail. Background Technology

[0002] With the rapid development of artificial intelligence and sensor technology, autonomous driving technology is gradually moving from the laboratory to large-scale commercial application. According to a Counterpoint Research report, the global market penetration rate of Advanced Driver Assistance Systems (ADAS) and autonomous vehicles has exceeded 65%, and it is projected to jump to 94% by 2035, marking the imminent arrival of a new era of widespread adoption for autonomous driving technology. However, complex weather conditions such as smog severely restrict the development and widespread adoption of this technology.

[0003] From an optical principle perspective, fog, as a typical aerosol medium, contains a large number of suspended tiny water droplets and ice crystals. This causes a strong light scattering effect on the incident light, resulting in a decrease in the contrast and color distortion of the images captured by the vehicle camera. This severely restricts the accuracy of target detection and greatly increases the probability of missed detections and false detections.

[0004] A more severe challenge lies in the stringent real-time requirements of autonomous driving systems. Traditional object detection algorithms often employ manual feature extraction methods, requiring researchers to meticulously construct discriminative feature descriptors using their expertise, such as Histogram of Oriented Gradients (HOG) and Local Binary Pattern (LBP). These methods suffer from three fundamental limitations: first, feature design heavily relies on human experience, making it difficult to cover complex and ever-changing real-world scenarios; second, handcrafted features lack robustness to weather-degraded images, exhibiting a sharp decline in expressive power under hazy conditions; and third, the multi-step feature extraction and classification process leads to low computational efficiency, failing to meet the millisecond-level response requirements of automotive-grade systems. Therefore, overcoming the challenge of real-time object detection in adverse weather conditions has become a core breakthrough for driving the leapfrog development of autonomous driving technology.

[0005] In recent years, deep learning has achieved groundbreaking developments in the field of computer vision. Deep learning-based object detection algorithms, with their superior performance, have gradually become mainstream. Currently, deep learning-based object detection algorithms are mainly divided into two-stage detection and one-stage detection. Two-stage object detection is represented by Region-based Convolutional Neural Networks (R-CNN), including Fast R-CNN and Faster R-CNN. The core idea of ​​this type of algorithm is to use a Region Proposal Network (RPN) to generate coarse candidate regions from the input image, and then, while ensuring sufficiently high accuracy and recall, use a fully connected network to classify each proposal and regress its bounding box to determine the specific category and precise location of the target. Although two-stage detection algorithms perform well in terms of detection accuracy, the generation and processing of candidate regions introduce significant computational overhead, which limits the real-time performance of the algorithm. In contrast, one-stage object detection algorithms, such as You Only Look Once (YOLO), Single Shot Multibox Detector (SSD), and RetinaNet, are more suitable for scenarios requiring rapid detection. One-stage detection simplifies into a unified end-to-end process, eliminating the need for additional candidate region generation. It directly performs feature extraction, object classification, and bounding box prediction on the input image, enabling rapid object localization and classification. This concise and efficient network model significantly reduces computational costs and improves algorithm speed while ensuring detection accuracy, giving it a significant advantage in autonomous driving technologies requiring rapid response.

[0006] To improve target detection accuracy in foggy conditions, Zhang Shuqing et al. improved the fifth-generation YOLO (YOLO version, YOLOv5) by employing the BoT3 module based on a self-attention mechanism and a lightweight hybrid attention HAM to enhance the model's ability to capture pedestrians and vehicles in foggy weather. Zhu Qin et al. proposed the spatiotemporal feature fusion module SF-Module and embedded it into the neck of the eighth-generation YOLO (YOLO version, YOLOv8). They used the multi-head self-attention mechanism from the Transformer model to strengthen feature extraction and fusion, and chose the WIou loss function as the regression loss for predicted bounding boxes to reduce harmful gradients generated by low-quality bounding boxes. Pan Zongyou et al. introduced the lightweight dynamic upsampling module Dysample to enhance the model's anti-interference ability based on YOLOv8, used the expandable residual attention module DWR to strengthen feature representation, and replaced the standard detection head with the lightweight shared convolutional detection head LSCD, significantly improving the accuracy and recall of foggy detection.

[0007] Due to the diversity and complexity of target information in different practical application scenarios, current fog detection algorithms still have some limitations, such as the blurring of background textures and severe loss of detailed information such as edge contours in image data, as well as insufficient multi-scale feature fusion. Therefore, it is necessary to improve and optimize the current target detection algorithms in terms of detection accuracy, detection efficiency, and generalization ability. Summary of the Invention

[0008] To address the issues of detail loss, decreased image contrast, and insufficient multi-scale feature fusion faced by autonomous driving technology in target detection tasks under blurred and complex scenarios such as foggy weather, this paper proposes a detail-enhanced channel recalibration fog detection method, DISE-YOLO, based on the 11th generation YOLO (YOLO Version, YOLOv11). This method aims to overcome the shortcomings of current defogging algorithms, further improve the accuracy and speed of fog detection for autonomous vehicles, and promote the development of autonomous driving technology.

[0009] To achieve the above objectives, the present invention provides the following solution: A channel recalibration fog detection method based on detail enhancement includes: Construct an image dataset of haze scenes; the labels in the image dataset include: pedestrians, buses, cars, motorcycles, and bicycles; The image dataset is input into the improved YOLOv11 neural network model for model training to obtain a fog detection model. The improved YOLOv11 neural network model includes: replacing the first layer standard convolution of the backbone network with a detail enhancement convolution module, introducing an iterative attention feature fusion mechanism in the neck network, and embedding a squeeze excitation module before each detection head in the head network. The autonomous vehicle performs detection in foggy weather based on the aforementioned fog detection model.

[0010] Optionally, inputting the image dataset into the improved YOLOv11 neural network model includes: The image is input into the backbone network, and after detail enhancement and multi-scale feature extraction, a multi-scale feature map is output. The multi-scale feature map is input into the neck network, and after feature fusion and enhancement processing, a high-quality fused feature map is output. The fused feature map is input into the detection head, and after channel recalibration and target detection prediction, the target detection result in the foggy scene is output.

[0011] Optionally, inputting the image into the backbone network includes: The image enters the backbone network, where detail enhancement features are obtained through the detail enhancement convolutional module; The detail enhancement features are sequentially passed through several convolutional modules and dynamically configurable modules. The output multi-scale semantic features are then fed into a fast spatial pyramid pooling layer to generate feature representations with multi-scale receptive fields. The feature representation with multi-scale receptive field enters the channel parallel spatial attention mechanism to obtain the multi-scale feature map.

[0012] Optionally, the detail enhancement convolution module adopts a five-branch parallel architecture, including one ordinary convolution and four differential convolutions; The differential convolution includes: central differential convolution, angular differential convolution, horizontal differential convolution, and vertical differential convolution; The central difference convolution and the angle difference convolution rearrange the learned kernel convolution, the horizontal difference convolution and the vertical difference convolution respectively display the encoding horizontal and vertical gradient priors, and the four difference convolutions learn useful gradient information to enhance representation and generalization ability. The five features are added and fused to obtain the final output.

[0013] Optionally, inputting the multi-scale feature map into the neck network includes: The neck network receives the multi-scale feature map and upsamples the high-level semantic feature map to align it with the size of the shallow feature map. The upsampled feature map is concatenated with a shallow feature map of the same scale in the neck network to fuse details and semantic information. The concatenated feature maps are then passed through an additional C3k2_iAFF module to obtain high-quality fused feature maps. The C3k2_iAFF module incorporates dynamic adaptive characteristics into the feature fusion process by introducing an iterative attention feature fusion mechanism. The detection results of the C3k2_iAFF module are used as input to the detection head.

[0014] Optionally, the C3k2_iAFF module adopts a two-stage attention mechanism design; In the first stage, the initial fusion of feature maps is achieved through multi-scale channel attention MS-CAM, which simultaneously aggregates local details and global contextual information; In the second stage, an iterative attention feature fusion mechanism is adopted to further improve the quality of the fused features.

[0015] Optionally, the iterative attention feature fusion mechanism is as follows: in, This represents the initial feature fusion result. This represents multi-scale channel attention module operation. Indicates initial feature integration, It is the Sigmoid activation function. and These represent local feature extraction and global feature extraction, respectively. Represents element-wise multiplication. This indicates the results of the first phase of fusion. This represents the result of the second-stage fusion, i.e., the high-quality features of the final fusion output.

[0016] Optionally, inputting the fused feature map into the detection head includes: Before each Detect is input, feature maps of different scales are first processed by the SEBlock module for global information compression and channel weight recalibration. The SEBlock module is integrated into the output positions of the large, medium and small target detection layers of the YOLOv11 model, so that the model performs channel recalibration on the feature maps before generating the final detection results.

[0017] Optionally, the SEBlock module includes: a squeezing operation and an excitation operation; For the input feature map, the squeezing operation first generates channel statistics through global average pooling, compresses the global spatial information, and compresses it into a channel descriptor; The activation operation forms two bottlenecks with fully connected layers around the nonlinearity through a simple gating mechanism: a dimensionality reduction layer and a dimensionality increase layer. Channel weights are generated by combining a nonlinear activation function. Finally, the output feature map is recalibrated by element-wise multiplication.

[0018] The beneficial effects of this invention are as follows: This invention is based on a channel recalibration fog detection method with detail enhancement. This method constructs DISE-YOLO based on YOLOv11. In the Backbone, the first-layer standard convolution is replaced with a detail enhancement convolution module DEConv. A five-branch parallel architecture is used to achieve collaborative modeling of intensity and gradient, significantly improving the model's ability to capture details such as edges and textures. An iterative attention fusion mechanism iAFF is introduced in the Neck to enhance multi-scale context aggregation capabilities and solve the problem of spurious interference caused by scale bias. Finally, a squeezing excitation module SEBlock is added before each Detect module to adaptively calibrate the channel feature responses of target detection layers at different scales (large, medium, and small), enhancing the synergy between channels.

[0019] This invention aims to overcome the shortcomings of current defogging algorithms, further improve the accuracy and speed of autonomous vehicle detection in foggy weather, and promote the development of autonomous driving technology. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the DISE-YOLO network model structure according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the DEConv module structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the iAFF module structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the SEBlock structure according to an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] This embodiment proposes a channel recalibration fog detection method based on detail enhancement, including: Construct an image dataset of haze scenes; the labels in the image dataset include: pedestrians, buses, cars, motorcycles, and bicycles; The image dataset is input into the improved YOLOv11 neural network model for model training to obtain a fog detection model. The improved YOLOv11 neural network model includes: replacing the first layer standard convolution of the backbone network with a detail enhancement convolution module, introducing an iterative attention feature fusion mechanism in the neck network, and embedding a squeeze excitation module before each Detect. The autonomous vehicle performs detection in foggy weather based on the aforementioned fog detection model.

[0025] Specifically, in this embodiment, to accurately and quickly identify various targets in foggy images, this invention combines artificial intelligence technology, using YOLOv11 as the base network. Compared to previous one-stage algorithms, YOLOv11, with its smaller network structure and more layers of feature information utilization, achieves higher recognition accuracy. Based on the advantages and disadvantages of current fog detection algorithms, the RTTS dataset was selected as the experimental dataset, and an improved DISE-YOLO fog detection method was proposed. The main improvements are as follows: (1) In the backbone network, the first layer of the standard convolution is replaced by the detail-enhanced convolution (DEConv) module. The five-branch parallel architecture is used to achieve the collaborative modeling of intensity and gradient, which enhances the model's ability to perceive high-frequency details such as edges and textures, thereby enhancing the model's feature extraction ability under dense fog conditions.

[0026] (2) Introducing an iterative attentional feature fusion mechanism (iAFF) in the neck enhances the adaptive aggregation of cross-scale contextual information, improves the feature fusion method of YOLOv11, and effectively suppresses pseudo-feature interference caused by scale differences.

[0027] (3) Embed a squeeze-and-excitation block (SEBlock) before each Detect, and through global information compression and channel weight recalibration, make the model's response to key feature channels better and enhance the model's sensitivity to occluded targets and small targets under low visibility conditions.

[0028] This embodiment constructs a DISE-YOLO fog detection method based on YOLOv11. In the Backbone, the first-layer standard convolution is replaced with a detail enhancement convolution module DEConv. A five-branch parallel architecture is used to achieve collaborative modeling of intensity and gradient, which significantly improves the model's ability to capture details such as edges and textures. An iterative attention fusion mechanism iAFF is introduced in the Neck to enhance the ability to aggregate multi-scale contexts to solve the problem of false interference caused by scale bias. Finally, a squeezing excitation module SEBlock is added before each Detect to adaptively calibrate the channel feature responses of target detection layers of different scales (large, medium and small) and enhance the synergy between channels.

[0029] The dataset used in this embodiment is the RTTS (Real-world Task-driven Testing Set) dataset proposed by the RESIDE project team in 2018. This dataset is a benchmark dataset designed specifically for low-visibility object detection tasks in the field of computer vision. All image data were crawled from the Internet and cover diverse and complex haze scenes. It contains 5 categories of labels: 'person', 'bus', 'car', 'motorbike', and 'bicycle', totaling 4,322 images. The dataset is divided in an 8:1:1 ratio. The training set contains 3,457 images with the following number of labels: 9172, 2079, 19914, 951, and 560, respectively. The validation set contains 432 images with the following number of instances for the 5 labels: 1013, 259, 2708, 124, and 62, respectively.

[0030] Furthermore, inputting the image dataset into the improved YOLOv11 neural network model includes: The image is input into the backbone network, and after detail enhancement and multi-scale feature extraction, a multi-scale feature map is output. The multi-scale feature map is input into the neck network, and after feature fusion and enhancement processing, a high-quality fused feature map is output. The fused feature map is input into the detection head, and after channel recalibration and target detection prediction, the target detection results in foggy weather are output.

[0031] Furthermore, inputting the image into the backbone network includes: The image enters the backbone network, where detail enhancement features are obtained through the detail enhancement convolutional module; The detail enhancement features are sequentially passed through several convolutional modules and dynamically configurable modules. The output multi-scale semantic features are then fed into a fast spatial pyramid pooling layer to generate feature representations with multi-scale receptive fields. The feature representation with multi-scale receptive field enters the channel parallel spatial attention mechanism to obtain the multi-scale feature map.

[0032] For fog detection, this embodiment uses a YOLOv11 neural network as the base network and makes improvements, such as... Figure 1 As shown, the improved network takes the dataset images as input images and feeds them into the DISE-YOLO algorithm. The images enter the backbone and first pass through the detail enhancement convolution module DEConv.

[0033] The detail enhancement convolution module DEConv employs a five-branch parallel architecture, comprising one regular convolution (Convolution, Conv) and four difference convolutions (DC): Central Difference Convolution (CDC), Angular Difference Convolution (ADC), Horizontal Difference Convolution (HDC), and Vertical Difference Convolution (VDC). It simultaneously focuses on both intensity and gradient dimensions through a multi-path feature parallelism strategy. Its topology is as follows: Figure 2 As shown. The Conv branch is mainly responsible for semantic modeling of intensity and texture. The center difference convolution (CDC) and angle difference convolution (ADC) rearrange the learned kernel convolutions, saving computational cost and memory consumption while enhancing edge detection capabilities. The horizontal difference convolution (HDC) and vertical difference convolution (VDC) explicitly encode horizontal and vertical gradient priors, respectively, learning useful gradient information to enhance representation and generalization capabilities. Finally, the five features are added and fused to obtain the final output, as shown in the following formula: (1); in, ( These correspond to the kernel weights of Conv, CDC, ADC, HDC, and VDC, respectively. This represents a convolution operation. To optimize inference efficiency, DEConv also introduces a parameter re-parameter technique, which treats the parallel convolution kernel as a single standard convolution kernel during the training phase. During reasoning, formula (2) is satisfied: (2); The feature map output from the DEConv module is fed into several convolutional modules (Conv) and a dynamically configurable module (C3k2). The dynamic parameters in the first two sets of C3k3 modules are set to False, ensuring the bottleneck module (Bottleneck) retains its original structure. The dynamic parameters in the latter two sets of C3k2 modules are set to True, automatically switching the bottleneck module to the local feature enhancement module (C3k). Then, it enters a fast spatial pyramid pooling (SPPF) layer to generate feature representations with multi-scale receptive fields. This is followed by a cross-channel parallel spatial attention (C2PSA) mechanism, which is then fed into the neck network.

[0034] (3) The Neck receives the multi-scale feature maps from the Backbone and enlarges the size of the high-level semantic feature maps through the Upsample operation to align them with the size of the shallow feature maps. The upsampled feature maps are then concatenated with the shallow feature maps of the same scale in the Backbone to fuse details and semantic information. The three concatenated feature maps are then processed by an additional C3k2_iAFF module.

[0035] This module introduces the Iterative Attention Feature Fusion (iAFF) mechanism, integrating dynamic adaptive features into the feature fusion process. This addresses the issues of simple initial integration in ordinary feature fusion methods, dynamically integrating inconsistent features. Its structure diagram is shown below. Figure 3 As shown.

[0036] Ordinary convolutional neural networks often use simple addition or cascading for feature fusion. While this method provides the model with a fixed linear aggregation of feature maps, this simple combination, which does not consider the degree of adaptation between sub-feature maps, has obvious flaws in context awareness. Given two sub-feature maps X and Y to be fused, both of which are of channel size (C),... Height (H) Width (W) (as symbol) (represented), the fusion feature F is abstracted into mathematical expressions as shown in formulas (3) and (4): (3); (4); In most cases, such fusion schemes can only partially perceive the input feature map, relying solely on high-level feature maps to train the model's learning ability, and lack the ability to perceive dynamic changes in feature content.

[0037] iAFF is an extension of the basic Attentional Feature Fusion (AFF) mechanism, optimizing the initial feature integration by adding Multi-scale Channel Attention (MS-CAM). While standard AFF directly uses simple operations (such as addition) to initially integrate features, iAFF uses the output of the first AFF module as the initial integration input for the second AFF module, thereby dynamically generating more appropriate attention weights. The core idea of ​​C3k2_iAFF lies in its two-stage attention mechanism design.

[0038] In the first stage, initial feature mapping fusion is achieved through MS-CAM. This module simultaneously aggregates local details and global contextual information, helping the network focus on objects with less background clutter. In the second stage, an iterative attention feature fusion strategy is employed to further improve the quality of the fused features. The mathematical expression is as follows: (5); (6); (7); in, This represents the initial feature fusion result. This represents multi-scale channel attention module operation. Indicates initial feature integration, It is the Sigmoid activation function. and These represent local feature extraction and global feature extraction, respectively, implemented through pointwise convolution (PWConv) and global average pooling (GAP). Represents element-wise multiplication. This indicates the results of the first phase of fusion. This represents the result of the second-stage fusion, i.e., the high-quality features of the final fusion output.

[0039] From a module implementation perspective, this invention utilizes Bottleneck_iAFF to achieve feature reconstruction. An attention weight learning mechanism is introduced into the Neck, which maintains the gradient flow of residual learning while enhancing the adaptability of feature selection, thus overcoming the limitations of traditional feature fusion methods.

[0040] (4) The detection results of the C3k2_iAFF module are used as input to the Head part. Before each Detect is input, the feature maps of three different scales will first pass through the SEBlock module for global information compression and channel weight recalibration. Its structure diagram is as follows. Figure 4 As shown, this module consists of a squeeze operation and an excitation operation, which are responsible for global information embedding and adaptive feature recalibration, respectively.

[0041] For the input feature map X ( The Squeeze operation first generates channel statistics through Global Average Pooling (GAP), compressing the global spatial information into a single channel descriptor. This is equivalent to aggregating the spatial information of each channel into a scalar, enabling the network to better integrate contextual information from within and outside the global receptive field.

[0042] To utilize the aggregated information during compression in a timely manner, the SE mechanism performs an excitation operation. Excitation uses a simple gating mechanism to create two bottlenecks around the nonlinearity, each with a fully connected layer (FC): a dimensionality reduction layer and a dimensionality increase layer. Channel weights are generated using a nonlinear activation function (such as ReLU). Finally, the output feature map F is recalibrated through element-wise multiplication. The formula is as follows: (8); (9); (10); (11); (12); Where z represents the channel descriptor generated by the extrusion operation, and each element They all contain global information for the corresponding channel c. and Each layer consists of a dimensionality reduction layer and a dimensionality increase layer, reducing the dimension from c to c. Then change the dimension from Revert to c, with the reduction rate r set to 16 by default.

[0043] In foggy images, the imbalance between channels is exacerbated, with some channels dominated by fog noise. Integrating the SEBlock module into the output positions of the large, medium, and small target detection layers in the YOLOv11 model allows the model to recalibrate the feature maps before generating the final detection results, focusing more on channels that still retain target information, thus significantly improving fog detection performance.

[0044] Ultimately, the network predicts multiple bounding boxes for each grid cell on the feature map. Each bounding box contains center coordinates, width and height, confidence score, and probability of each class.

[0045] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A channel recalibration fog detection method based on detail enhancement, characterized in that, include: Construct an image dataset of haze scenes; the labels in the image dataset include: pedestrians, buses, cars, motorcycles, and bicycles; The image dataset is input into the improved YOLOv11 neural network model for model training to obtain a fog detection model. The improved YOLOv11 neural network model includes: replacing the first layer standard convolution of the backbone network with a detail enhancement convolution module, introducing an iterative attention feature fusion mechanism in the neck network, and embedding a squeeze excitation module in front of each detection head in the head network. The autonomous vehicle performs detection in foggy weather based on the aforementioned fog detection model.

2. The channel recalibration fog detection method based on detail enhancement according to claim 1, characterized in that, Inputting the image dataset into the improved YOLOv11 neural network model includes: The image is input into the backbone network, and after detail enhancement and multi-scale feature extraction, a multi-scale feature map is output. The multi-scale feature map is input into the neck network, and after feature fusion and enhancement processing, a high-quality fused feature map is output. The fused feature map is input into the detection head, and after channel recalibration and target detection prediction, the target detection result in the foggy scene is output.

3. The channel recalibration fog detection method based on detail enhancement according to claim 2, characterized in that, Inputting images into the backbone network includes: The image enters the backbone network, where detail enhancement features are obtained through the detail enhancement convolutional module; The detail enhancement features are sequentially passed through several convolutional modules and dynamically configurable modules. The output multi-scale semantic features are then fed into a fast spatial pyramid pooling layer to generate feature representations with multi-scale receptive fields. The feature representation with multi-scale receptive field enters the channel parallel spatial attention mechanism to obtain the multi-scale feature map.

4. The channel recalibration fog detection method based on detail enhancement according to claim 3, characterized in that, The detail-enhancing convolutional module adopts a five-branch parallel architecture, including one ordinary convolution and four differential convolutions; The differential convolution includes: central differential convolution, angular differential convolution, horizontal differential convolution, and vertical differential convolution; The central difference convolution and the angle difference convolution rearrange the learned kernel convolution, the horizontal difference convolution and the vertical difference convolution respectively display the encoding horizontal and vertical gradient priors, and the four difference convolutions learn useful gradient information to enhance representation and generalization ability. The five features are added and fused to obtain the final output.

5. The channel recalibration fog detection method based on detail enhancement according to claim 2, characterized in that, Inputting the multi-scale feature map into the neck network includes: The neck network receives the multi-scale feature map and upsamples the high-level semantic feature map to align it with the size of the shallow feature map. The upsampled feature map is concatenated with a shallow feature map of the same scale in the neck network to fuse details and semantic information. The concatenated feature maps are then passed through an additional C3k2_iAFF module to obtain high-quality fused feature maps. The C3k2_iAFF module incorporates dynamic adaptive characteristics into the feature fusion process by introducing an iterative attention feature fusion mechanism. The detection results of the C3k2_iAFF module are used as input to the detection head.

6. The channel recalibration fog detection method based on detail enhancement according to claim 5, characterized in that, The C3k2_iAFF module adopts a two-stage attention mechanism design. In the first stage, the initial fusion of feature maps is achieved through multi-scale channel attention. This module simultaneously aggregates local details and global contextual information. In the second stage, an iterative attention feature fusion mechanism is adopted to further improve the quality of the fused features.

7. The channel recalibration fog detection method based on detail enhancement according to claim 1, characterized in that, The iterative attention feature fusion mechanism is as follows: in, This represents the initial feature fusion result. This represents multi-scale channel attention module operation. Indicates initial feature integration, It is the Sigmoid activation function. and These represent local feature extraction and global feature extraction, respectively. Represents element-wise multiplication. This indicates the results of the first phase of fusion. This represents the result of the second-stage fusion, i.e., the high-quality features of the final fusion output.

8. The channel recalibration fog detection method based on detail enhancement according to claim 2, characterized in that, Inputting the fused feature map into the detection head includes: Before each Detect is input, feature maps of different scales are first processed by the squeezing excitation module for global information compression and channel weight recalibration. The squeezing excitation module is integrated into the output positions of the large, medium and small target detection layers of the YOLOv11 model, so that the model performs channel recalibration on the feature maps before generating the final detection results.

9. The channel recalibration fog detection method based on detail enhancement according to claim 8, characterized in that, The extrusion excitation module includes: extrusion operation and excitation operation; For the input feature map, the squeezing operation first generates channel statistics through global average pooling, compresses the global spatial information, and compresses it into a channel descriptor; The activation operation forms two bottlenecks with fully connected layers around the nonlinearity through a simple gating mechanism: a dimensionality reduction layer and a dimensionality increase layer. Channel weights are generated by combining a nonlinear activation function. Finally, the output feature map is recalibrated by element-wise multiplication.

Images