Real-time foggy day target detection method based on GhostConv feature extraction

Abstract

The invention discloses a real-time foggy day target detection method based on GhostConv feature extraction, and relates to the field of image target detection. The method comprises the following steps: preprocessing data; feature extraction: carrying out feature extraction on the preprocessed image by adopting a BackBone backbone network, wherein the BackBone backbone network comprises a phantom multi-scale convolution module; feature fusion; performing feature decoding; the loss function correction model is used for respectively calculating classification loss and positioning loss of the feature map of each input decoupling detection head by utilizing a loss function to obtain total loss, and network parameters of corresponding parts of the model are calculated and adjusted through back propagation gradient; and target detection: carrying out target detection on a to-be-detected foggy day image by adopting the corrected model. According to the method, a phantom multi-scale convolution module is adopted, the model feature extraction capability is improved under the condition that semantic information loss is reduced as much as possible, and shape and complete IoU overlapping loss is proposed.

Description

Technical Field

[0001] The invention belongs to the technical field of image target detection, and in particular relates to a real-time foggy target detection method based on GhostConv feature extraction. Background Art

[0002] In the field of computer vision, object detection is an important branch that aims to classify, locate and regress the target box at the same time. In foggy weather, there are problems such as low visibility of the target object and the object and background are often obscured by fog, which makes the performance of the target detection algorithm often less than ideal. In real scenarios, foggy object detection has a wide range of application scenarios, such as autonomous driving, industrial monitoring systems, etc.

[0003] In recent years, there are two main frameworks for target detection algorithms: DETR (Detection Transformer) series algorithms and YOLO (You Only Look Once) series algorithms. In recent years, the YOLO series of algorithms has developed rapidly, from the initial YOLOv1 to the current YOL Ov10. During this period, they compared the advantages and disadvantages of algorithms with anchor boxes and algorithms without anchor boxes, discussed the difference between non-maximum suppression (NMS), and derived from a detection head to a multi-scale detection head and then to today's multi-scale decoupled detection head. These methods have made great breakthroughs in detection performance.

[0004] The target detection algorithm based on the YOLO series is one of the hot topics that has attracted much attention. Among them, YOLOv10 is the latest algorithm in the YOLO series. It has achieved the best performance on the COCO dataset, and the number of parameters and calculations are reduced compared to YOLOv8. Based on YOLOv8, YOLOv10 cancels NMS, uses a one-to-one and one-to-many dual-label assignment strategy in the model training stage, and uses a consistent matching metric formula to calculate the difference between the two label assignment trainings, and then uses the metric score to make the detection performance of the one-to-one label assignment as close as possible to the detection performance of the one-to-many label assignment. Only a one-to-one label assignment strategy is used during model inference to reduce the amount of model calculation and speed up the inference speed. However, when using the YOLOv10 target detection algorithm for foggy target detection, the following problems still exist:

[0005] 1) Although the YOLOv10 target detection algorithm proposed a lightweight classification head, spatial decoupled channel downsampling and large kernel convolution method based on YOLOv8, while reducing the parameters and computational complexity of the model while keeping the accuracy as unaffected as possible, and made a balance between the performance and training cost of the model as much as possible. However, in foggy weather, light will be scattered and absorbed by water droplets and tiny particles suspended in the air, resulting in a significant reduction in the contrast of the image. This reduces the visual difference between the target object and the background, making it difficult for the model network to distinguish the target object features from the image background, and it is difficult to locate the exact position of the target object and it also has a certain impact on the recall rate. These problems have not been well solved in YOLOv10.

[0006] 2) Bounding box regression loss plays an important role in the field of target detection as an important indicator to measure detection accuracy. Most current target detection algorithms usually consider the geometric relationship between the true box and the predicted box, such as the IoU between the detection box and the true box, the ratio of the center distance to the diagonal of the minimum circumscribed rectangle, and the aspect ratio. The loss is calculated using information such as the position and shape of the detection box, while ignoring the influence of the inherent properties of the box itself, such as its shape and scale, on the bounding box regression. Specifically: In the existing YOLOv10 method, if the same offset is used, two loss values ​​with large differences will be obtained along the long side offset and the short side offset, which is very unfavorable for loss calculation and adjustment of model parameters.

[0007] 3) When designing the consistent matching metric formula, the YOLOv10 network multiplies the target space prior, the regression score, and the intersection over union (IoU) of the detection box. This makes the metric score sensitive to the regression score and IoU. Specifically, when one of the two is larger and the other is smaller, the metric score obtained by multiplying them will also be smaller. The detection performance under the one-to-one label assignment strategy needs to use the metric score to approach the detection performance under the one-to-many label assignment strategy. This phenomenon will cause the metric score to fluctuate greatly, seriously affecting the detection performance under the one-to-one label assignment strategy, and is very unfavorable for the training of one-to-one label assignment detection.

[0008] In addition, people are currently paying much attention to the research on Ghost Conv. Network modules based on GhostConv are also emerging in an endless stream, such as: C3Ghost, GhostBottleneck and GhostBlock. GhostConv significantly reduces the computational burden and parameter amount of the model by introducing Ghost Filters and reducing the number of feature channels. Moreover, GhostConv is not only suitable for static image processing, but can also be applied to video analysis, speech recognition and other fields. This wide applicability makes GhostConv a general optimization tool in multiple fields.

[0009] The present invention proposes a real-time foggy target detection method based on GhostConv feature extraction, which adopts the ghost multi-scale convolution GMSConv module to improve the model feature extraction capability while minimizing the loss of semantic information as much as possible, and adopts shape and complete IoU overlap loss to take into account the shape and size information of the border itself to make the overlap loss part more reasonable and reduce the fluctuation of the metric score. Summary of the invention

[0010] The purpose of the present invention is to provide a real-time foggy target detection method based on GhostConv feature extraction, so as to solve the problems in the prior art of foggy target detection based on the YOLOv10 target detection algorithm, such as the limitations of image feature extraction, the detection performance under the one-to-one label assignment strategy is affected by the metric score, and the loss function ignores the influence of inherent attributes such as the shape and scale of the border itself on the border regression.

[0011] To achieve the above purpose, the technical solution adopted by the present invention is:

[0012] The first aspect of the present invention proposes a real-time foggy target detection method based on GhostConv feature extraction, comprising the following steps:

[0013] S1. Data preprocessing: collecting image data, performing image enhancement on the collected images, and fixing the image size to a uniform size;

[0014] S2, feature extraction: The BackBone backbone network is used to extract features from the preprocessed image. The phantom multi-scale convolution module is added to the BackBone backbone network to map the semantic information of the RGB three channels to a high-dimensional space to obtain feature maps of multiple scales;

[0015] S3, feature fusion: Neck network is used to splice and fuse feature maps of multiple scales to obtain feature information of multiple scales;

[0016] S4, feature decoding: Decoupled Head is used to decode feature information of different sizes, each feature information generates a positioning information and classification information, and the positioning information and classification information are translated to obtain classification and positioning results;

[0017] S5, loss function correction model; train the real-time foggy target detection model based on GhostConv feature extraction, use the loss function to calculate the classification loss and positioning loss of each input Decoupled Head decoupled detection head feature map to obtain the total loss; based on the final loss value, adjust the network parameters of the corresponding part of the model through back propagation gradient calculation;

[0018] S6. Target detection: using the calibrated real-time foggy target detection model based on GhostConv feature extraction, target detection is performed on the foggy image to be detected.

[0019] Preferably, the image enhancement in S1 includes performing image enhancement by random cropping, random rotation, and random occlusion.

[0020] Preferably, the phantom multi-scale convolution module in S2 is specifically as follows:

[0021] If the input feature is F1, first perform two phantom convolutions on the input feature F1 with the same number of channels and a convolution kernel size of 1×1, and obtain F2 and F3 respectively; then concatenate F2 and F3 together to obtain F4, and then restore the number of channels of F4 through phantom convolution to obtain F5; finally, add F5 and F1 through residual link to obtain the final output; the process of the phantom multi-scale convolution module is as follows

[0022] F4=Concat(GhostConv(F1), (GhostConv(GhostConv(F1))))

[0023] Output=GhostConv(F4)+F1

[0024] Among them, GhostConv is a ghost convolution operation.

[0025] Preferably, the phantom convolution is specifically as follows:

[0026] If the input feature is S1, first perform a Conv operation on the input feature S1 with a convolution kernel size of 1×1 and half the number of channels to obtain S2; then use a grouped Conv operation with the same number of channels and a convolution kernel size of 3×3 to obtain S3; then concatenate S2 and S3 to obtain the final output feature tensor s4, The Con v process is expressed as follows:

[0027] Output=SiLu(BatchNorm(Conv2d(Input)))

[0028] Among them, Conv2d represents a normal two-dimensional convolution.

[0029] Preferably, the BackBone backbone network in S2 is as follows:

[0030] After the image is input into the BackBone backbone network, the semantic information is mapped into a high-dimensional space through two layers of Conv, and then input into the phantom multi-scale convolution module. Two phantom multi-scale convolution modules are added to the shallow network; the BackBone backbone network generates feature maps of multiple scales (P3, P4, P5).

[0031] Preferably, the Neck network in S3 is as follows:

[0032] The Neck network is composed of a feature pyramid. Through the feature pyramid, P3, P4, and P5 are first upsampled and then downsampled. During the sampling process, the extracted P3, P4, and P5 feature maps are synthesized into a new tensor without loss through splicing, and then channel adjustment and information fusion are performed.

[0033] Preferably, the Decoupled Head in S4 is specifically as follows:

[0034] After feature fusion, feature information of multiple scales is obtained. Three feature maps with different widths and heights are selected. The three feature maps are combined for target classification and positioning. The decoding process is shown as follows:

[0035] Bbox=Conv(Conv(Conv2d(Feature map)))

[0036] Cls=Conv(Conv(Conv2d(Feature map)))

[0037] Among them, Bbox and Cls represent the positioning and classification results of the decoupling head, and Conv2d represents an ordinary two-dimensional convolution.

[0038] Preferably, the loss function in S5 uses shape and complete IoU overlap loss to calculate the positioning loss, and its formula can be expressed as:

[0039]

[0040]

[0041] Among them, B p , B gt is the predicted box and the real box, c is the diagonal length of the minimum circumscribed matrix of the predicted box and the real box; ω w and ω h are the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the real frame; w gt 、h gt Indicates the length and width of the real box.

[0042] Preferably, the classification loss in the loss function in S5 is calculated by binary cross entropy, and its formula is expressed as:

[0043]

[0044] Where N is the number of samples, y i is the true label of the i-th sample, p i is the predicted probability of the ith sample.

[0045] Preferably, the real-time foggy target detection model based on GhostConv feature extraction in S5 includes a BackBone backbone network, a Neck network, and a Decoupled Head decoupled detection head.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) The present invention aims at the problem that the backbone network of the YOLOv10 target detection algorithm still has limitations in its ability to extract image features when facing the target detection task in foggy scenes; the present invention adopts a ghost multi-scale convolution GMSConv module, which includes a ghost convolution GhostConv module, to achieve the extraction of semantic features on multi-size convolution kernel sizes and multi-depth semantic information, thereby improving the model feature extraction capability while minimizing the loss of semantic information as much as possible.

[0048] (2) Although the loss function of the existing YOLOv10 in the present invention takes into account the intersection-over-union ratio of the detection box and the true box, the ratio of the center distance of the box to the diagonal of the minimum circumscribed rectangle, and the aspect ratio of the box, these are the relative shape and relative position information of the box, and ignore the impact of the inherent attributes of the box itself, such as the scale, on the box regression. The present invention proposes a shape and complete IoU overlap loss, which also takes into account the shape and size information of the box itself, making the overlap loss part more reasonable and accelerating the convergence speed of the model during training.

[0049] (3) The present invention aims to solve the problem that the detection performance under the one-to-one label assignment strategy is affected by the metric score. The present invention reduces the metric score fluctuation and the impact on the detection performance under the one-to-one label assignment strategy by proposing shape and complete IoU overlap losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a structural block diagram of the real-time foggy target detection model based on GhostConv feature extraction in the present invention;

[0051] Figure 2 It is a structural block diagram of the phantom multi-scale convolution module in the present invention;

[0052] Figure 3 It is a structural block diagram of the phantom convolution in the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Embodiment 1:

[0055] The real-time foggy target detection method based on GhostConv feature extraction can be divided into five stages: data preprocessing, image feature extraction, image feature fusion, information decoding and prediction, and loss function correction. The details are as follows:

[0056] Step 1: data preprocessing.

[0057] In order to cope with inputs of different image sizes, it is convenient to input images into the network for feature extraction and prediction, enhance the robustness of the model, and perform data preprocessing.

[0058] First, image enhancement is required, including but not limited to random cropping, blurring, random occlusion, translation and flipping. These enhancement methods can effectively improve the robustness of the model. Then fix images of different sizes to a uniform size. In order to prevent the image from being distorted, it is generally chosen to scale the image according to the original ratio and blacken the empty pixels. Then stitch several images of uniform size together to further enhance the image. Finally, fix the size of the image to be input to 640×640, and then normalize each pixel. This operation can fix the pixels of each channel in a certain range (such as: between 0 and 1 or -1 and 1), and does not change their size order when changing the value. This method can effectively improve the computing efficiency of the computer.

[0059] Step 2: feature extraction.

[0060] The BackBone backbone network is used for image feature extraction. The preprocessed image is input into the backbone network to map the semantic information of the RGB three channels into a high-dimensional space. Then, feature maps of multiple scales (P3, P4, P5) are generated. These feature maps come from different levels of network structures, so they contain semantic information at different levels, which facilitates the network to identify target objects of different sizes.

[0061] The BackBone backbone network uses the Ghost multi-scale convolution GMSConv (Ghost multi-scale convolution) module. The function of this network module is to extract semantic features based on multi-size convolution kernel sizes and multi-depth semantic information, and to improve the model feature extraction capability while minimizing the loss of semantic information. The module structure is as follows: Figure 2 shown.

[0062] In this embodiment, the preprocessed image is input into the backbone network, the semantic information of the RGB three channels is mapped to a high-dimensional space, and then feature maps of three scales are generated.

[0063] Specifically, after the preprocessed image is input into the backbone network, it passes through two layers of Conv to map the semantic information into a high-dimensional space, and then enters the GMSConv module. Its structure is as follows: Figure 2 As shown. Assuming that the input feature is F1, the GMSConv module first performs two ghost convolutions (GhostConv) on the input features with the same number of channels and a convolution kernel size of 1×1 to obtain F2 and F3. This method can extract deep semantic information, thereby accurately distinguishing target objects from background information, and extracting those target features that are difficult to capture. F2 and F3 are then spliced ​​together to obtain F4. Because the levels of GhostConv are different, the tensor F4 contains semantic information at different levels, and the splicing method allows the semantic information to be completely preserved. The number of channels is then restored through GhostConv to obtain F5. This operation reduces model parameters while preventing the loss of important semantic information. Finally, F5 and F1 are added through the residual connection to obtain the final output. The residual connection can effectively prevent the model gradient from disappearing, which is beneficial to the adjustment of parameters during model training. The process of GMSConv is expressed as follows

[0064] F4=Concat(GhostConv(F1), (GhostConv(GhostConv(F1))))

[0065] Output=GhostConv(F4)+F1

[0066] The main working principle of the ghost convolution module is: assuming the input feature is S1, the ghost convolution (GhostConv) module first performs a Conv operation on the input feature with a convolution kernel size of 1×1 and half the number of channels to obtain S2, where the 1×1 convolution kernel can aggregate information features between channels. Then use a grouped Conv operation with the same number of channels and a convolution kernel size of 3×3 to obtain S3, where each channel is divided into a group. Then S2 and S3 are concatenated to obtain the final output feature tensor S4; the module structure is as follows Figure 3 The Conv process is shown as follows

[0067] Output=SiLu(BatchNorm(Conv2d(Input))).

[0068] In the entire backbone network, considering the trade-off between the change of model parameters and the ability of feature extraction, only two GMSConv modules are added to the shallow network. In the shallow network, the feature map obtained is small and contains rich potential semantic information, which not only reduces the number of model parameters and calculations, but also fully extracts the potential information in the high-dimensional space.

[0069] Step three: feature fusion.

[0070] The Neck network is used for image feature fusion. After the image passes through the backbone network, the low-dimensional semantic information is expanded to a high-dimensional space. The Neck network is mainly composed of feature pyramids, which fully integrate feature information during upsampling and downsampling. It aims to improve the expressiveness and adaptability of features, so as to better capture the semantic information and spatial relationship of the target object.

[0071] Through the feature pyramid method, the feature information extracted by the backbone is first upsampled and then downsampled. During the sampling process, the P3, P4, and P5 feature maps extracted by the backbone network are synthesized into a new tensor without loss by splicing, and then input into the module for channel adjustment and information fusion. It aims to improve the expressiveness and adaptability of features, so as to better capture the semantic information and spatial relationship of the target object.

[0072] Step 4: Feature decoding and prediction.

[0073] Decoupled Head is used for classification and positioning. Detection tensors of different sizes are input into the decoupled detection head. After processing, each detection tensor will obtain a classification tensor and a positioning tensor. The final classification and positioning results can be obtained by parsing these two tensors. In the training phase of YOLOv10, two label assignment strategies, one-to-one and one-to-many, are used. The balance between the two requires the use of a compensating directional consistent matching metric. In the inference phase, only a one-to-one label assignment strategy is used.

[0074] In the feature fusion stage, many feature maps are generated, and then three of them are selected as prediction samples. These three feature maps have different widths and heights. The larger the feature map, the more effective it is for detecting small targets. Because the larger the feature map, the smaller the area of ​​the original image corresponding to each element, and vice versa. Then the three feature maps are combined for target classification and positioning. The specific decoding process can be expressed as follows:

[0075] Bbox=Conv(Conv(Conv2d(Feature map)))

[0076] Cls=Conv(Conv(Conv2d(Feature map)))

[0077] Among them, Bbox and Cls represent the positioning and classification results of the decoupling head. Conv2d represents a common two-dimensional convolution.

[0078] Step 5: Loss function correction.

[0079] The above is the whole process of the model inference stage. When the model is in the training stage, it is necessary to calculate the feature map classification loss and positioning loss of each input detection head separately, and then combine them to get the total loss of the model. It should be noted that YOLOv10 will train two label assignment strategies at the same time during the training stage, and the total loss is obtained by combining the losses of the two label assignment strategies. After the model obtains the final loss value, it adjusts the network parameters of the corresponding part through back-propagation gradient calculation to achieve the purpose of learning.

[0080] Specifically, in this embodiment, the classification loss is calculated by binary cross entropy (BCE), and its formula can be expressed as:

[0081]

[0082] Where N is the number of samples, y i is the true label of the i-th sample, p i is the predicted probability of the ith sample.

[0083] Specifically, in this embodiment, in order to take into account the shape and size information of the bounding box itself, make the overlap loss more reasonable, and accelerate the convergence speed of the model during training, the present invention proposes shape and complete IoU overlap loss.

[0084] This loss improves CIoU (Complete IoU). During model training, CIoU only considers the relative shape and position information of the border, ignoring the impact of the border's own attributes on the judgment result. CIoU can be expressed as:

[0085]

[0086] Among them, B p , B gt is the predicted box and the real box, ρ(b, b gt ) is the distance between the center of the predicted box and the real box, and c is the diagonal length of the minimum circumscribed matrix of the predicted box and the real box. From the formula, we can see that whether it is the ratio of the center distance of the border to the diagonal of the minimum circumscribed rectangle or the difference in the aspect ratio of the border, these information are all relative information of the border, and do not take into account the impact of the border's own attributes on the degree of overlap.

[0087] In the present invention, the network uses the shape and complete IoU overlap loss (SCIoU) to calculate the positioning loss according to the tensor obtained by the detection head, and the regression loss is calculated using SCIoU (Shape And Complete IoU), which can be expressed as:

[0088]

[0089] Among them, B p , B gt is the predicted box and the real box, c is the diagonal length of the minimum circumscribed matrix of the predicted box and the real box; ω w and ω h are the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the real frame; w gt 、h gt Indicates the length and width of the real box.

[0090] The calculated losses are used to calculate the gradients of each level of the model, thereby adjusting the model parameters appropriately.

[0091] The present invention adopts the above method to train and obtain a real-time foggy target detection model based on GhostConv feature extraction, and the model includes a BackBone backbone network, a Neck network, and a Decoupled Head decoupled detection head.

[0092] When using the real-time foggy target detection model based on GhostConv feature extraction for target detection, specifically: in the feature extraction stage, the foggy image is input into the backbone network, the semantic information of the RGB three channels is mapped to the high-dimensional space, and then feature maps of multiple scales are obtained. In the feature fusion stage, the image feature information extracted from the backbone network is fused through upsampling and downsampling of the feature pyramid. In the information decoding stage, the multiple feature information after fusion are combined into a feature sequence, which is then input into the detection head for decoding. Finally, each feature information will generate a positioning information and classification information, and the classification and positioning results of the model can be obtained by translating the information. Models of size n such as Figure 1 As shown in the figure, models of different sizes only have different modules that make up the network, and the overall process remains unchanged.

[0093] Experimental verification:

[0094] Based on the above invention content, the following technical effect evaluation is proposed:

[0095] Performance Evaluation,The present paper is experimented and evaluated on the RTTs dataset.,This dataset contains 4423 high-resolution real foggy images, of which the training set contains 4323 images and the test set contains 100 images.,This dataset includes five categories: pedestrians, cars, buses, bicycles and motorcycles.

[0096] The benchmark model YOLOv10 has multiple parameter versions. Considering the usage scenario, the model is required to achieve real-time detection. Therefore, this invention selects two models with the smallest number of parameters, n and s, as the benchmark model. The test results on the RTTs dataset are shown in Table 1:

[0097] Table 1 Comparative test results on the RTTs dataset

[0098] Model Parameters GFLOPs(G) P(％) R(％) mAP50(%) YOLOv10-n 2686366 8.2 67.0 48.3 55.2 YOLOv8-n 3006623 8.1 68.3 54.8 58.5 Ours-n 2861422 11.6 70.0 57.5 61.5 YOLOv10-s 8038830 24.5 70.2 51.7 60.1 YOLOv8-s 11127519 28.4 72.8 50.9 59.3 Ours-s 8229486 28.4 73.1 57.2 65.6

[0099] At the n scale, compared with the baseline model, although the number of parameters and the amount of calculation of the present invention are increased by 6.5% and 41.5% respectively, the average accuracy, average recall rate and the average accuracy of fifty retrieval results are increased by 3, 9.2 and 6.3 percentage points respectively.

[0100] Under the s-size, compared with the baseline model, although the number of parameters and the amount of calculation of the present invention are increased by 2.4% and 15.9% respectively, the accuracy, recall rate and the average accuracy of fifty retrieval results are increased by 2.9, 5.5 and 5.5 percentage points respectively.

[0101] Whether it is the n model or the s model, the number of parameters and the amount of calculation have increased, but if Ours-n is compared with YOLOv10-s, the number of parameters and the amount of calculation are reduced by 66.4% and 52.7% respectively, and the average accuracy of the fifty retrieval results increased by 1.4 percentage points.

[0102] The above description is only used to help understand the method of the present invention and its core essence, but the protection scope of the present invention is not limited thereto. For those skilled in the art in the art, equivalent replacement or change according to the technical solution and inventive concept of the present invention within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A real-time foggy target detection method based on GhostConv feature extraction, characterized in that: The steps include: S1. Data preprocessing: collecting image data, performing image enhancement on the collected images, and fixing the image size to a uniform size; S2, feature extraction; The BackBone backbone network is used to extract features from the preprocessed images. A phantom multi-scale convolution module is added to the BackBone backbone network to map the semantic information of the RGB three channels into a high-dimensional space to obtain feature maps of multiple scales. S3, feature fusion: Neck network is used to splice and fuse feature maps of multiple scales to obtain feature information of multiple scales; S4, feature decoding: Decoupled Head is used to decode feature information of different sizes, each feature information generates a positioning information and classification information, and the positioning information and classification information are translated to obtain classification and positioning results; S5, loss function correction model; train the real-time foggy target detection model based on GhostConv feature extraction, use the loss function to calculate the classification loss and positioning loss of each input Decoupled Head decoupled detection head feature map to obtain the total loss; based on the final loss value, adjust the network parameters of the corresponding part of the model through back propagation gradient calculation; S6. Target detection: using the calibrated real-time foggy target detection model based on GhostConv feature extraction, target detection is performed on the foggy image to be detected.

2. The real-time foggy target detection method based on GhostConv feature extraction according to claim 1 is characterized in that: The phantom multi-scale convolution module in S2 is as follows: If the input feature is F1, first perform two phantom convolutions on the input feature F1 with the same number of channels and a convolution kernel size of 1×1, and obtain F2 and F3 respectively; then concatenate F2 and F3 together to obtain F4, and then restore the number of channels of F4 through phantom convolution to obtain F5; finally, add F5 and F1 through residual connection to obtain the final output; the process of the phantom multi-scale convolution module is expressed as follows: F4=Concat(GhostConv(F1),(GhostConv(GhostConv(F1)))) Output=GhostConv(F4)+F1 Among them, GhostConv is a ghost convolution operation.

3. The real-time foggy target detection method based on GhostConv feature extraction according to claim 2 is characterized in that: The phantom convolution is specifically as follows: If the input feature is S1, first perform a Conv operation on the input feature S1 with a convolution kernel size of 1×1 and half the number of channels to obtain S2; then use a grouped Conv operation with the same number of channels and a convolution kernel size of 3×3 to obtain S3; then concatenate S2 and S3 to obtain the final output feature tensor S4; the Conv process is expressed as follows: Output=SiLu(BatchNorm(Conv2d(Input))) Among them, Conv2d represents a normal two-dimensional convolution.

4. The real-time foggy target detection method based on GhostConv feature extraction according to claim 3 is characterized in that: The BackBone backbone network in S2 is as follows: After the image is input into the BackBone backbone network, the semantic information is mapped into a high-dimensional space through two layers of Conv, and then input into the phantom multi-scale convolution module. Two phantom multi-scale convolution modules are added to the shallow network; the BackBone backbone network generates feature maps of multiple scales (P3, P4, P5).

5. The real-time foggy target detection method based on GhostConv feature extraction according to claim 4 is characterized in that: The S3 Neck network is as follows: The Neck network is composed of a feature pyramid. Through the feature pyramid, P3, P4, and P5 are first upsampled and then downsampled. During the sampling process, the extracted P3, P4, and P5 feature maps are synthesized into a new tensor without loss through splicing, and then channel adjustment and information fusion are performed.

6. The real-time foggy target detection method based on GhostConv feature extraction according to claim 5 is characterized in that: The Decoupled Head in S4 is specifically as follows: After feature fusion, feature information of multiple scales is obtained. Three feature maps with different widths and heights are selected. The three feature maps are combined for target classification and positioning. The decoding process is shown as follows: Bbox=Conv(Conv(Conv2d(Featuremap))) Cls=Conv(Conv(Conv2d(Featuremap))) Among them, Bbox and Cls represent the positioning and classification results of the decoupling head, and Conv2d represents an ordinary two-dimensional convolution.

7. A real-time foggy target detection method based on GhostConv feature extraction according to any one of claims 1-6, characterized in that: The loss function in S5 uses shape and complete IoU overlap loss to calculate the positioning loss, and its formula can be expressed as: Among them, ω w and ω h are the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the real frame; w gt 、h gt Indicates the length and width of the real box.

8. The real-time foggy target detection method based on GhostConv feature extraction according to claim 7 is characterized in that: The classification loss in the loss function in S5 is calculated by binary cross entropy.

9. The real-time foggy target detection method based on GhostConv feature extraction according to claim 8, characterized in that: The real-time foggy target detection model based on GhostConv feature extraction in S5 includes a BackBone backbone network, a Neck network, and a Decoupled Head decoupled detection head.

Images