Target detection task-driven image defogging method

Abstract

The invention relates to the technical field of image processing, solves the technical problem that a traditional image defogging method is poor in performance in a high-quality restoration application scene, and particularly relates to a target detection task-driven image defogging method, which comprises the following steps of: inputting a foggy day image into a pre-trained target detection network to obtain a preliminary target detection prediction result; generating a detection guide image based on the target detection prediction result; constructing an image defogging network, and respectively inputting the detection guide image and the foggy day image into the image defogging network to obtain defogged images; introducing a high-order task loss function which deeply combines a target detection task and a defogging task into the target detection network; and inputting the defogged image into the target detection network to generate a final target detection result. The method can achieve the high-quality restoration of the image under a severe weather condition, and remarkably improves the accuracy of a target detection task.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular, to an image dehazing method driven by a target detection task. Background Art

[0002] Images collected by outdoor vision systems are often severely troubled by the degradation of contrast and visibility in extreme weather conditions such as fog, heavy rain, and sandstorms. This kind of degradation erodes the image quality in multiple dimensions, weakens the contrast and clarity, distorts the colors, erases details, and even buries key information, bringing great obstacles to subsequent feature extraction, analysis, and understanding. In target detection tasks, degraded images make target features blurred and difficult to distinguish, severely impacting the accuracy and robustness of detection algorithms, resulting in frequent missed detections and false detections. Taking military target detection as an example, battlefield smoke or rain curtains will significantly reduce the image clarity and block targets in the atmosphere, causing a sharp decline in detection accuracy. This data distortion is fatal in real applications such as military. Therefore, extracting effective and robust visual features in extreme weather is crucial for subsequent detection tasks.

[0003] The academic and industrial communities have paid attention to visual feature representation and target detection technologies in extreme interference scenarios, especially image restoration and target detection problems in adverse weather conditions. Current research can be roughly divided into three categories: one is to directly train and test based on foggy images; the second is to preprocess foggy images with existing dehazing algorithms and then detect; the third is to explore a unified paradigm to connect low-order image dehazing with high-order target detection tasks. For example, DSNet, TogetherNet, and DH-YOLO learn clean and clear features by adding a feature restoration branch, and the detection backbone network shares weights with the feature extraction layer of this branch to improve the detection effect. However, most existing methods have obvious limitations. Some methods only focus on the dehazing task and ignore the improvement of detection accuracy by the dehazing results; others overemphasize detection performance and ignore the quality of the restored image, performing poorly in application scenarios that require high-quality restoration results. How to achieve an accurate joint optimization mechanism between low-order and high-order tasks is still a complex and unsolved problem. Therefore, designing a unified optimization paradigm has become an extremely important research direction in this field. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an image dehazing method driven by a target detection task, which solves the technical problem that traditional image dehazing methods perform poorly in high-quality restoration application scenarios.

[0005] To solve the above technical problems, the present invention provides the following technical solution: An image dehazing method driven by a target detection task, the method comprising the following steps:

[0006] S1. Input the foggy image into the pre-trained object detection network to obtain the preliminary object detection prediction results;

[0007] S2. Generate a detection guidance image based on the object detection prediction results, where the pixel values of the target regions are set to the class ID + 1, and the background regions are set to 0;

[0008] S3. Construct an image dehazing network, and input the detection guidance image and the foggy image into the image dehazing network respectively to obtain the dehazed image;

[0009] The image dehazing network uses a U-shaped architecture with skip connections as the backbone network, including a downsampling module for extracting fused features from the foggy image and the detection guidance image, and an upsampling module for gradually restoring the details and features in the foggy image and generating the dehazed image, where multiple layers are set for the downsampling module and the upsampling module;

[0010] S4. Introduce a high-order task loss function that deeply combines the object detection task and the dehazing task in the object detection network;

[0011] S5. Input the dehazed image into the object detection network to generate the final object detection results.

[0012] Further, in step S1, the specific process includes:

[0013] S11. Select YOLOv7-tiny as the detector for the object detection task for pre-training, and freeze the parameters of the detector during the process of updating the image dehazing network;

[0014] S12. Input the foggy image into the above-frozen detector to obtain the preliminary object detection prediction results.

[0015] Further, each layer of the downsampling module and the upsampling module integrates a feature enhancement module with physical perception;

[0016] The feature enhancement module consists of feature extractors in two stages. In each stage, the input features will pass through a batch normalization layer for normalization to improve the generalization ability, and finally the output features and the input features are obtained through a residual connection to get the dehazing features.

[0017] Further, in the multi-layer downsampling module and upsampling module, a guidance fusion module for extracting shallow information of the original features at the coarse-grained level is introduced in the downsampling module located in the second layer; a guidance attention module for extracting deep information of the original features at the fine-grained level is introduced in the upsampling module located in the second layer.

[0018] Further, in step S3, the process of obtaining the dehazed image includes:

[0019] S31. Input the foggy image into a 3×3 convolutional layer for feature extraction, and sequentially pass through multiple downsampling modules to extract features from the foggy image and the detection guidance image, and output the fused features.

[0020] S32. Input the features extracted by the downsampling module into a symmetric upsampling module, and use the feature enhancement module to gradually restore the details and features of the foggy image to obtain the defogged image.

[0021] S33. Send the defogged image output by the upsampling module into a 3×3 convolutional layer to generate the restored image after defogging.

[0022] Further, the two stages of the feature enhancement module include:

[0023] The first stage includes:

[0024] Input the normalized features into a pointwise convolutional layer and a depth convolutional layer to extract feature x.

[0025] Input the normalized features into a pointwise convolutional layer and a Sigmoid function as the gating signal for feature x.

[0026] Use a pointwise convolutional layer to apply the gating signal to the adjusted feature representation of feature x.

[0027] The second stage includes:

[0028] Input the output features of the first stage into the channel attention module to obtain the attention-weighted feature representation.

[0029] Input the attention-weighted features into the physical perception defogging module. The physical perception defogging module introduces the atmospheric scattering model and is divided into two branches to respectively predict the parameters of the model, namely the global atmospheric light A and the reciprocal of the transmittance.

[0030] Branch one uses global average pooling to accelerate the calculation and reduce redundant information in the feature space. Then, it passes through two pointwise convolutional layers to increase the number of channels, and uses GELU as the activation function during this process to obtain the predicted parameters.

[0031] Branch two uses a pointwise convolutional layer to increase the number of channels, then inputs it into a 3×3 convolutional layer to extract deep features and uses GELU as the activation function to obtain the predicted parameters.

[0032] Based on the predicted parameters and According to the formula Calculate the interpretable defogging features where I hazy(x) is the atomization feature representation at feature x in the feature space; is the transmittance estimate corresponding to feature x in the feature space.

[0033] Furthermore, a guided fusion module is introduced into the downsampling module of the second layer, and the processing process includes:

[0034] Extract the guiding features g of the detection guiding image through point-by-point convolutional layers and depth-wise convolutional layers;

[0035] Perform upsampling to keep the size consistency between the original feature f and the guided feature g;

[0036] The dehazing features output by the guide feature g and the feature enhancement module Add together to get the fusion features;

[0037] After the fusion features are extracted by point-by-point convolution and depth-wise convolution, the weights are adjusted through the Sigmoid function, and the original feature f representation is retained through the residual connection, thereby extracting the shallow information of the original feature f at a coarse-grained level.

[0038] Furthermore, a guided attention module is introduced into the upsampling module of the second layer, and the processing process includes:

[0039] Apply point-by-point convolutional layers to extract features on the original feature f and the guided feature g respectively and add and fuse the features;

[0040] The attention score is calculated using two ReLU activation functions and point-by-point convolutional layers;

[0041] The original feature f is multiplied by the attention score to enhance the feature expression of the key area and used as the final output feature, thereby extracting the deep information of the original feature f at a fine level.

[0042] Furthermore, in step S4, it includes:

[0043] Use MAE loss as image restoration loss L res To optimize the predicted dehazed image, the expression of image restoration loss is:

[0044]

[0045] Where N is the total number of training samples; J and Represent the clear image and the predicted restored image respectively;

[0046] Use the detection loss of the original target detection network as the target detection loss L res , the expression of target detection loss is:

[0047] L det =λbox L box + λ obj L obj + λ cls L cls

[0048] Wherein, L box 、L obj and L cls represent the positioning loss, confidence loss, and classification loss respectively; λ box 、λ obj and λ cls represent the corresponding weight coefficients;

[0049] The haze removal task and the object detection task are weighted by the weight hyperparameter λ to promote the update of the haze removal process in a direction more conducive to detection. The expression of the high-order task loss function is:

[0050] L total = L res + λL det

[0051] Wherein, L res is the object detection loss; L res is the image restoration loss.

[0052] With the above technical solutions, the present invention provides an image haze removal method driven by an object detection task, which at least has the following beneficial effects:

[0053] 1. The present invention realizes the high-quality restoration of images under bad weather conditions and significantly improves the accuracy of the object detection task through the guidance fusion module, guidance attention module, physical perception feature enhancement module, and training strategy guided by the high-order task loss.

[0054] 2. The present invention inputs the detection-guided images into the guidance fusion module and the guidance attention module of the image haze removal network respectively, guides the haze removal focus to the area containing the object of interest from two levels of roughness and fineness, and then improves the high-order scene understanding ability of the haze removal model to generate high-quality haze removal features.

[0055] 3. The present invention introduces the high-order task loss function of the object detection network, deeply combines the object detection task with the haze removal task, and drives the optimization of the feature extraction of the image haze removal network by the object detection task, so that the haze removal process is updated in a direction conducive to object detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0057] Figure 1 This is the flowchart of the image defogging method in the present invention;

[0058] Figure 2 This is the flowchart of obtaining the target detection result by using the image defogging network in the present invention;

[0059] Figure 3 This is the network structure diagram of the image defogging network of the present invention;

[0060] Figure 4 This is the network structure diagram of the guiding fusion module and the guiding attention module in the present invention. Detailed implementation manners

[0061] To make the above 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 implementation manners. Thereby, the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0062] This embodiment proposes an image defogging method driven by a target detection task. Through a guiding fusion module, a guiding attention module, a physical perception feature enhancement module, and a training strategy guided by a high-order task loss, high-quality restoration of images under adverse weather conditions is achieved, and at the same time, the accuracy of the target detection task is significantly improved. As Figure 1 shown, the method includes the following steps:

[0063] S1. Input the foggy image into the pre-trained target detection network to obtain a preliminary target detection prediction result. The specific process includes:

[0064] S11. Select YOLOv7-tiny as the detector for the target detection task for pre-training, and freeze the parameters of the detector during the process of updating the image defogging network;

[0065] S12. Input the foggy image into the above frozen detector to obtain a preliminary target detection prediction result.

[0066] S2. Generate a detection guiding image similar to a mask image based on the target detection prediction result, where the pixel value of the target area is set to the class ID + 1, and the background area is set to 0.

[0067] S3. Construct an image defogging network, and input the detection guiding image and the foggy image into the image defogging network respectively to obtain a defogged image, as Figure 2 shown. Specifically, by inputting the foggy image into a defogging network integrated with a physical perception feature enhancement module (PFEB), a physical prior is introduced in the feature space. As Figure 3As shown in the figure, the image dehazing network with a U-shaped architecture having skip connections includes a multi-layer downsampling module (encoder) and an upsampling module (decoder), and each layer integrates a physically-aware feature enhancement module (PFEB). Specifically:

[0068] The image dehazing network uses a U-shaped architecture with skip connections as the backbone network, including a downsampling module and an upsampling module composed of multiple layers stacked. Each layer of the downsampling module and the upsampling module integrates a physically-aware feature enhancement module (PFEB). Among them, the downsampling module is used to extract features from the foggy image and the detection guidance image and output fused features, and the upsampling module is used to gradually restore the details and features in the foggy image to generate a dehazed image. In this embodiment, the process of obtaining the dehazed image includes:

[0069] S31: The foggy image is input into a 3×3 convolutional layer for feature extraction, and the foggy image and the detection guidance image are successively passed through a multi-layer downsampling module for feature extraction to output fused features;

[0070] S32: The features extracted by the downsampling module are input into a symmetric upsampling module, and the physically-aware feature enhancement module (PFEB) is used to gradually restore the details and features of the foggy image to obtain a dehazed image;

[0071] S33: The dehazed image output by the upsampling module is sent into a 3×3 convolutional layer to further refine the features and generate a finally dehazed restored image.

[0072] As a further innovative solution, the feature enhancement module (PFEB) in this embodiment consists of two-stage feature extractors. In each stage, the input features will pass through a batch normalization layer for normalization to improve the generalization ability, and finally the output features and the input features are obtained through a residual connection to obtain dehazing features Specifically, stage one includes:

[0073] The normalized features are input into a pointwise convolutional layer and a depth convolutional layer for extracting feature x; the normalized features are input into a pointwise convolutional layer and a Sigmoid function as the gating signal for feature x; a pointwise convolutional layer is used to apply the gating signal to adjust the feature representation of feature x.

[0074] Stage two includes:

[0075] The output features of stage one are input into a channel attention module (CA) to obtain an attention-weighted feature representation; the attention-weighted features are input into a physically-aware dehazing module (PFEB), and the physically-aware dehazing module (PFEB) introduces an atmospheric scattering model and is divided into two branches to respectively predict the parameters of the model, namely the global atmospheric light A and the reciprocal of the transmittance

[0076] Branch 1 uses global average pooling to accelerate calculations and reduce redundant information in the feature space. Then, two layers of pointwise convolutional layers are used to increase the number of channels. During this process, GELU is used as the activation function to obtain the prediction parameters.

[0077] To prevent the large amount of information loss caused by global average pooling in Branch 2, the number of channels is first increased using a pointwise convolutional layer, and then the input is fed into a 3×3 convolutional layer to extract deep features and GELU is used as the activation function to obtain the prediction parameters.

[0078] Based on the prediction parameters and According to the formula the interpretable defogging features are calculated. where I hazy (x) is the foggy feature representation at feature x in the feature space; is the transmittance estimate corresponding to feature x in the feature space.

[0079] In the multi-layer downsampling module and upsampling module, a Guided Fusion Block (GFB) for extracting shallow information of the original features at the coarse-grained level is introduced in the second-layer downsampling module; a Guided Attention Block (GAB) for extracting deep information of the original features at the fine-grained level is introduced in the second-layer upsampling module. Specifically, a Guided Fusion Block (GFB) is introduced in the second-layer downsampling module, as Figure 4 shown. The processing process includes:

[0080] The guiding features g of the detection guiding image are extracted through a pointwise convolutional layer and a depth convolutional layer; an upsampling operation is performed to keep the size consistency of the original features f and the guiding features g; the guiding features g and the defogging features output by the Feature Enhancement Block (PFEB) are added to obtain the fused features; after the fused features are extracted through pointwise convolution and depth convolution, the weights are adjusted through the Sigmoid function, and the original features f are retained through residual connection, so as to extract the shallow information of the original features f at the coarse-grained level.

[0081] Furthermore, a Guided Attention Block (GAB) is introduced in the second-layer upsampling module, as Figure 4 shown. The processing process includes:

[0082] Pointwise convolutional layers are respectively applied to the original features f and the guiding features g to extract features and add the features for fusion; two ReLU activation functions and a pointwise convolutional layer are used to calculate the attention scores; the original features f are multiplied by the attention scores to enhance the feature expression of the key regions and serve as the final output features, so as to extract the deep information of the original features f at the fine-grained level.

[0083] In this embodiment, the detection guidance image is input into the guidance fusion module (GFB) and the guidance attention module (GAB) of the image dehazing network respectively, and the dehazing focus is guided to the area containing the object of interest from two levels of roughness and fineness, so as to improve the high-order scene understanding ability of the dehazing model and generate high-quality dehazing features.

[0084] S4. Introduce a high-order task loss function that deeply combines the object detection task and the dehazing task in the object detection network. In this embodiment, by introducing the high-order task loss function of the object detection network, the object detection task and the dehazing task are deeply combined, and the object detection task is used to drive the optimization of the feature extraction of the image dehazing network, so that it is updated in the direction beneficial to object detection, including:

[0085] Use the MAE loss as the image restoration loss L res to optimize the predicted dehazed image. The expression of the image restoration loss is:

[0086]

[0087] where N is the total number of training samples; J and represent the clear image and the predicted restored image respectively;

[0088] Use the detection loss of the original object detection network (i.e., YOLOv7-tiny) as the object detection loss L res , and the expression of the object detection loss is:

[0089] L det = λ box L box + λ obj L obj + λ cls L cls

[0090] where L box , L obj and L cls represent the localization loss, the confidence loss and the classification loss respectively; λ box , λ obj and λ cls represent the corresponding weight coefficients;

[0091] Weigh the dehazing task and the object detection task through the weight hyperparameter λ, and promote the dehazing process to be updated in the direction more beneficial to detection. According to the ablation experiment on the loss weight, the best result can be obtained when λ is set to 0.4. The expression of the high-order task loss function is:

[0092] L total = L res + λLdet

[0093] In the formula, L res is the target detection loss; L res is the image restoration loss.

[0094] S5. Input the dehazed image into the target detection network to generate the final target detection result. Specifically, input the dehazed image into the pre-trained target detection model YOLOv7-tiny with frozen parameters to generate the final target detection result.

[0095] In the present invention, by respectively inputting the detection guidance image into the guidance fusion module and the guidance attention module of the image dehazing network, the dehazing focus is guided to the area containing the object of interest from two levels of rough and fine, thereby improving the high-order scene understanding ability of the dehazing model to generate high-quality dehazing features.

[0096] In the present invention, by introducing the high-order task loss function of the target detection network, the target detection task and the dehazing task are deeply combined, and the feature extraction of the image dehazing network is optimized driven by the target detection task, so that the dehazing process is updated in the direction beneficial to target detection.

[0097] Those of ordinary skill in the art can understand that all or part of the steps in implementing the above method embodiments can be completed by instructing relevant hardware through a program. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the above embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0099] The above embodiments have introduced the present invention in detail. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An image dehazing method driven by an object detection task, characterized in that, The method includes the following steps: S1. Input the foggy image into the pre-trained object detection network to obtain the preliminary object detection prediction results; S2. Generate a detection guidance image based on the object detection prediction results, where the pixel values of the target regions are set to the class ID + 1, and the background regions are set to 0; S3. Construct an image dehazing network, and input the detection guidance image and the foggy image into the image dehazing network respectively to obtain the dehazed image; The image dehazing network uses a U-shaped architecture with skip connections as the backbone network, including a downsampling module for extracting fused features from the foggy image and the detection guidance image, and an upsampling module for gradually restoring the details and features in the foggy image and generating the dehazed image. The downsampling module and the upsampling module are provided with multiple layers; S4. Introduce a high-order task loss function that deeply combines the object detection task and the dehazing task in the object detection network; S5. Input the dehazed image into the object detection network to generate the final object detection results.

2. The image defogging method according to claim 1, wherein In step S1, the specific process includes: S11. Select YOLOv7-tiny as the detector for the object detection task for pre-training, and freeze the parameters of the detector during the process of updating the image dehazing network; S12. Input the foggy image into the above-frozen detector to obtain the preliminary object detection prediction results.

3. The image defogging method according to claim 1, characterized in that, Each layer of the downsampling module and the upsampling module integrates a feature enhancement module with physical perception; The feature enhancement module consists of two-stage feature extractors. In each stage, the input features will go through a batch normalization layer for normalization to improve the generalization ability, and finally the output features and the input features are obtained through a residual connection to get the dehazing features.

4. The image defogging method according to claim 3, wherein In the multiple-layer downsampling module and upsampling module, a guidance fusion module for extracting shallow information of the original features at the coarse-grained level is introduced in the second-layer downsampling module; a guidance attention module for extracting deep information of the original features at the fine-grained level is introduced in the second-layer upsampling module.

5. The image defogging method according to claim 3, wherein In step S3, the process of obtaining the dehazed image includes: S31. Input the foggy image into a 3*3 convolutional layer for feature extraction, and sequentially pass through the multiple-layer downsampling module to extract features from the foggy image and the detection guidance image and output the fused features; S32. Input the features extracted by the downsampling module into the symmetric upsampling module, and use the feature enhancement module to gradually restore the details and features of the foggy image to obtain the dehazed image; S33. Send the dehazed image output by the upsampling module into a 3*3 convolutional layer to generate the restored image after dehazing.

6. The image defogging method according to claim 3, wherein The two stages of the feature enhancement module include: The first stage includes: Input the normalized features into a pointwise convolutional layer and a depth convolutional layer to extract the feature x; Input the normalized features into a pointwise convolutional layer and a Sigmoid function as the gating signal for the feature x; Use a pointwise convolutional layer to apply the gating signal to the adjusted feature representation of the feature x; The second stage includes: Input the output features of the first stage into the channel attention module to obtain the attention-weighted feature representation; The features weighted by attention are input into the physically-aware defogging module, which introduces the atmospheric scattering model and is divided into two branches to predict the parameters of the model, namely the global atmospheric light A and the reciprocal of the transmittance Branch 1 uses global average pooling to accelerate calculations and reduce redundant information in the feature space. After that, two layers of pointwise convolutional layers are used to increase the number of channels, and GELU is used as the activation function during this process to obtain the prediction parameters Branch two uses a pointwise convolutional layer to increase the number of channels, then inputs a 3*3 convolutional layer to extract deep features and uses GELU as the activation function to obtain the prediction parameters Based on prediction parameters and According to the formula Calculate the interpretable defogging feature Where I hazy (x) is the fogging feature representation at feature x in the feature space; Is the transmittance estimate corresponding to feature x in the feature space.

7. The image defogging method according to claim 4, wherein Introduce a guiding fusion module in the downsampling module of the second layer. The processing process includes: Extract the guiding feature g of the detection guiding image through a pointwise convolutional layer and a depth convolutional layer; Perform an upsampling operation to maintain the dimensional consistency of the original feature f and the guiding feature g; Add the guiding feature g and the defogging feature output by the feature enhancement module to obtain a fused feature; After the fused feature is extracted through pointwise convolution and depth convolution, the weight is adjusted by the Sigmoid function, and the original feature f representation is retained through a residual connection, so as to extract the shallow information of the original feature f at the coarse-grained level.

8. The image defogging method according to claim 4, wherein Introduce a guiding attention module in the upsampling module of the second layer. The processing process includes: Apply a pointwise convolutional layer to the original feature f and the guiding feature g respectively to extract features and add the features for fusion; Calculate the attention score using two ReLU activation functions and a pointwise convolutional layer; Multiply the original feature f by the attention score to enhance the feature expression of the key region and use it as the final output feature, so as to extract the deep information of the original feature f at the fine-grained level.

9. The image defogging method according to claim 1, characterized in that In step S4, it includes: Using the MAE loss as the image restoration loss L res to optimize the predicted dehazed image, the expression of the image restoration loss is as follows: where N is the total number of training samples; J and represent the clear image and the predicted restored image, respectively; Use the detection loss of the original object detection network as the object detection loss L res , and the expression of the object detection loss is: L det = λ box L box + λ obj L obj + λ cls L cls where, L box , L obj and L cls represent the localization loss, confidence loss, and classification loss respectively; λ box , λ obj and λ cls represent the corresponding weight coefficients; Weigh the dehazing task and the object detection task through the weight hyperparameter λ to promote the update of the dehazing process in a direction more conducive to detection. The expression of the high-order task loss function is: L total = L res + λL det Where L res is the object detection loss; L res is the image restoration loss.

Images