Low-illumination target detection method based on double-layer optimization

Abstract

The invention discloses a low-illumination target detection method based on a double-layer optimization theory. According to the method, the enhancement quality and the detection precision are remarkably improved. According to the method, firstly, a double-layer learning architecture is adopted to bridge enhancement and detection tasks, and a low-illumination image enhancement task and a target detection task are organically combined together; secondly, enhancing network parameters as a main optimization target, and detecting network parameters as constraint conditions; and the enhanced network generates an image which is beneficial to machine perception. And finally, in order to improve the training efficiency, providing a method based on an implicit function theory to approximately solve a corresponding gradient. The training cost is effectively reduced, and the training speed is increased. According to the method, the target detection precision in the weak light scene is remarkably improved, and the method can be applied to target detection tasks in various weak light scenes, such as night pedestrian recognition, automatic driving, security monitoring and the like.

Description

Technical Field

[0001] The present invention relates to the field of low-light object detection, and in particular to a low-light target detection method based on double-layer optimization. Background Art

[0002] Low-light object detection is a challenging task in the field of computer vision and has received widespread attention. Mainstream low-light object detection methods usually cascade the enhancement network and the detection network, and apply the existing object detection network to low-light scenes by narrowing the difference between low-light and normal-light images in the image domain and feature domain. However, these methods focus on improving the quality of feature extraction by the detector, while ignoring the intrinsic connection between the low-light image enhancement task and the object detection task. Since the enhancement network uses the evaluation indicators of human vision and the detection network uses the evaluation indicators of machine vision, due to the inconsistency of the evaluation indicators, some image enhancement methods may destroy some useful scene information, such as edges and details, in order to provide visually friendly enhanced images, resulting in a decrease in the performance of the detection algorithm. Summary of the invention

[0003] In view of the above problems existing in the prior art: The present invention proposes a low-light target detection method and system based on a double-layer architecture for target detection tasks in low-light environments, aiming to improve both the low-light enhancement effect and the accuracy of target detection.

[0004] The present invention adopts the following technical solutions to achieve the above-mentioned purpose:

[0005] A low-light target detection method based on double-layer optimization includes the following steps:

[0006] S1, collect low-light scene images and corresponding normal-light images, low-light denoised images, and anchor frame images, perform preprocessing and generate label images;

[0007] S2. Establish a low-light enhancement network based on the DNF network, wherein the DNF network includes a low-light codec unit and a normal-light codec unit; and is used to process the input low-light scene image to improve the image quality and obtain a target image F. target , making it closer to the label image;

[0008] S3, based on the YOLOv5 algorithm, establish a target detection network for the enhanced target image F target The data is iteratively trained so that it can recognize and locate targets in images, output target categories and position regression boxes;

[0009] S4. Use a two-layer learning architecture to bridge enhancement and detection tasks, jointly optimize the low-light enhancement network-detection network, iteratively train, calculate network gradients and update network parameters, and obtain an ideal model of the low-light enhancement network-detection network to improve enhancement and detection performance;

[0010] S5. After collecting low-light images of actual scenes and adjusting the size for preprocessing, the trained ideal model is input to automatically locate the target in the image and output the target category and position regression box.

[0011] The anchor frame image is a position frame that is manually annotated in advance according to different target categories.

[0012] The preprocessing includes performing a Mosaic data enhancement operation on a low-light scene image, a normal-light image, a low-light denoised image, and an anchor frame image, and scaling the image to a uniform size; the low-light scene image is used as the original data to be restored; and the preprocessed anchor frame image is used as a label of the target detection network;

[0013] The method of making a label image includes enhancing the contrast of a low-light scene image based on local histogram equalization; adjusting the color of a normal-light image based on a color space conversion method; and using the color-adjusted normal-light image and the contrast-adjusted low-light scene image as labels for a subsequent enhancement network.

[0014] The low-light codec unit includes a RAW feature extraction module and a RAW feature restoration module; the RAW feature extraction module uses the CID denoising module to sequentially encode and extract defect features The RAW feature restoration module uses the CID denoising module to decode and restore the denoised defect feature map in sequence The last CID denoising module outputs the restored low-light image

[0015] The normal light codec unit includes a RAW feature extraction module and a RAW feature restoration module;

[0016] The RAW feature extraction module includes the GDM gating module and the CID denoising module; the GDM gating module is based on defect features. And the defect feature map after denoising Perform gate processing and output fusion features After the CID denoising module Further assign weights to extract source features F source ;

[0017] The RAW feature restoration module uses the MCC color restoration module to restore the F source Calculate the target image F targetAfter being processed step by step, the final MCC color restoration module outputs a normal light image

[0018] The low light enhancement loss function is:

[0019]

[0020] Among them, y rgb represents the ground truth of the normal light image, is the normal light image output by the color restoration decoder, y raw represents the ground truth of the low-light denoised image, Represents the low-light denoised image output by the RAW decoder, is the discriminator loss, and λ is the weight of the discriminator loss.

[0021] The target detection network includes a YOLOv5 network consisting of a backbone, a neck, and a detection head, and a recognition module;

[0022] Combining the bounding box loss, object confidence loss, and category loss, we define the object detection loss:

[0023]

[0024] in, represents the classification loss, λ cls is the weight of classification loss, is the confidence loss, λ obj is the weight of confidence loss, is the rectangular box loss, λ loc is the weight of the rectangular box loss.

[0025] The two-layer learning architecture is used to bridge the enhancement and detection tasks, jointly optimize the low-light enhancement network-detection network, iteratively train, calculate the network gradient and update the network parameters, including:

[0026] a. Define the two-level optimization method as:

[0027]

[0028] in, represents the loss of the enhanced network, represents the loss of the target detection network, ω represents the parameters of the low-light enhancement network, θ represents the parameters of the target detection network, x represents the low-light image input to the low-light enhancement network, y represents the normal light image input to the target detection network, θ * Represents the optimal target detection network parameters;

[0029] b. Use the approximate gradient solution method based on implicit function to calculate the gradient of the enhanced network parameters

[0030] c. Use the gradient update strategy based on the one-step forward model to update the enhancement network parameters θ and the detection network parameters φ to obtain the ideal model.

[0031] The present invention has the following beneficial effects and advantages:

[0032] 1. Enhancement quality improvement: Compared with the state-of-the-art methods, this method achieves superior enhancement effects, as evidenced by quantitative indicators such as PSNR, SSIM, and LPIPS. In addition, this method is highly consistent with human vision in visual perception and can produce more natural and realistic enhancement effects.

[0033] 2. Improved detection accuracy: This method significantly improves the object detection performance in low-light scenes, achieving a detection mAP of 37.9% on the LOD dataset.

[0034] 3. Effective Two-layer Learning Strategy: The proposed two-layer learning framework effectively exploits the intrinsic relationship between enhancement and detection, thus achieving better performance in both tasks.

[0035] 4. Efficient gradient calculation: The gradient calculation method based on implicit function theory reduces the computational complexity while maintaining accuracy. While ensuring accuracy, the Hessian matrix is ​​converted into a product operation of first-order derivatives, reducing the amount of calculation.

[0036] 5. Has broad application prospects: This method can be applied to various image enhancement and target detection tasks and has broad application prospects. This method is versatile and can be applied to various image enhancement and target detection tasks, such as image enhancement in low light conditions, night video surveillance, autonomous driving, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the overall flow chart of the method of the present invention;

[0038] Figure 2 It is a schematic diagram of the low-light enhancement network composition;

[0039] Figure 3 It is a schematic diagram of the target detection network composition;

[0040] Figure 4 It is a schematic diagram of the gradient update process of the double-layer architecture network parameters; DETAILED DESCRIPTION

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0043] The present invention aims at the target detection task in a low-light environment and proposes a low-light target detection method and system based on a double-layer architecture, aiming to improve both the low-light enhancement effect and the target detection accuracy.

[0044] The specific process is as follows Figure 1 As shown, it includes: construction of data preprocessing module, construction of low-light image enhancement network, construction of target detection network, modeling of double-layer learning framework, a solution strategy based on implicit function theory, and a gradient solution method based on one-step forward model.

[0045] 1. Data Preprocessing

[0046] Module composition

[0047] The data that needs to be preprocessed include low-light images, normal-light images, low-light denoised images, and target detection outlines. Low-light images are the input of the entire network. When low-light images enter the enhancement network, they first pass through the denoising part of the network to obtain denoised low-light images, then undergo color restoration to obtain normal-light images, and finally pass through the target detection network to obtain detection results. The normal-light images and low-light denoised images in the dataset are used as the ground truth of the enhancement network, and the outlines are used as the ground truth of the target detection network.

[0048] Specific implementation steps

[0049] 1. First, use Mosaic data enhancement on low-light images, normal-light images, low-light denoised images, and frame images, then perform adaptive image scaling and crop to 512*512 size. After the above steps, all data have the same size.

[0050] 2. For low-light images, a method based on local histogram equalization (LHE) is used to enhance the contrast of low-light images, making the image details clearer.

[0051] 3. For the normal light image, in order to make the normal light image more closely match the low light image, a method based on color space conversion (CSC) is used to adjust the color of the normal light image. CSC is a commonly used image processing technology that can convert an image from one color space to another color space to make the color of the normal light image closer to the color of the low light image.

[0052] 2. Low-light Enhanced Network

[0053] Module composition

[0054] The network includes CID denoising module, GDM gating module and MCC color restoration module.

[0055] 1. CID denoising module. Figure 2 As shown in the figure, the RAW encoder E and RAW decoder D are implemented by stacking multiple channel-independent denoising (CID) modules in the low-light image enhancement network. The design of the CID block is based on the following two prior knowledge: 1) There is signal-independent noise in the low-light image in RAW format, and the noise follows a zero-mean distribution; 2) Since the signals of different channels have low inherent correlation in the RAW domain, the distribution of noise between channels tends to be independent. According to the above priors, deep convolution with a larger convolution kernel is introduced in the CID block for denoising. Specifically, for the input feature F in , the output feature F after the channel-independent denoising block out It can be expressed as:

[0056] F out =MLP(DConv7(F in ))+F in

[0057] Where DConv7 is a depthwise convolution with a 7×7 kernel. The MLP is implemented by two point-wise convolutional layers and a GELU nonlinear function.

[0058] 2. Matrix Color Correction (MCC) block to perform global color enhancement and local refinement, e.g. Figure 2 As shown. For the sRGB decoder D, multiple MCC blocks are stacked for color correction. Given the input source feature F source ∈R C×H×W , first generate Q∈R through a 1×1 convolutional layer C×HW , K∈R C×HW ,V∈R C×HW vector, and then perform a 3×3 convolution. Finally, the transformation matrix M∈R is obtained by matrix multiplication C×C . This step can be expressed as:

[0059] Feature vector [Q, K, V] = Flatten(DConv3((PCOnv(Fsource )))

[0060] M=Softmax(Q·K T / λ)

[0061] Where λ is a stable value, Flatten is a flattening layer operation, DConv3 and PConv represent the depthwise convolution and pointwise convolution with a kernel of 3×3, respectively. Then the matrix M is used to perform feature-level color space conversion. target =M·V can obtain the target features after color transformation.

[0062] 3. The GFM gating module is used for noise feedback and estimation of initial denoising features. In the gating module, we expect useful information to be adaptively selected and merged along the spatial and channel dimensions. To improve efficiency, point-wise convolution and depth-wise convolution are used to aggregate channel and local content information, respectively. Then, the mixed features are split into two blocks along the channel dimension and then activated by the GELU nonlinear function. The process of GFM gating can be expressed as:

[0063] Feature Vector

[0064]

[0065] Among them, DConv3 and PConv represent the depth convolution and point-wise convolution with a kernel of 3×3, respectively, and ⊙ represents the Hadamard product, that is, the multiplication of corresponding elements of the matrix. is the feature obtained after the low-light image passes through the preprocessing module. It is the feature output by the CID module in the RAW encoder. is the corresponding fusion feature.

[0066] Specific implementation steps

[0067] 1. The low-light image after data preprocessing first enters the U-Net denoising network. A four-layer denoising module is used in the encoder to extract multi-scale features, and a jump connection is introduced to fuse the features of the encoder and decoder. The low-light denoised image is obtained from the decoder output of the denoising network, and the loss is performed on it compared with the ground truth of the low-light denoised image.

[0068] 2. Input the features extracted from the denoising network into the gating module, and input the denoising feature maps of different layers with different weights into the color restoration encoder for encoding, and then fuse the features of the encoder and decoder through skip connections. The decoder performs color restoration, and the final output is the restored normal light image.

[0069] Loss Design

[0070] In addition to the L1 loss commonly used in the U-Net network, the low-light enhancement loss function also introduces a discriminator to assist in discriminating the visual quality of the enhanced image. The loss function of the low-light enhancement network is as follows:

[0071]

[0072] Among them, y rgb represents the ground truth of the normal light image, is the normal light image output by the color restoration decoder, y raw represents the ground truth of the low-light denoised image, Represents the low-light denoised image output by the RAW decoder, is the discriminator loss, and λ is the weight of the discriminator loss.

[0073] 3. Object Detection Network

[0074] Module composition

[0075] like Figure 3 As shown, it includes a backbone network, a neck, a detection head, and a detection part.

[0076] 1. The backbone of the network is used to extract features. The backbone processes the input image through a series of convolutional layers and pooling layers, gradually reducing the size of the feature map while increasing the number of channels. The purpose of this is to retain and extract important features in the image. The feature map extracted by the backbone will be passed to the subsequent feature pyramid network and detection head for processing.

[0077] 2. The neck of the network is the network module between the backbone and the detect head. Based on the features extracted by the backbone network, feature fusion and upsampling operations are further performed to provide more advanced semantic information and the ability to adapt to images of different scales. Through the neck part, the fusion of shallow graphic features and deep semantic features can be achieved.

[0078] 3. The detection head of the network is responsible for further processing the extracted features and generating the final output. The detection head can fuse and transform features of different scales, which helps capture higher-level semantic information and contextual relationships. The head receives feature maps from the 17th, 20th, and 23rd layers. The resolution of the feature maps is getting lower and lower, and the receptive field is getting larger. Although more advanced semantic information can be learned, details will also be lost. In order to detect objects of different sizes, the detect module is designed to achieve this.

[0079] 4. The detection part (detect) is responsible for decoding the network output and generating the final target detection box and category prediction results for the final target detection and prediction stage. This module generates prediction results for the position, category and confidence of objects in the input image.

[0080] Specific implementation steps

[0081] 1. The output of the enhanced network is used as the input of the target detection network, and passes through the backbone network, neck, detection head, and detection part in turn to finally obtain the detection result.

[0082] Loss Design

[0083] The total object detection loss is defined by combining the bounding box loss, object confidence loss, and category loss. The total object detection loss is defined as:

[0084]

[0085] in, represents the classification loss, λ cls is the weight of classification loss, is the confidence loss, λ obj is the weight of confidence loss, is the rectangular box loss, λ loc is the weight of the rectangular box loss.

[0086] 4. Two-layer Optimized Architecture

[0087] like Figure 4 As shown in the figure, firstly, a two-layer learning architecture is used to bridge the enhancement and detection tasks, and the low-light image enhancement task and the target detection task are organically combined. Secondly, the enhancement network parameters are taken as the main optimization target, and the detection network parameters are taken as constraints. This allows the enhancement network to generate images that are conducive to machine perception. Finally, in order to improve the training efficiency, the present invention proposes a method based on implicit function theory to approximate the corresponding gradient. This effectively reduces the training cost and improves the training speed.

[0088] Problem Definition

[0089] The two-level optimization method can be expressed as:

[0090]

[0091] in represents the loss of the augmented network. represents the loss of the target detection task, ω represents the parameters of the low-light enhancement network, θ represents the parameters of the target detection network, x represents the low-light image input to the low-light enhancement network, y represents the normal light image input to the target detection network, and θ * Represents the optimal target detection network parameters.

[0092] According to the loss formula in step 3, calculate the loss of the enhanced network and detect network loss

[0093] Compute the gradients of the boosted network parameters and the gradient of the detection network parameters An approximate gradient solution method based on implicit function is used; a gradient update strategy based on a one-step forward model is used to update the enhancement network parameters θ and the detection network parameters φ.

[0094] Solution strategy

[0095] An approximate solution strategy based on implicit function theory. Convert two second-order Hessian matrices into a simple product operation of first-order derivatives. Only a few first-order gradients need to be calculated, without calculating the Hessian matrix and its inverse. The specific solution process is as follows:

[0096] According to the chain rule, the gradient of E can be expressed as:

[0097]

[0098] Among them, the first term represents the direct gradient of the enhancement module, and the second term represents the indirect gradient of the enhancement module for the target detection task. The indirect gradient G(θ) can be further expressed as the following two components:

[0099]

[0100] The direct gradient can be obtained by simple calculation. Solving the indirect gradient G(θ) is a challenging problem because we must calculate the optimal solution θ of the underlying problem. * The rate of change relative to the upper-level problem variable ω. Existing methods use different approximation techniques to approximate the indirect gradient However, the existing methods cannot meet the low cost and high accuracy requirements for solving this model. Therefore, this paper proposes an effective strategy to approximate the indirect gradient

[0101] First, use the implicit function theorem to calculate the indirect gradient θ * is a solution to the subproblem, so it satisfies:

[0102]

[0103] We can take the first-order derivative of ω to get Further assuming that the lower layer target θ(ω) is second-order differentiable, based on implicit function theory, we can get:

[0104]

[0105] Assumptions is transposable. Substituting in, we can get the indirect gradient G(θ):

[0106]

[0107] However, The calculation involves computing the Hessian matrix and the inverse of the Hessian matrix In practical applications, it is quite challenging to directly calculate and store the Hessian matrix. Inspired by the Gauss-Newton method, the Hessian matrix can be approximated as:

[0108]

[0109] This method transforms two second-order Hessian matrices into a simple product operation of first-order derivatives, greatly improving the computational efficiency. After the above derivation, the indirect gradient G(θ) can be approximated by the following formula:

[0110]

[0111] In this way, only a few first-order gradients need to be calculated, instead of the Hessian matrix and its inverse. The network can be trained to approach the optimum, thus providing a more accurate response.

[0112] Training strategy

[0113] The complete enhanced network training gradient process can be summarized as the following steps: Initialize the parameters of the enhanced network. Randomly select a sample from the training set, including the original image and the target image. Input the original image into the enhanced network to generate the enhanced image. Calculate the loss function value between the enhanced image and the target image. Calculate the gradient of the enhanced network. Update the parameters of the enhanced network.

[0114] step:

[0115] 1. Use a warm-up strategy at the beginning of training to gradually increase the learning rate so that the model can learn more general features at the beginning of training and avoid focusing on the details in the training data too early, thereby improving the generalization ability of the model. Specifically, the warm-up strategy can be divided into the following steps: 1. Set the initial learning rate. Choose a small initial learning rate, usually a small fraction of the preset target learning rate. 2. Linearly increase the learning rate. As training progresses, gradually increase the learning rate until the preset target learning rate is reached. 3. Maintain the target learning rate. After reaching the target learning rate, keep the learning rate unchanged and continue training the model.

[0116] 2. Calculate the gradient of the enhanced network and update the parameters of the enhanced network. Approximately calculate the gradient of the detection network and update the parameters of the detection network.

[0117] A gradient update strategy based on a one-step forward model uses a warm-up strategy at the beginning of training to gradually increase the learning rate, so that the model can learn more general features at the beginning of training, avoid paying too much attention to the details in the training data, and thus improve the generalization ability of the model; when updating the parameters of the enhancement network and the target detection network, only one step is updated, the training efficiency is high, and all training data can be effectively utilized. At the same time, the strategy includes passing in low-light images and corresponding normal-light images as well as loss functions, calculating the gradient of the enhancement network, updating the parameters of the enhancement network once, approximately calculating the gradient of the detection network, and updating the parameters of the detection network once.

[0118] The above is a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A low-light target detection method based on double-layer optimization, characterized in that: The following steps are involved: S1, collect low-light scene images and corresponding normal-light images, low-light denoised images, and anchor frame images, perform preprocessing and generate label images; S2. Establish a low-light enhancement network based on the DNF network, wherein the DNF network includes a low-light codec unit and a normal-light codec unit; and is used to process the input low-light scene image to improve the image quality and obtain a target image F. targrt , making it closer to the label image; S3, based on the YOLOv5 algorithm, establish a target detection network for the enhanced target image F target The data is iteratively trained so that it can recognize and locate targets in images, output target categories and position regression boxes; S4. Use a two-layer learning architecture to bridge enhancement and detection tasks, jointly optimize the low-light enhancement network-detection network, iteratively train, calculate network gradients and update network parameters, and obtain an ideal model of the low-light enhancement network-detection network to improve enhancement and detection performance; S5. After collecting low-light images of actual scenes and adjusting the size for preprocessing, the trained ideal model is input to automatically locate the target in the image and output the target category and position regression box.

2. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The anchor frame image is a position frame that is manually annotated in advance according to different target categories.

3. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The preprocessing includes performing a Mosaic data enhancement operation on a low-light scene image, a normal-light image, a low-light denoised image, and an anchor frame image, and scaling the image to a uniform size; the low-light scene image is used as the original data to be restored; and the preprocessed anchor frame image is used as a label of the target detection network; The method of making a label image includes enhancing the contrast of a low-light scene image based on local histogram equalization; adjusting the color of a normal-light image based on a color space conversion method; and using the color-adjusted normal-light image and the contrast-adjusted low-light scene image as labels for a subsequent enhancement network.

4. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The low-light codec unit includes a RAW feature extraction module and a RAW feature restoration module; The RAW feature extraction module uses the CID denoising module to encode and extract defect features in sequence The RAW feature restoration module uses the CID denoising module to decode and restore the denoised defect feature map in sequence The last CID denoising module outputs the restored low-light image 5. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The normal light codec unit includes a RAW feature extraction module and a RAW feature restoration module; The RAW feature extraction module includes the GDM gating module and the CID denoising module; the GDM gating module is based on defect features. And the defect feature map after denoising Perform gate processing and output fusion features After the CID denoising module Further assign weights to extract source features F source ; The RAW feature restoration module uses the MCC color restoration module to restore the F source Calculate the target image F target After being processed step by step, the final MCC color restoration module outputs a normal light image 6. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The low light enhancement loss function is: Among them, y rgb represents the ground truth of the normal light image, is the normal light image output by the color restoration decoder, y raw represents the ground truth of the low-light denoised image, Represents the low-light denoised image output by the RAW decoder, is the discriminator loss, and λ is the weight of the discriminator loss.

7. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The target detection network includes a YOLOv5 network consisting of a backbone, a neck, and a detection head, and a recognition module; Combining the bounding box loss, object confidence loss, and category loss, we define the object detection loss: in, represents the classification loss, λ cls is the weight of classification loss, is the confidence loss, λ obj is the weight of confidence loss, is the rectangular box loss, λ loc is the weight of the rectangular box loss.

8. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that: The two-layer learning architecture is used to bridge the enhancement and detection tasks, jointly optimize the low-light enhancement network-detection network, iteratively train, calculate the network gradient and update the network parameters, including: a. Define the two-level optimization method as: in, represents the loss of the enhanced network, represents the loss of the target detection network, ω represents the parameters of the low-light enhancement network, θ represents the parameters of the target detection network, x represents the low-light image input to the low-light enhancement network, y represents the normal light image input to the target detection network, θ * Represents the optimal target detection network parameters; b. Use the approximate gradient solution method based on implicit function to calculate the gradient of the enhanced network parameters c. Use the gradient update strategy based on the one-step forward model to update the enhancement network parameters θ and the detection network parameters φ to obtain the ideal model.

Images