A low-light target detection method based on double-layer optimization

Abstract

The application discloses a low-light target detection method based on a double-layer optimization theory, which significantly improves the enhancement quality and detection accuracy. The method first adopts a double-layer learning architecture to bridge the enhancement and detection tasks, and organically combines the low-light image enhancement task and the target detection task. Secondly, the enhancement network parameters are taken as the main optimization target, and the detection network parameters are taken as the constraint condition. The enhancement network generates images that are conducive to machine perception. Finally, in order to improve the training efficiency, a method based on the implicit function theory is proposed to approximately solve the corresponding gradient. The training cost is effectively reduced, and the training speed is improved. The application significantly improves the target detection accuracy in weak light scenes, and can be applied to various target detection tasks in weak light scenes, such as night pedestrian recognition, automatic driving, security monitoring and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of low-light object detection, and particularly relates to a low-light object detection method based on double-layer optimization. BACKGROUND

[0002] Low-light object detection is a challenging task in the field of computer vision and has attracted extensive attention. The mainstream low-light object detection method usually cascades an enhancement network and a detection network to narrow the gap between weak-light and normal-light images in the image domain and the feature domain, and applies the existing object detection network to the weak-light scene. However, these methods focus on improving the quality of the features extracted by the detector, and ignore the internal relationship between the weak-light image enhancement task and the object detection task. Since the enhancement network uses the evaluation index of human visual perception, and the detection network uses the evaluation index of machine visual perception, due to the inconsistency of the evaluation indexes, some image enhancement methods may destroy some useful scene information such as edges and details in order to provide visually friendly enhanced images, thereby causing the performance of the detection algorithm to decrease. SUMMARY

[0003] In view of the above problems in the prior art, the present application proposes a low-light object detection method and system based on a double-layer architecture for the object detection task in a low-light environment, aiming to improve both the effect of low-light enhancement and the accuracy of object detection.

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

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

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

[0007] S2, establish a low-light enhancement network based on a DNF network, wherein the DNF network comprises a low-light encoder-decoder unit and a normal-light encoder-decoder unit, and is used for processing the input low-light scene image to improve the image quality and obtain a target image F target , so as to make it closer to the label image;

[0008] S3, establish a target detection network based on a YOLOv5 algorithm, which is used for iteratively training the enhanced target image F target , so that the target detection network can recognize and locate the target in the image, output the target category and the position regression frame, and output the target category and the position regression frame.

[0009] S4. A two-layer learning architecture is used to bridge the enhancement and detection tasks. The low-light enhancement network and the detection network are jointly optimized, iteratively trained, network gradients are calculated, and network parameters are updated to obtain an ideal model of the low-light enhancement network and the detection network, which is used to improve both enhancement and detection performance.

[0010] S5. After collecting low-light images of actual scenes and resizing them 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 the low-light scene image, the normal-light image, the low-light denoised image, and the 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 for the object detection network;

[0013] The label image production includes enhancing the contrast of the low-light scene image based on local histogram equalization; adjusting the color of the normal-light image based on a color space conversion method; and the normal-light image after color adjustment and the low-light scene image after contrast adjustment are used 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 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 the defect feature 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 target, after 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] Wherein, 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 denoising image, represents the low-light denoising image output by the RAW decoder, is the discriminator loss, and λ is the weight of the discriminator loss.

[0021] The target detection network comprises a YOLOv5 network composed of a backbone, a neck and a head, and an identification module.

[0022] In combination with the bounding box loss, the object confidence loss and the class loss, the target detection loss is defined as:

[0023]

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

[0025] The dual-layer learning architecture is used to bridge the enhancement and detection tasks, and the low-light enhancement network and the detection network are jointly optimized, iteratively trained, network gradient is calculated and network parameters are updated, comprising:

[0026] a. The dual-layer optimization method is defined as:

[0027]

[0028] Wherein, represents the loss of the enhancement 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 into the low-light enhancement network, y represents the normal light image input into the target detection network, and θ * represents the optimal target detection network parameters;

[0029] b. Using an implicit function-based approximate gradient solving method to calculate the gradient of the enhanced network parameters

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

[0031] The present application has the following advantages and benefits:

[0032] 1. Enhanced quality improvement: Compared with the most advanced methods, this method achieves more excellent enhancement effect, which is proved 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 effect.

[0033] 2. Detection accuracy improvement: 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 double-layer learning strategy: The proposed double-layer learning framework effectively utilizes the internal relationship between enhancement and detection, thereby 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. By converting the Hessian matrix to first-order derivative multiplication operation while ensuring accuracy, the computational load is reduced.

[0036] 5. Wide application prospect: This method can be applied to various image enhancement and object detection tasks, and has a wide application prospect. This method is universal and can be applied to various image enhancement and object detection tasks, such as image enhancement in low-light conditions, night video monitoring, automatic driving, etc. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is the overall flowchart of the method of the present application;

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

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

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

[0041] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific implementation method of the present application is described in detail below in combination with the drawings. In the following description, a large number of specific details are set forth in order to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of the present application, so the present application is not limited to the specific implementation disclosed below.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0043] The present application is directed to the task of target detection in low light environment, proposes a low light target detection method and system based on double-layer architecture, the purpose is to improve the effect of low light enhancement and improve the accuracy of target detection.

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

[0045] I. Data preprocessing

[0046] Module composition

[0047] The data that needs to be preprocessed includes low light images, normal light images, low light denoising images, and target detection frame images. The low light image is the input of the entire network. The low light image enters the enhancement network first, and then passes through the network denoising part to obtain the denoised low light image, and then passes through the color restoration to obtain the normal light image, and finally passes through the target detection network to obtain the detection result. The normal light images and low light denoising images in the data set are used as the ground truth of the enhancement network, and the frame image is used as the ground truth of the target detection network.

[0048] Specific implementation steps

[0049] 1. First, the low light image, normal light image, low light denoising image, and frame image are subjected to Mosaic data enhancement, and then subjected to adaptive image scaling and cropped 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, so that the image details are clearer.

[0051] 3. For normal-light images, to make the normal-light images more matching with low-light images, a color space conversion (CSC) based method is used to adjust the color of the normal-light images. CSC is a commonly used image processing technique that can convert an image from one color space to another, making the color of the normal-light images closer to the color of the low-light images.

[0052] II. Low-light enhancement network

[0053] Module composition

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

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

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

[0057] where DConv7 is a deep convolution with a 7x7 kernel. MLP is implemented by two point-wise convolution layers and a GELU non-linear function.

[0058] 2. The matrix color correction (MCC) block is used to perform global color enhancement and local refinement, as shown in Figure 2 . 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 C×HW , K∈R C×HW , V∈R C×HW vectors through a 1x1 convolution layer, and then perform a 3x3 convolution. Finally, the transformation matrix M∈R C×C is obtained by matrix multiplication. This step can be represented as:

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

[0060] M = Softmax(Q K T / λ)

[0061] where λ is a stabilizing number, Flatten is a flatten layer operation, DConv3 and PConv represent depthwise convolution and pointwise convolution with 3x3 kernel respectively. Then the color space conversion at feature level is performed by matrix M. The target feature after color conversion can be obtained by F target = M V.

[0062] 3. GFM gating module is used for noise feedback and estimating initial denoised feature. In the gating module, we expect useful information to be adaptively selected and merged along the spatial and channel dimensions. To improve efficiency, pointwise and depthwise convolution are used to aggregate channel and local content information, respectively. Then the mixed feature is split into two blocks along the channel dimension and activated by GELU nonlinear function. The process of GFM gating can be represented as:

[0063] feature vector

[0064]

[0065] where DConv3 and PConv represent depthwise convolution and pointwise convolution with 3x3 kernel respectively, and represents Hadamard product, i.e. multiplication of corresponding elements of matrices. is the feature obtained after preprocessing the low-light image, 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 type denoising network, and four layers of denoising modules are used in the encoder to extract multi-scale features, and skip connections are introduced to fuse the features of the encoder and the decoder. The low-light denoised image is obtained from the output of the decoder of the denoising network, and the loss is calculated with the ground truth of the low-light denoised image.

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

[0069] Loss design

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

[0071]

[0072] where 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, is the low-light denoised image output by the RAW decoder, is the discriminator loss, and λ is the weight of the discriminator loss.

[0073] III. Target detection network

[0074] Module composition

[0075] As shown in Figure 3 , 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 part 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 preserve and extract important features in the image. The feature map extracted by the backbone is passed to the subsequent feature pyramid network and detection head for processing.

[0077] 2. The neck of the network is a network module between the backbone and the detect. Based on the features extracted by the backbone, further feature fusion and upsample operations (Upsample) are performed to provide higher-level semantic information and the ability to adapt to different scale images. Through the neck part, the fusion of shallow graphical features and deep semantic features can be achieved.

[0078] 3. The detection head of the network is responsible for further processing of the extracted features and generating the final output result. The detection head (head) can fuse and convert features of different scales, which helps to capture higher-level semantic information and contextual relationships. The head receives feature maps from layers 17, 20, and 23, with increasingly lower resolution and increasingly larger receptive fields. Although it can learn higher-level semantic information, it will also lose details. 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 frame and class prediction results for the final target detection and prediction stage, which produces the prediction results of the position, class and confidence of the objects in the input image.

[0080] Specific implementation steps

[0081] 1. The output of the enhancement network is taken as the input of the target detection network, which sequentially passes through the backbone network, the neck, the detection head, and the detection part to finally obtain the detection result.

[0082] Loss design

[0083] The total target detection loss is defined in combination with the bounding box loss, object confidence loss and class loss. The total target detection loss is defined as:

[0084]

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

[0086] Four, double-layer optimization architecture

[0087] As shown in Figure 4 , first, a double-layer learning architecture is adopted to bridge the enhancement and detection tasks, and the low-light image enhancement task and the target detection task are organically combined together. Secondly, the enhancement network parameters are taken as the main optimization target, and the detection network parameters are taken as the constraint condition. The enhancement network generates images that are conducive to machine perception. Finally, in order to improve the training efficiency, the present application proposes a method based on the implicit function theory to approximately solve the corresponding gradient. Effectively reduce the training cost, and improve the training speed.

[0088] Definition of the problem

[0089] The double-layer optimization method can be expressed as:

[0090]

[0091] wherein represents the loss of the enhancement 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 into the low-light enhancement network, y represents the normal light image input into the target detection network, and θ * represents the optimal target detection network parameters.

[0092] According to the step three loss formula, the loss of the enhancement network is calculated And the loss of the detection network

[0093] The gradient of the enhancement network parameter is calculated And the gradient of the detection network parameter An approximate gradient solving method based on implicit function is used; the gradient update strategy based on one-step forward model is used to update the enhancement network parameter θ and the detection network parameter φ.

[0094] Solving strategy

[0095] An approximate solving strategy based on implicit function theory. Two second-order Hessian matrices are converted into simple product operations of first-order derivatives. Only a few first-order gradients need to be calculated, and the Hessian matrix and its inverse do not need to be calculated. The specific solving process is as follows:

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

[0097]

[0098] Where 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] For the direct gradient, it can be calculated simply. Solving the indirect gradient G(θ) is a challenging problem because we have to calculate the change rate of the lower problem variable ω * with respect to the upper problem variable ω. Existing methods use different approximation techniques to approximate the calculation of the indirect gradient But the existing method cannot meet the low cost and high precision requirements of solving the model. Therefore, the present application proposes an effective strategy to approximate the indirect gradient

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

[0102]

[0103] The first-order derivative of ω can be obtained Further assume that the lower target θ(ω) is twice differentiable, and based on the implicit function theory, the following can be obtained:

[0104]

[0105] Assume is transposable. Substituting, the indirect gradient G(θ) is:

[0106]

[0107] However, The calculation of involves the calculation of 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, an approximation of the Hessian matrix can be written as:

[0108]

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

[0110]

[0111] In this way, only a few first-order gradients need to be calculated, without the need to calculate the Hessian matrix and its inverse. The network can be trained to approximate the optimal, thus providing a more accurate response.

[0112] Training strategy

[0113] The complete training gradient process of the enhanced network can be summarized as follows: 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 an 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 the 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, avoiding focusing too early on the details of the training data, 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 it reaches the preset target learning rate. 3. Keep 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, update the parameters of the enhanced network once. Approximate the gradient of the detection network, update the parameters of the detection network once.

[0117] A gradient update strategy based on one-step forward model uses a preheating strategy in the early training stage, gradually increases the learning rate, enables the model to learn more general features in the early training stage, avoids paying too much attention to details in the training data, and thus improves 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 higher, and all training data can be effectively utilized. Meanwhile, the strategy includes passing in low-light images and corresponding normal light images, a loss function, 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 the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, and not to limit the present application, any modification, equivalent replacement and improvement within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A low-light target detection method based on double-layer optimization, characterized in that, The method comprises the following steps: S1, collecting low-light scene image, corresponding normal light image, low-light denoising image and anchor frame image, and performing preprocessing and making label image; S2, establish low light enhancement network based on DNF network, the DNF network includes low light codec unit, normal light codec unit; for processing input low light scene image to improve image quality to obtain target image F targrt to make it closer to the label image; S3, a target detection network is established based on a YOLOv5 algorithm, which is used for the enhanced target image F target The data is iteratively trained so that it can recognize and locate the target in the image, output the target category and position regression frame; S4, bridging the enhancement and detection tasks by using a double-layer learning architecture, jointly optimizing, iteratively training, calculating network gradient and updating network parameters of the low-light enhancement network-detection network, obtaining an ideal model of the low-light enhancement network-detection network, and improving the enhancement and detection performance; S5, collecting a low-light image of an actual scene, adjusting the size for preprocessing, inputting the ideal model after training, automatically positioning the target in the image, and outputting the target category and position regression frame.

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 labeled by a person according to different target categories in advance.

3. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that, The preprocessing comprises Mosaic data enhancement operation and image scaling to a uniform size on the low-light scene image, the normal light image, the low-light denoising image and the anchor frame image; the low-light scene image is used as original data to be restored; and the anchor frame image after preprocessing is used as a label of the target detection network. The making of the label image comprises enhancing the contrast of the low-light scene image based on local histogram equalization, and adjusting the color of the normal light image based on color space conversion; the normal light image after adjusting the color and the low-light scene image after adjusting the contrast are used as labels of the 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 encoder-decoder unit comprises a RAW feature extraction module and a RAW feature restoration module. The RAW feature extraction module sequentially encodes and extracts defect features by using the CID denoising module The RAW feature restoration module sequentially decodes and restores the defect feature map after denoising by using the CID denoising module 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 encoder-decoder unit comprises a RAW feature extraction module and a RAW feature restoration module. The RAW feature extraction module comprises a GDM gating module and a CID denoising module; the GDM gating module performs gating processing on the defect feature based on the defect feature and the defect feature map after denoising , and outputs a fusion feature After the CID denoising module further assigns weights to extract a source feature F source ; The RAW feature restoration module utilizes the MCC color restoration module to restore the F source The target image F target is calculated, and is processed step by step, and 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 illumination enhancement loss function is: where y rgb ground truth representing normal light image, y is the normal light image output by the color restoration decoder, raw ground truth representing low light denoised image, y is 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 comprises a YOLOv5 network composed of a backbone, a neck and a head, and a recognition module. The target detection loss is defined by combining the boundary frame loss, the object confidence loss and the category loss: wherein, represents a classification loss, λ cls is a weight for the classification loss, represents a confidence loss, λ obj is a weight for the confidence loss, represents a bounding box loss, λ loc is a weight for the bounding box loss.

8. The low-light target detection method based on double-layer optimization according to claim 1, characterized in that, The joint optimization, iterative training, network gradient calculation and network parameter updating of the low-light enhancement network-detection network by using the double-layer learning architecture to bridge the enhancement and detection tasks comprise: a. defining the double-layer optimization method as: wherein, represents a loss of the enhancement network, represents a loss of the target detection network, ω represents parameters of the low-light enhancement network, θ represents parameters of the target detection network, x represents a low-light image input to the low-light enhancement network, y represents a normal-light image input to the target detection network, θ * represents optimal target detection network parameters; b. Using an implicit function based approximate gradient solving method to calculate the gradient of the enhancement network parameters c. updating the enhancement network parameters θ and the detection network parameters φ by using a gradient update strategy based on a one-step forward model, to obtain an ideal model.

Images