A low-light image defogging method based on a lightweight deep neural network

Abstract

The application discloses a low-light image defogging method based on a lightweight deep neural network, mainly solves the problems of difficult fog image feature extraction, lack of effective correction of image color cast, and complex defogging network structure and more resource occupation in a low-light scene. The method comprises the following steps: constructing a training and test dataset containing synthetic low-light fog images and real low-light fog images; constructing an end-to-end lightweight deep neural network for low-light image defogging, including a lightweight multi-level feature fusion submodule and a lightweight channel attention submodule; constructing a network target loss function; training the network using the constructed dataset; inputting the foggy image in the low-light scene into the trained network to obtain a defogged image. The application has the advantages of better correction of image color cast, recovery of image details, less resource occupation, small parameter quantity and low computational complexity while maintaining the contrast of the restored image.

Description

Technical Field

[0001] This invention relates to an image dehazing method, and more particularly to a low-light image dehazing method based on a lightweight deep neural network. Background Technology

[0002] Image dehazing is a crucial technology in surveillance and security, as well as image enhancement. In foggy or hazy weather conditions, visibility is severely reduced, leading to significant degradation in the quality of images acquired by visual systems. This is especially true in low-light, foggy or hazy scenes such as early morning, evening, and night, where the impact of fog on image quality is amplified, resulting in lower image contrast, blurred details, and impaired performance for advanced visual tasks such as target detection, tracking, and recognition. Therefore, image dehazing in low-light, foggy or hazy scenes has significant importance and practical value.

[0003] Existing low-light dehazing algorithms can be broadly categorized into two types: traditional methods and deep learning methods. Currently, traditional algorithms are more prevalent. For example, Jiang et al. proposed a low-light image dehazing method based on guided image filtering theory, which effectively reduces the interference of various artificial light sources in nighttime fog scenes (Jiang B, Meng H, Ma X, et al. Nighttime image Dehazing with modified models of color transfer and guided image filter[J]. Multimedia tools and applications, 2018, 77(3):3125-3141). Yang et al. designed a new illumination model, introducing a structure-preserving optimization flow based on retinal theory to obtain ambient lighting, and proposed a variation-based robust nighttime image dehazing flow to better address the non-global atmospheric light problem in nighttime scenes (C.-H. Yang, Y.-H. Lin and Y.-C. Lu, "A Variation-Based Nighttime Image Dehazing Flow With a Physically Valid Illumination Estimator and a Luminance-Guided Coloring Model," in IEEE). Access, vol.10, pp.50153-50166, 2022, doi:10.1109 / ACCESS.2022.3173330. However, with the development of deep learning, numerous experiments have shown that in the field of image restoration, deep learning methods have stronger generalization ability and enhancement effects than traditional methods, and are more easily integrated with high-level vision tasks (object detection, recognition, segmentation).Typical dehazing algorithms under normal lighting conditions generally achieve excellent processing results. For example, Dong et al. proposed a multi-scale boosted dehazing network that combines enhancement strategies and back-projection techniques for image dehazing (DONG H, PAN J, XIANG L, et al. Multi-scale boosted dehazing network with dense feature fusion[C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2020:2157-2167); Chen et al. proposed a gated information aggregation network to remove fog from images and used smooth dilation techniques to repair and restore the grid effect in the image (CHEN D, HE M, FANQ, et al. Gated context aggregation network for image dehazing and deraining[C] / / 2019IEEE winter conference on applications of computer vision (WACV). IEEE, 2019:1375-1383); Qin et al. proposed a feature fusion network based on hybrid attention (QIN X, WANG Z, BAIY, et al. FFA-Net: Feature fusion attention network for single imagedehazing[C] / / Proceedings of the AAAI Conference on Artificial Intelligence.2020,34(07):11908-11915. These algorithms perform well in normal lighting dehazing tasks, but they still have shortcomings, mainly in the following two aspects: (1) In nighttime non-single light source scenarios, such as the common yellow streetlights at night, existing methods lack effective correction for image color shift, and the image is prone to severe color shift; (2) The models often consume a lot of memory and have a slow processing speed, which is often unfavorable for application in actual deployment due to limited computing resources. Summary of the Invention

[0004] Purpose of the invention: The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a method for dehazing low-light images based on a lightweight deep neural network.

[0005] To address the aforementioned technical problems, this invention discloses a method for dehazing low-light images based on a lightweight deep neural network, comprising the following steps:

[0006] Step 1: Construct a training dataset and a test dataset containing synthetic low-light fog images and real low-light fog images;

[0007] Step 2: Construct an end-to-end lightweight deep neural network for dehazing low-light images;

[0008] Step 3: Construct the target loss function of the end-to-end lightweight deep neural network for low-light image dehazing;

[0009] Step 4: Train the end-to-end lightweight deep neural network for low-light image dehazing constructed in Step 2 using the training dataset constructed in Step 1.

[0010] Step 5: Input the foggy image in the low-light scene into the end-to-end lightweight deep neural network for low-light image dehazing trained in step 4 to obtain the dehazed image, thus completing the low-light image dehazing based on the lightweight deep neural network.

[0011] Beneficial effects:

[0012] First, the network of this invention, by introducing a lightweight multi-level feature fusion module and a lightweight channel attention module, not only ensures good dehazing effect, but also has the advantages of low resource consumption, small number of parameters and low computational load.

[0013] Secondly, by introducing content loss and color loss to jointly optimize the network, the network can better correct image color cast, improve visual effect, restore image details and compensate for the impact of low light on dehazing while maintaining the contrast of the restored image. Attached Figure Description

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0015] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0016] Figure 2 This is a schematic diagram of the lightweight multi-level feature fusion submodule structure.

[0017] Figure 3 This is a schematic diagram of the lightweight channel attention submodule structure.

[0018] Figure 4 A schematic diagram of a lightweight deep neural network architecture for dehazing low-light images.

[0019] Figure 5 This is a schematic diagram comparing the fogging datasets of the present invention.

[0020] Figure 6 This is a schematic diagram illustrating the processing effect of the present invention on the synthesized fog image.

[0021] Figure 7 This is a schematic diagram illustrating the processing effect of the present invention on a real fog image. Detailed Implementation

[0022] The specific embodiments and effects of the present invention will be further described below with reference to the accompanying drawings.

[0023] like Figure 1 As shown, the specific implementation steps of the present invention are as follows:

[0024] Step 1: Construct training and testing datasets containing synthetic low-light fog images and real low-light fog images. Two different datasets were selected from publicly available low-light dehazing datasets to train and test the network: the HDP-Net dataset and the 3R dataset.

[0025] 1.1) The artificially fogged images publicly available in the HDP dataset contain artifacts. To alleviate or even avoid these artifacts, the algorithm is re-applied to fog the clear images in the HDP dataset to obtain a new HDP fogged dataset. This invention uses a method of synthesizing fog maps based on image depth maps for fogging. First, the depth map d(x) of the clear image is estimated using the depth estimation network proposed by Li et al. (see LI B, REN W, FU D, et al. Benchmarking single-image dehazing and beyond[J].IEEE Transactions on Image Processing,2018,28(1):492-505). Then, given the known image depth map, a new fogged dataset is constructed by combining the atmospheric scattering model I(x)=L(x)η(x)R(x)t(x)+L(x)η(x)(1-t(x)).

[0026] Where x is the pixel index, I(x) is the low-light, foggy image, L(x) and η(x) are the light intensity and color cast of the light source, R(x) is the reflectivity, and t(x) is the haze transmittance. The global transmittance value t(x) = e is obtained from the depth map d(x) of the image. -βd(x) Furthermore, the image after fogging is obtained, where β is set to 1.3. At this point, the generated fog image is closer to the real fog image. For example... Figure 5 As shown, the original low-light, foggy image is... Figure 5 (a) and new low-light foggy images Figure 5(b) shows a comparison between the image in the HDP dataset after fogging according to the present invention and the image in the original dataset.

[0027] 1.2) The new HDP fog dataset and 3R dataset are used to construct training and testing datasets containing synthetic low-light fog maps and real low-light fog maps.

[0028] Step 2: Establish a lightweight multi-level feature fusion module. The following are used in the construction process: Convolutional layer, see Fukushima, K., 1980. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics, 36(4), 193-202; ReLU activation function layer, see R Hahnloser, HS Seung (2001). Permitted and Forbidden Sets in Symmetric Threshold-Linear Networks. NIPS2001; BN (BatchNorm) normalization layer, see Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[J]. arXiv preprint arXiv:1502.03167,2015; Ghost convolutional layer, see HAN K, WANG Y, TIAN Q, et al. Ghostnet: More features from cheap operations[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition.2020:1580-1589; shuffle operation, see ZHANG recognition.2018:6848-6856; Sigmoid function see Han, Jun; Morag, Claudio.The influence of the sigmoid function parameters on the speed of backpropagation learning: From Natural to Artificial NeuralComputation., 1995: 195–201. .

[0029] like Figure 2 As shown, this step is implemented as follows:

[0030] 2.1) Establish a separate convolutional submodule. Convolutional layers with kernel size of 3*1 and stride of 1, ReLU activation function layers, convolutional layers with kernel size of 1*3 and stride of 1, BN normalization layers and ReLU activation function layers are stacked sequentially. The features output by the stacked layers and the original input features are concatenated along the channel dimension. The concatenated result is then fed into the next module in the network.

[0031] 2.2) Establish the GSC (Ghost Shuffle Convolution Block) sub-module: A convolutional layer with a kernel size of 1*1 and a stride of 1, a BN normalization layer, and a ReLU activation function layer are sequentially stacked to obtain feature a. Feature a is then fed into a Ghost convolutional layer with a kernel size of 3*3, a BN normalization layer, and a ReLU activation function layer, which are sequentially stacked to obtain feature b. Feature a and feature b are then concatenated along the channel dimension. Finally, a shuffle operation is performed, and the result is fed into the next module.

[0032] 2.3) The input of the lightweight multi-level feature fusion module is fed into the separable convolution sub-module. The output result is concatenated with the initial input dimension of the module and then fed into the GSC sub-module. The output result is concatenated with the upper-level input and the initial input in the same dimension. That is, the original input feature a is concatenated with the input feature a in the channel dimension through the output b of the separable convolution sub-module to obtain feature c. Feature c is input into the GSC sub-module to obtain feature d. The original input feature a, feature c and feature d are concatenated in the channel dimension to obtain the result e, which is the output of the lightweight multi-level feature fusion module.

[0033] Step 3: Establish a lightweight channel attention module:

[0034] like Figure 3 As shown, this module includes grouped convolutional layers, SE submodules, convolutional layers, and shuffle layers, where:

[0035] The grouped convolutional layer consists of cascaded grouped convolutions with a kernel size of 3*3 and BN normalized layers;

[0036] The SE submodule sequentially stacks an average pooling layer, a convolutional layer with a kernel size of 1*1, a BN normalization layer, a ReLU activation function layer, another convolutional layer with a kernel size of 1*1, and a Sigmoid activation function layer. It then performs a pixel-wise multiplication operation between the features output from the stacked layers and the original input features, and feeds the result of the operation into the next module in the network.

[0037] The convolutional layer consists of cascaded convolutional layers with a kernel size of 1*1 and BN normalized layers;

[0038] The structural relationship between the parts is as follows: the grouped convolutional layer, the SE submodule, the convolutional layer, and the shuffle layer are connected in series and then added to the original input as residuals. The result of the residual addition is the module output.

[0039] Step 4: Using the open-source PyTorch deep learning framework (programmed in Python by Facebook), build a lightweight low-light dehazing network using various modules:

[0040] like Figure 4 As shown, the lightweight low-light dehazing network can be divided into an encoding end and a decoding end, both of which include four layers, and the layers are connected by pixel-by-pixel addition.

[0041] The first layer of the encoder takes the original features as input to a convolutional layer to obtain feature a1. Feature a1 is then input to two sequentially stacked lightweight channel attention submodules to output feature b1. Feature a1 and feature b1 are then added pixel by pixel to obtain feature c1, which is the output of the first layer of the encoder. The convolutional layer has padding of 5, a kernel size of 11*11, and a stride of 1.

[0042] The second layer of the encoder takes the output of the first layer (i.e., feature c1) as input. Feature c1 is input into a convolutional layer and a lightweight multi-level feature fusion submodule to obtain feature a2. Feature a2 is then input into two sequentially stacked lightweight channel attention submodules to output feature b2. Feature a2 and feature b2 are added pixel by pixel to obtain feature c2, which is the output of the second layer. The convolutional layer has padding of 1, a kernel size of 3*3, and a stride of 2.

[0043] The third and fourth layers of the encoding are the same as the second layer, resulting in outputs c3 and c4. The convolutional layers have padding of 1, a kernel size of 3*3, and a stride of 2.

[0044] The fourth layer of the decoder takes the output of the fourth layer of the encoder (i.e., feature c4) as input d4 and sequentially superimposes it through 11 lightweight channel attention submodules to obtain feature e4. It then adds feature e4 to feature d4 pixel by pixel to obtain feature f4, which is input to the lightweight multi-level feature fusion submodule to obtain the decoder output g4.

[0045] The input to the third layer of the decoder is the pixel-wise sum of the output c3 from the third layer of the encoder and the output g4 from the fourth layer of the decoder. This sum is passed through a deconvolution layer to obtain d3, which is then sequentially stacked through two lightweight channel attention submodules to obtain feature e4. This e4 is then added pixel-wise to feature d3 to obtain feature f3, which is input to a lightweight multi-level feature fusion submodule to obtain the decoder output g3. The deconvolution layer has a padding of 1, a kernel size of 3*3, and a stride of 2.

[0046] The second and third layers of the decoding end also use the output c of this layer and the output g of the previous layer as input, pixel by pixel, and obtain the decoding outputs g2 and g1 through the same operation as the third layer of the decoding end. The deconvolution layer has a padding of 1, a kernel size of 3*3, and a stride of 2.

[0047] The first layer output g1 at the decoding end is followed by a convolutional layer with padding of 1, a kernel size of 3*3, and a stride of 1 to obtain the final output of the entire lightweight low-light dehazing network.

[0048] Step 5: Construct the network objective loss function.

[0049] Use L cont Content loss and L color Color loss is used to jointly optimize the network, and the overall loss function is:

[0050] Loss = L cont +L color

[0051] The formula for calculating content loss is as follows:

[0052]

[0053] Where n is the number of data points, X is the fog-free image predicted by the network in this method, and Y is the corresponding clear fog-free image.

[0054] Color loss is evaluated using the CIEDE2000 color difference formula, a standard color evaluation formula for object colors provided by the International Commission on Illumination to the color-using industry. See SHARMA G, WU W, DALAL E N. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations [J]. Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre de laCouleur, 2005, 30(1):21-30. The calculation formula is as follows:

[0055]

[0056] Where, ΔE 00 Indicates the total color difference; This method uses a network to predict the color values ​​of haze-free images in the CIELAB (also known as L*a*b* color space, where L* represents perceived brightness and a* and b* represent the four unique colors of human vision: red, green, blue, and yellow) color space. The corresponding clear, haze-free image has its color value in the CIELAB color space; ΔL ′ ΔC represents the brightness difference between the predicted fog-free image and the corresponding sharp fog-free image. ′ Indicates chroma difference, ΔH ′ Indicates hue difference; k L ,k C ,k H S is the parameter weighting factor, with a value of 1; L ,S C ,S H Weighting function for correcting color space uniformity; R T A rotation function to correct the deflection of the principal axis of the tolerance ellipse in the blue region of the color space.

[0057] Step 6: Train the network constructed in Step 2 using the dataset from Step 1.

[0058] 6.1) The training image set is divided into multiple batches, each batch consisting of eight paired image groups. The first image group is input simultaneously for training to obtain the initial weights W for each convolutional operation of the network in this paper. n and bias value B n The weights D of the deconvolution operation n And the estimated corresponding low-light, fog-free image group.

[0059] 6.2) Substitute the set of clear, fog-free images and the estimated corresponding low-light, fog-free images into the total loss formula to calculate the total loss value corresponding to the training images;

[0060] 6.3) Using the Adam optimizer in the PyTorch architecture, which is used to manage and update the learnable parameter values ​​in the model, update the network parameters, take minimizing the total loss value as the objective, set the initial learning rate to 0.0002, and change it to 0.1 times the previous value every 100 rounds, and update the parameters in the low-light dehazing network.

[0061] 6.4) Input the second set of images into the network after the first parameter update, and repeat steps (4a) to (4c) to obtain the network after the second parameter update; and so on, until the last set of images is input into the network after the previous update, to obtain a lightweight low-light dehazing network after one training.

[0062] 6.5) Input all image groups into the network after one training iteration to obtain the network after the second training iteration; repeat this process until all image groups have been input 400 times to complete the training of the lightweight low-light dehazing network.

[0063] Step 7: Dehaze using a lightweight, low-light dehazing network:

[0064] The foggy image in the low-light scene is input into the network trained in step 6 to obtain the dehazed image. The images before and after dehazing are compared, and the network's running speed, computational load, and parameter count are also compared.

[0065] The effects of this invention will be further illustrated below with simulation experiments:

[0066] I. Test Conditions and Methods

[0067] 1. Simulation experimental conditions:

[0068] CPU: Intel Core i9 10900X, clock speed 3.7GHz, RAM 64GB;

[0069] Graphics card: NVIDIA GeForce RTX 3090;

[0070] Operating system: Ubuntu 18.04, CUDA version 11.3, PyTorch version 1.10, Python version 3.8.

[0071] 2. Test images: The test dataset constructed in step (1);

[0072] 3. Simulation content and result analysis:

[0073] Simulation Test 1: The algorithm of this invention was tested on a synthetic dataset, and the results are as follows. Figure 6 As shown, where:

[0074] Figure 6 a is a composite fog image of four images;

[0075] Figure 6 b is to use the network of the present invention to Figure 6 The result of defogging a synthetic fog image;

[0076] Figure 6 c is a clear image without fog;

[0077] from Figure 6 It can be seen that, in terms of subjective effect, the network processing result of this invention is close to a real and clear image.

[0078] Simulation Test 2: The algorithm of this invention was tested on a real dataset, and the results are as follows. Figure 7 As shown, where:

[0079] Figure 7 a represents four real low-light fog images;

[0080] Figure 7 b is to use the network of the present invention to Figure 7 The result of defogging the fog map;

[0081] from Figure 7 It can be seen that, in terms of subjective effect, the defogging result achieved by the network of this invention is significantly clearer.

[0082] Simulation Test 3: The dataset was cropped into images of uniform size of 512×512 pixels. Three existing dehazing algorithms were tested and compared, namely MSBDNet (a multi-scale enhanced dehazing network with dense feature fusion), GCANet (a gated context aggregation network for image dehazing and rain removal), and FFANet (a feature fusion attention network for single image dehazing), to evaluate the network's speed, computational load, and parameter count.

[0083] Algorithm Name Runtime / s Computational complexity / M Parameter quantity / M MSBDNet 0.037 24536 28.7 GCANet 0.021 18397 0.7 FFANet 0.015 126706 2.0 Algorithm of this invention 0.02 3407 1.1

[0084] From the perspective of resource consumption and running speed, the network computation in this paper is minimal, and the number of parameters is second only to GCANet, while the running speed is only 5 milliseconds slower than FFANet. Considering these two factors, the algorithm of this invention is superior to other comparative algorithms overall.

[0085] Based on the above test results, the lightweight network-based image low-light dehazing network of the present invention achieves the goal of dehazing images in low-light scenes while reducing resource consumption and improving running speed.

[0086] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content, particularly a low-light image dehazing method based on a lightweight deep neural network, and some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0087] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0088] This invention provides a concept and method for low-light image dehazing based on lightweight deep neural networks. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for dehazing low-light images based on lightweight deep neural networks, characterized in that, Includes the following steps: Step 1: Construct a training dataset and a test dataset containing synthetic low-light fog images and real low-light fog images; Step 2: Construct an end-to-end lightweight deep neural network for dehazing low-light images; Step 3: Construct the target loss function of the end-to-end lightweight deep neural network for low-light image dehazing; Step 4: Train the end-to-end lightweight deep neural network for low-light image dehazing constructed in Step 2 using the training dataset constructed in Step 1. Step 5: Input the foggy image in the low-light scene into the end-to-end lightweight deep neural network for low-light image dehazing trained in step 4 to obtain the dehazed image, thus completing the low-light image dehazing based on the lightweight deep neural network. The end-to-end lightweight deep neural network for low-light image dehazing described in step 2 uses an encoder-decoder architecture. The encoder transforms the input into features, and the decoder transforms these features into the desired target, which is the dehazed image. The encoder-decoder architecture consists of four layers, each composed of a convolutional layer, a lightweight multi-level feature fusion module, a lightweight channel attention module, and a deconvolutional layer. The specific construction method is as follows: Step 2-1: Establish a lightweight multi-level feature fusion module; Step 2-2: Establish a lightweight channel attention module; Steps 2-3: Construct the end-to-end lightweight deep neural network for low-light image dehazing; The specific method for establishing the lightweight multi-level feature fusion module described in step 2-1 is as follows: Step 2-1-1: Establish separate convolutional submodules and GSC submodules respectively; Step 2-1-2: The input of the lightweight multi-level feature fusion module is fed into the split convolution sub-module. The output result is then dimensionally concatenated with the initial input of the lightweight multi-level feature fusion module and fed into the GSC sub-module. The output result is then dimensionally concatenated with the input of the GSC sub-module and the initial input to form the lightweight multi-level feature fusion module. The GSC submodule described in step 2-1-1 has the following structure: Feature a is obtained by sequentially stacking a convolutional layer with a kernel size of 1*1 and a stride of 1, a BN normalization layer, and a ReLU activation function layer. Feature a is then fed into a Ghost convolutional layer with a kernel size of 3*3, a BN normalization layer, and a ReLU activation function layer, which are then stacked sequentially to obtain feature b. Feature a and feature b are concatenated along the channel dimension, and finally a shuffle operation is performed to obtain the output of the GSC submodule.

2. The low-light image dehazing method based on a lightweight deep neural network according to claim 1, characterized in that, The method for obtaining the synthetic low-light fog image described in step 1 is as follows: Based on clear images in the existing dataset, a method for synthesizing fog maps from image depth maps is used for fogging. Given the depth map of the image, combined with an atmospheric scattering model, the processing procedure is as follows: ； in, It is a pixel index. It is a low-light, foggy image. and It refers to the light intensity and color deviation of the light source. It's reflectivity. This refers to the transmittance of haze; the global transmittance value is obtained from the depth map. The image after fogging is obtained. .

3. The low-light image dehazing method based on a lightweight deep neural network according to claim 2, characterized in that, The structure of the separated convolutional submodule described in step 2-1-1 includes: A convolutional layer with a kernel size of 3*1 and a stride of 1, a first ReLU activation function layer, a convolutional layer with a kernel size of 1*3 and a stride of 1, a BN normalization layer, and a second ReLU activation function layer are sequentially stacked as a stacking layer. The features output by the stacking layer and the original input features of the separated convolutional submodule are concatenated along the channel dimension. The concatenated result is the output of the separated convolutional submodule.

4. The low-light image dehazing method based on a lightweight deep neural network according to claim 3, characterized in that, The method for establishing a lightweight channel attention module as described in step 2-2 includes: The lightweight channel attention module includes: a grouped convolutional layer, an SE submodule, a convolutional layer, and a shuffle layer. The grouped convolutional layer consists of cascaded grouped convolutions with a kernel size of 3*3 and BN normalized layers. The SE submodule includes: sequentially stacking an average pooling layer, a convolutional layer with a kernel size of 1*1, a BN normalization layer, a ReLU activation function layer, a convolutional layer with a kernel size of 1*1, and a Sigmoid activation function layer, and performing a pixel-wise multiplication operation on the output features of the stacked layer and the original input features of the SE submodule. The result of the above operation is the output of the SE submodule. The convolutional layer consists of a cascade of convolutional layers with a kernel size of 1*1 and BN normalized layers; The structure of the lightweight channel attention module is as follows: the grouped convolutional layer, the SE submodule, the convolutional layer and the shuffle layer are connected in series and then added to the original input of the lightweight channel attention module. The result of the residual addition is the output of the lightweight channel attention module.

5. The low-light image dehazing method based on a lightweight deep neural network according to claim 4, characterized in that, The method for constructing the end-to-end lightweight deep neural network for low-light image dehazing, as described in steps 2-3, includes: using convolutional units, residual modules, a constructed lightweight multi-level feature fusion module, and a lightweight channel attention module within the PyTorch framework to construct an encoder-decoder structure, further forming an end-to-end lightweight deep neural network for low-light image dehazing. The network comprises a four-layer encoder-decoder structure, with layers connected by pixel-wise addition. The specific method is as follows: The first layer of the encoder takes the original features as input to a convolutional layer to obtain feature a1. Feature a1 is then input to two sequentially stacked lightweight channel attention submodules to output feature b1. Feature a1 and feature b1 are then added pixel by pixel to obtain feature c1, which is the output of the first layer of the encoder. The convolutional layer has padding of 5, a kernel size of 11*11, and a stride of 1. The second layer of the encoder takes the output of the first layer, i.e., feature c1, as input. Feature c1 is input into a convolutional layer and a lightweight multi-level feature fusion submodule to obtain feature a2. Feature a2 is input into two sequentially stacked lightweight channel attention submodules to output feature b2. Feature a2 and feature b2 are added pixel by pixel to obtain feature c2, which is the output of the second layer. The convolutional layer padding is 1, the convolutional kernel size is 3*3, and the stride is 2. The third and fourth layers of the encoding end have the same structure as the second layer, and the outputs are c3 and c4 respectively. The convolutional layer padding is 1, the convolutional kernel size is 3*3, and the stride is 2. The fourth layer of the decoder takes the output of the fourth layer of the encoder, i.e., feature c4, as input d4. It is sequentially superimposed through 11 lightweight channel attention sub-modules to obtain feature e4. It is then added to feature d4 pixel by pixel to obtain feature f4. It is then input into the lightweight multi-level feature fusion sub-module to obtain the output g4 of the decoder. The input to the third layer of the decoder is the pixel-wise addition of the output c3 of the third layer of the encoder and the output g4 of the fourth layer of the decoder. After passing through a deconvolution layer, d3 is obtained. Then, after passing through two lightweight channel attention submodules in sequence, feature e4 is obtained. It is added to feature d3 pixel by pixel to obtain feature f3. The input to the lightweight multi-level feature fusion submodule is used to obtain the output g3 of the decoder. The deconvolution layer padding is 1, the convolution kernel size is 3*3, and the convolution stride is 2. The second and first layers of the decoding end also use the output c of this layer and the output g of the upper layer pixel by pixel as input. After the same operation as the third layer of the decoding end, the decoding end outputs g2 and g1 are obtained respectively. The deconvolution layer padding is 1, the convolution kernel size is 3*3, and the convolution stride is 2. The first layer output g1 at the decoding end is followed by a convolutional layer with padding of 1, a kernel size of 3*3, and a stride of 1 to obtain the final output of the entire end-to-end lightweight deep neural network used for dehazing low-light images.

6. The low-light image dehazing method based on a lightweight deep neural network according to claim 5, characterized in that, The specific method for constructing the target loss function of the end-to-end lightweight deep neural network for low-light image dehazing described in step 3 is as follows: Content loss and color loss The end-to-end lightweight deep neural network for low-light image dehazing is jointly optimized, and the overall loss function is: ； The content loss calculation method is as follows: ； in, It is the number of data points. This is the dehazed image obtained by the i-th end-to-end lightweight deep neural network using low-light image dehazing after predicting foggy images in the training dataset. The corresponding i-th clear and fog-free image is the clear image mentioned in step 1; Color loss is calculated using the following method: ； in, The total color difference between the haze-free image and the clear haze-free image predicted by the end-to-end lightweight deep neural network used for dehazing low-light images; The color value in CIELAB space for the haze-free image predicted by the end-to-end lightweight deep neural network used for dehazing low-light images. This represents the color value of the corresponding clear, haze-free image in the CIELAB color space. This represents the brightness difference between the predicted fog-free image and the corresponding clear fog-free image. Indicates poor chroma. Indicates hue difference; This is the parameter weighting factor, with a value of 1; A weighting function for correcting color space uniformity; A rotation function to correct the deflection of the principal axis of the tolerance ellipse in the blue region of the color space.

7. The low-light image dehazing method based on a lightweight deep neural network according to claim 6, characterized in that, The specific training methods described in step 4 include: Step 4-1: Divide the training dataset mentioned in Step 1 into batches, with each batch consisting of eight paired image groups. Simultaneously input the first image group for training to obtain the initial weights of each convolutional operation in the end-to-end lightweight deep neural network used for low-light image dehazing. and bias value Weights of deconvolution operation And the estimated corresponding low-light, fog-free image set; Step 4-1: Substitute the clear images from the existing dataset used in Step 1 when synthesizing the low-light fog image with the estimated corresponding low-light fog-free image into the calculation of the total loss value corresponding to the image used for training. Step 4-2: Update the network parameters using the Adam optimizer, with the goal of minimizing the total loss. The initial learning rate is set to 0.0002, and it is changed to 0.1 times the previous rate every 100 rounds. Update the parameters in the end-to-end lightweight deep neural network for low-light image dehazing. Step 4-3: Input the second image group into the end-to-end lightweight deep neural network for low-light image dehazing after the first parameter update, and repeat steps 4-1 to 4-3 to obtain the network after the second parameter update; and so on, until the last image group of this batch is input into the network after the previous update to obtain the end-to-end lightweight deep neural network for low-light image dehazing after one training. Step 4-4: Input all image groups of the next batch into the network that has completed one training iteration to obtain the network after the second training iteration; and so on, until all image groups of all batches have been input 400 times to complete the training of the end-to-end lightweight deep neural network for dehazing low-light images.

Images