A Single Image Dehazing Method and System Based on Neural Network

Abstract

The present invention discloses a single-image defogging method and system based on a neural network; including: acquiring a foggy image of a target environment; preprocessing the foggy image of the target environment, and extracting features from the foggy image of the target environment during preprocessing to obtain a first feature; sequentially extracting features from the preprocessed foggy image of the target environment through multiple-scale convolutional kernels and an encoder-decoder to obtain a second feature; performing a mapping process on the first feature and the second feature to obtain a mapping map; concatenating the foggy image of the target environment with the mapping map, and obtaining an estimated value of the combined parameter based on a convolutional layer; inputting the estimated value of the combined parameter into a combined parameter model to output a defogged image; realizing the estimation of unified parameters with a lightweight network, obtaining a relatively excellent and general defogging effect, and finally the obtained defogged image can preserve more details, effectively avoiding the problems of detail loss and color distortion of common defogging algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of image defogging, and particularly relates to a single-image defogging method and system based on a neural network. Background Art

[0002] In a foggy or hazy scene, the images captured by optical sensors are affected by the scattering of atmospheric particles, resulting in missing details, dull colors, reduced brightness, and decreased image contrast and color fidelity in the images. This directly affects people's visual perception of the images, and also affects the subsequent processing of the images, and even affects the operation of various systems relying on optical imaging instruments, such as satellite remote sensing systems, aerial photography systems, outdoor monitoring and target recognition systems, etc. Therefore, image defogging is a very meaningful task. According to the atmospheric scattering model, to achieve image defogging, it is necessary to accurately estimate the atmospheric layer coefficient and the transmittance. These two important physical parameters are very difficult to accurately estimate, which is a major challenge in the field of image defogging.

[0003] Dr. He Kaiming's dark channel theory can estimate the atmospheric layer coefficient and the transmittance based on the dark channel map of a foggy image, thereby achieving image defogging. However, the dark channel theory does not perform well in processing images containing large sky areas or large white buildings, etc.; the fast visibility restoration (FVR) method is different from the method of the dark channel theory. It does not estimate the transmittance and the atmospheric light value, and pays more attention to the latter part of the formula. Since it does not estimate the parameter values, the defogging speed is relatively fast, but the effect is poor; the atmospheric light adaptive restoration algorithm ATM pays more attention to the estimation of the atmospheric light. The atmospheric light value can be obtained based on the fact that the pixels of each patch are distributed on the same line in the RGB color space. However, the images defogged by this method can effectively avoid the influence of the deviation of the atmospheric light value on color distortion; the automatic color level (AL) automatically defines the brightest and darkest pixels in each channel as white and black, and then redistributes the pixel values between them proportionally. It can relatively simply process foggy images, but the effect is not particularly obvious.

[0004] With the rapid development of deep learning, people have gradually applied deep learning to image defogging. Deep learning is mainly divided into supervised learning and unsupervised learning. Thus, there are two neural network training methods for image defogging: one is to use a convolutional neural network (CNN) for model estimation, and the other is to use a generative adversarial network (GAN) to generate defogged images.

[0005] In the application of the convolutional neural network method for haze removal, a relatively typical method is the end-to-end training model adopted by DehazeNet. It uses a neural network to estimate the transmittance in the atmospheric degradation model. The input of the model is a hazy image, and the output is the estimated value of the transmittance, and then the image is restored. This network proposes a new non-linear activation function to improve the quality of the haze-free image. FFA-NET (Feature Fusion Attention Network for Single Image Dehazing) proposes an end-to-end feature fusion attention network (FFA-Net) for directly restoring haze-free images. Among them, the FFA network architecture takes into account that different channel features contain completely different weighted information, and the haze distribution on different image pixels is uneven, and proposes a new feature attention (FA) module, which combines the channel attention and pixel attention mechanisms. FA processes different features and pixels unequally, which provides additional flexibility for processing different types of information and expands the representation ability of CNNs. MSCNN generates a rough transmission matrix by training the input hazy image and then refines it, and errors are inevitably generated. Gated Context Aggregation Network for Image Dehazingand Deraining 2019WACV uses the GAN network to achieve end-to-end image haze removal. The focus of this article is to solve the problem of grid artifacts, and there has been a great improvement in the indicators of PSNR and SSIM.

[0006] GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing. This paper proposes an end-to-end trainable CNN, namely GridDehazeNet. GridDehazeNet consists of three modules: a preprocessing module, a backbone module, and a postprocessing module. The trainable preprocessing module can generate more diverse and targeted inputs compared to manually selected preprocessing methods. The backbone module implements a new attention-based multi-scale estimation, effectively alleviating the bottleneck problems often encountered in traditional multi-scale methods. The postprocessing module helps reduce defects in the final output. However, in this paper, AOD Net does not separately estimate the transmission matrix and the atmospheric light coefficient. Instead, it unifies the two parameters into one parameter through formula transformation and finally directly generates a clear image through a lightweight CNN. Moreover, this paper has the following drawbacks: The preprocessing module consists only of simple convolutions and residual dense blocks, and some details may be lost during data processing; The backbone module is an enhanced version of GridNet originally proposed for semantic segmentation tasks. It adopts a grid structure of three rows and six columns and also introduces a large number of attention mechanism modules for feature extraction of feature maps, with a relatively large computational amount and an unsatisfactory operation speed; The postprocessing module is symmetric and the same as the preprocessing module in structure, only consisting of simple convolutions and residual dense blocks. Due to its relatively simple network structure, it cannot better estimate the unified parameter, and the achieved dehazing effect is not ideal.

[0007] Therefore, how to draw on the ideas of the above AOD network, not separately estimate the two important parameters of transmittance and atmospheric layer coefficient, and at the same time improve the preprocessing module, backbone module, and postprocessing module in network design to achieve an optimized dehazing effect has become a key issue in current research. Summary of the Invention

[0008] In view of the above problems, the present invention provides a single-image dehazing method and system that at least solves some of the above technical problems, realizes estimating unified parameters with a lightweight network, obtains a relatively excellent and general dehazing effect, and the finally obtained dehazed image can preserve more details, effectively avoiding the problems of detail loss and color distortion in common dehazing algorithms.

[0009] On the one hand, an embodiment of the present invention provides a single-image dehazing method based on a neural network, including:

[0010] S1. Obtain a foggy image of the target environment;

[0011] S2. Preprocess the foggy image of the target environment and extract features from the foggy image of the target environment during preprocessing to obtain a first feature;

[0012] S3. Sequentially perform feature extraction on the preprocessed foggy image of the target environment through convolutional kernels of multiple scales and an encoder-decoder to obtain a second feature;

[0013] S4. Perform mapping processing on the first feature and the second feature to obtain a mapping graph;

[0014] S5. Concatenate the foggy image of the target environment in S1 with the mapping graph, and obtain an estimated value of the merging parameter based on a convolutional layer;

[0015] S6. Input the estimated value of the merging parameter into a merging parameter model to output a defogged image.

[0016] Further, the S2 specifically includes:

[0017] Perform scale conversion on the foggy image of the target environment;

[0018] Divide the foggy image of the target environment after scale conversion into two partial images of the same size;

[0019] Perform feature extraction on the two partial images respectively to obtain a first feature; during the feature extraction process, collect convolutional kernels to perform convolutional processing on the two partial images;

[0020] Merge the two partial images after convolutional processing into a complete image.

[0021] Further, during the convolutional processing of the two partial images, the non-linear activation function adopts the PReLU function.

[0022] Further, in the S3, when performing feature extraction on the preprocessed foggy image of the target environment through convolutional kernels of multiple scales, add a pooling layer and a sampling layer after each convolutional layer.

[0023] Further, the S4 specifically includes:

[0024] Obtain a feature map based on the first feature and the second feature;

[0025] Perform mapping processing on the feature map through a convolutional layer, a pooling layer, and an upsampling layer to obtain a mapping graph of a preset size.

[0026] Further, the merging parameter model is expressed as:

[0027] J(x) = K(x) * I(x) - K(x) + b

[0028] where x represents an image; I(x) represents the foggy image of the target environment; J(x) represents the defogged image; K(x) represents the estimated value of the merging parameter; and b represents a fixed bias value.

[0029] Furthermore, the loss function of the merging parameter model is expressed as:

[0030]

[0031]

[0032]

[0033] Among them, J(x) represents the known dehazed image; represents the dehazed image output by the merging parameter model; represents the estimated value of the merging parameter; I(x) represents the foggy image of the target environment; b represents the fixed bias value.

[0034] On the other hand, an embodiment of the present invention provides a single-image dehazing system based on a neural network, which applies the above-mentioned single-image dehazing method based on a neural network; the system includes: an input module, a preprocessing module, a backbone module, a postprocessing module, and an output module;

[0035] The input module is used to obtain the foggy image of the target environment;

[0036] The preprocessing module is used to preprocess the foggy image of the target environment and extract features from the foggy image of the target environment during preprocessing to obtain a first feature;

[0037] The backbone module is used to sequentially extract features from the preprocessed foggy image of the target environment through multiple-scale convolutional kernels and an encoder-decoder to obtain a second feature;

[0038] The postprocessing module is used to perform mapping processing on the first feature and the second feature to obtain a mapping diagram; and concatenate the foggy image of the target environment with the mapping diagram, and obtain an estimated value of the merging parameter based on a convolutional layer;

[0039] The output module is used to input the estimated value of the merging parameter into the merging parameter model and output a dehazed image.

[0040] Compared with the prior art, a single-image dehazing method based on a neural network recorded in the present invention has the following beneficial effects:

[0041] In the preprocessing process of the present invention, features are extracted from the foggy image of the target environment. The advantage is that useless or unimportant features only account for a small part of the feature map, rather than being completely eliminated, ensuring that certain details are not lost.

[0042] In the present invention, the preprocessed hazy image of the target environment is subjected to feature extraction through a variety of scale convolution kernels and an encoder-decoder in sequence, which can achieve lightweight and perform convolution sampling using convolution kernels of different sizes, ensuring excellent accuracy in operations.

[0043] In the present invention, mapping processing is performed on the first feature and the second feature, and based on this, an estimated value of the combined parameter is obtained, ensuring that the model is suitable for image denoising in a variety of different environments and making the defogging effect more perfect.

[0044] Other features and advantages of the present invention will be described in the following specification, and in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the written specification, claims, and drawings.

[0045] The technical solutions of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0046] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0047] Figure 1 It is a schematic flowchart of a single-image defogging method based on a neural network provided by an embodiment of the present invention.

[0048] Figure 2 It is a framework diagram of a preprocessing network structure provided by an embodiment of the present invention.

[0049] Figure 3 It is a framework diagram of a CAB module structure provided by an embodiment of the present invention

[0050] Figure 4 It is a framework diagram of a CALayer structure provided by an embodiment of the present invention.

[0051] Figure 5 It is a schematic diagram of a multi-scale module structure provided by an embodiment of the present invention.

[0052] Figure 6 It is a schematic diagram of an encoder-decoder structure provided by an embodiment of the present invention.

[0053] Figure 7 It is a schematic diagram of a downsampling structure provided by an embodiment of the present invention.

[0054] Figure 8 It is a schematic diagram of an encoder structure provided by an embodiment of the present invention.

[0055] Figure 9Schematic diagram of the upsampling structure provided by the embodiment of the present invention.

[0056] Figure 10 Schematic diagram of the decoder structure provided by the embodiment of the present invention.

[0057] Figure 11 Schematic diagram of the structure of the non - linear mapping part provided by the embodiment of the present invention.

[0058] Figure 12 Schematic diagram of the process of generating a defogged image provided by the embodiment of the present invention.

[0059] Figure 13 Schematic diagram of the atmospheric imaging process provided by the related prior art.

[0060] Figure 14 Schematic diagram of the framework of the single - image defogging system based on neural network provided by the embodiment of the present invention. Detailed implementation manners

[0061] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0062] See Figure 1 As shown, the embodiment of the present invention provides a single - image defogging method based on neural network, which specifically includes the following steps:

[0063] S1. Obtain a foggy image of the target environment;

[0064] S2. Pre - process the foggy image of the target environment, and extract features from the foggy image of the target environment during pre - processing to obtain a first feature;

[0065] S3. Sequentially extract features from the pre - processed foggy image of the target environment through multiple - scale convolutional kernels and an encoder - decoder to obtain a second feature;

[0066] S4. Perform mapping processing on the first feature and the second feature to obtain a mapping graph;

[0067] S5. Concatenate the foggy image of the target environment in S1 with the mapping graph, and obtain a combined parameter estimation value based on a convolutional layer;

[0068] S6. Input the combined parameter estimation value into a combined parameter model to output a defogged image.

[0069] The above steps will be described in detail below.

[0070] In the above step S2, preprocessing is performed on the target hazy image obtained in step S1, specifically including: performing scale conversion on the target hazy image to convert it into a preset size, for example, in the embodiment of the present invention, it is converted into a size of 560×560, which is convenient for subsequent image processing; dividing the target hazy image after scale conversion into two parts of the same size from the middle; respectively performing feature extraction on these two parts of the images to obtain the first feature; during the feature extraction process, a 3×3 convolution kernel is used to perform convolution processing on these two parts of the images, and the non-linear activation function uses the PRelu function during this period. The purpose is to make some useless features occupy a smaller part instead of being completely eliminated, so that certain details will not be lost and the defogging effect can reach a satisfactory level. Finally, the two parts of the images after convolution processing are merged into a complete image.

[0071] The definition formula of the above convolution processing is expressed as:

[0072]

[0073] This formula (1) represents the convolution operation, where ω i represents the convolution kernel; τ i represents the image on which the convolution operation is performed; i represents the starting position where the convolution operation begins; j represents the moving distance of the convolution kernel on the pixel points of the image to be convolved; ε represents the maximum translation distance. A digital image can be regarded as a discrete function h(x, y) in a two-dimensional space. If the operation function is f(x, y), then the output image g(x, y) = f(x, y) * h(x, y); the operation process is that after h(x, y) is translated one by one and multiplied by each value of f(x, y), the superposition is obtained for convolution.

[0074]

[0075] This formula (2) represents the preprocessing operation process, where F(1) and F(2) respectively represent the two parts of the split images; G(1) and G(2) respectively represent the two parts of the images after convolution processing; G represents the merged complete image; Δ represents image segmentation; * represents the convolution operation; represents the concat operation.

[0076]

[0077] Formula (3) represents the expression of the PRelu activation function. Among them, a is initialized to 0.25. The PReLU activation function only adds a very small number of parameters, which means that the computational amount of the network and the risk of overfitting only increase a little; in particular, when the same a is used for different channels, the number of parameters will be further reduced.

[0078] The preprocessing network structure is shown in Figure 2 As shown. Further, in the embodiments of the present invention, a convolutional kernel is first added for convolution in this part, and then a CAB module (the result after convolution and pooling of the input is combined with the original input) is adopted, and a CALayer (Convolutional Averagepooling layer) is added in the CAB module; the CAB module structure is shown in Figure 3 As shown; the residual connection method in the CAB module can avoid the problems of gradient explosion and gradient disappearance in parameter calculation, and improve the computational efficiency of the network. The CALayer structure is shown in Figure 4 As shown; the CAlayer plays a role in imitating the human visual neural network to a certain extent, performs convolution with different weights on the features of different channels, and realizes the technical effect of retaining important features and ignoring useless features. Figures 2 - 4 The size of the convolutional kernel in is 3*3.

[0079] In the above step S3, first, multi-scale convolutional kernels are used to extract features from the preprocessed foggy target environment image. Specifically, in the embodiments of the present invention, convolutional layers with sizes of 11×11, 9×9, 7×7, 5×5, 3×3, and 1×1 are adopted to form a multi-scale module for feature extraction, and a pooling layer and an upsampling layer are added after each convolutional layer, so that the image size remains unchanged after each convolution; the multi-scale module structure is as Figure 5 shown; the feature extraction process can be expressed by formula (4):

[0080] C 2 (x) = f 1 (W 1 *G + b 1 ) (4)

[0081] where f 1 (.) represents the multi-scale process; G represents the image generated through the preprocessing process; W 1 and b 1 both represent the parameters in the multi-scale convolution process; * represents convolution.

[0082] The haze in the image can be regarded as image noise. Therefore, after multi-scale convolution, an image denoising network structure, that is, an encoder-decoder, is added in the embodiments of the present invention, and its structure is as Figure 6 shown; further feature extraction is performed through the encoder-decoder, where downsampling is performed through convolution operations, and the downsampling structure is as Figure 7 shown; the encoder structure is as Figure 8 shown; upsampling operations are performed through transposed convolution operations, and the upsampling structure is as Figure 9 shown; the decoder structure is asFigure 10 As shown in; the specific process is shown in formula (5):

[0083] V 3 (x 1 ) = f 2 (W 2 *x 1 +b 2 )

[0084] x 1 = V 2 (G) (5)

[0085] where f 2 (.) represents the encoding-decoding process; x 1 represents the image generated after multi-scale convolution; W 1 and b 1 both represent the parameters in the encoding-decoder convolution process; * represents convolution.

[0086] Record the features extracted by the multi-scale convolution kernels and the encoder-decoder as the second feature.

[0087] In the above step S4, obtain a feature map based on the above first feature and second feature; perform mapping processing on the feature map through a convolutional layer, a pooling layer, and an upsampling layer to obtain a mapping map with a preset size of 560×560.

[0088] In the above step S5, concatenate the foggy image of the target environment obtained in step S1 with the mapping map generated in step S4 to minimize the information lost through convolution and pooling, and finally obtain the final combined parameter estimation value through a convolutional layer. The structural schematic diagram of the non-linear mapping part is as Figure 11 shown;

[0089] The specific process of the above steps S4 - S5 is specifically shown in formula (6):

[0090]

[0091] where, f 3 (.) represents the mapping processing process; x 2 represents the image after encoder-decoder processing; W 3 and b 3 both represent the parameters in the mapping processing process; x 3 represents the mapping map; W 4 and b 4 both represent the parameters in the combined parameter process; represents the combined parameter estimation value; represents the concat operation; F represents the original foggy image of the target environment as input.

[0092] In the above step S6, the obtained combined parameter estimation value is input into the combined parameter model to generate a defogged image; for the schematic diagram of the defogged image generation process, see Figure 12 as shown;

[0093] The combined parameter model in step S6 is obtained through the following method:

[0094] Based on the existing atmospheric imaging process diagram (see Figure 13 as shown) for analysis: In a foggy environment, dust and small water droplets in the air will have two effects on image acquisition: First, light will have energy loss due to scattering when propagating through fog and small water droplets, and the degree of loss is proportional to the distance between the object and the human eye; Second, atmospheric light enters the acquisition device, and part of the atmospheric light will enter the image acquisition device after scattering, affecting image generation, and its influence is related to the distance;

[0095] The atmospheric scattering model is a physical model formed by a detailed description of the generation process of a foggy image. It divides the image generation process into two parts: atmospheric light and the object itself entering the human eye; it can be expressed as:

[0096] I(x) = J(x) * t(x) + A * (1 - t(x)) (7)

[0097] Among them, x represents the image; I(x) represents the foggy image of the target environment; J(x) represents the defogged image; t(x) represents the transmittance; A represents the atmospheric layer coefficient. In practical applications, I(x) is a known condition, while the transmittance and the atmospheric layer coefficient need to be estimated through I(x).

[0098] Among them, the transmittance is defined as:

[0099] t(x) = e -βd(x) (8)

[0100] Among them, β represents the atmospheric layer scattering coefficient; d(x) represents the distance between the object and the imaging device.

[0101] From the atmospheric scattering model formula (7), the generation formula of the defogged image can be obtained, expressed as:

[0102]

[0103] Among them, t 0 is a very small value to prevent the denominator from being 0.

[0104] Since most deep learning methods estimate the transmittance, then calculate the atmospheric coefficients, and finally substitute them into the formula to obtain the defogged image, two parameters need to be estimated, and the errors will accumulate. To avoid this situation, in the embodiments of the present invention, the two parameters are converted into one parameter to reduce the error. The conversion formula, that is, the combined parameter model, is expressed as:

[0105] J(x) = K(x) * I(x) - K(x) + b (10)

[0106] where x represents the image; I(x) represents the foggy image of the target environment; J(x) represents the defogged image; K(x) represents the estimated value of the combined parameter; b represents the fixed bias value, and usually b = 1.

[0107] The formula of K(x) is expressed as:

[0108]

[0109] Through formulas (10) and (11), the two parameters, the transmittance t(x) and the atmospheric coefficient A, are formed into one parameter K(x); based on this, only the estimated value of the combined parameter K(x) needs to be calculated, and then the estimated value of the combined parameter K(x) is input into the above combined parameter model (formula (10)), and the defogged image can be output.

[0110] The loss function of the above combined parameter model is expressed as:

[0111]

[0112]

[0113]

[0114] where J(x) represents the known defogged image; represents the defogged image output by the combined parameter model; represents the estimated value of the combined parameter; I(x) represents the foggy image of the target environment; b represents the fixed bias value.

[0115] In summary, in the embodiment of the present invention, a single image dehazing method based on a neural network is provided. When the preprocessing module extracts features, since the PRelu activation function is used, some useless features can be made to occupy a smaller part instead of being completely eliminated, so that some details can be preserved, and the finally obtained dehazed image has richer details; since the traditional atmospheric scattering model is not adopted in the present invention, but two parameters to be solved are converted into a parameter K(x) to be solved, and the entire algorithm directly estimates the parameter K(x), effectively avoiding error accumulation, thereby reducing the common color distortion in the image dehazing algorithm, and thus realizing bright image colors; finally, the non-linear module introduced in the present invention concatenates the input original image and the feature map after multi-layer convolution processing, minimizing the feature loss and error caused by the previous pooling and convolution operations to the greatest extent, effectively avoiding the common image darkening and color distortion in image dehazing, and improving the brightness of the dehazed image.

[0116] See Figure 14 As shown, in the embodiment of the present invention, a single image dehazing system based on a neural network is also provided, which applies the above-mentioned single image dehazing method based on a neural network; the system includes: an input module, a preprocessing module, a backbone module, a postprocessing module, and an output module; wherein, the input module is used to obtain a foggy image of the target environment; the preprocessing module is used to preprocess the foggy image of the target environment and extract features from the foggy image of the target environment during preprocessing to obtain a first feature; the backbone module is used to sequentially extract features from the preprocessed foggy image of the target environment through multiple scale convolutional kernels and an encoder-decoder to obtain a second feature; the postprocessing module is used to perform mapping processing on the first feature and the second feature to obtain a mapping graph; and concatenate the foggy image of the target environment with the mapping graph, and obtain a combined parameter estimation value based on a convolutional layer; the output module is used to input the combined parameter estimation value into a combined parameter model and output a dehazed image.

[0117] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A single image dehazing method based on neural network, characterized in that, comprising: S1. Obtain a hazy image of the target environment; S2. Preprocess the hazy image of the target environment, and extract features from the hazy image of the target environment during preprocessing to obtain a first feature; The S2 specifically includes: Perform scale conversion on the hazy image of the target environment; Divide the hazy image of the target environment after scale conversion into two parts of images with the same size; Extract features from the two parts of images respectively to obtain a first feature; during feature extraction, collect convolution kernels to perform convolution processing on the two parts of images; Merge the two parts of images after convolution processing into a complete image; S3. Sequentially extract features from the preprocessed hazy image of the target environment through multiple scale convolution kernels and an encoder-decoder to obtain a second feature; S4. Perform mapping processing on the first feature and the second feature to obtain a mapping graph; The S4 specifically includes: Obtain a feature map based on the first feature and the second feature; Perform mapping processing on the feature map through a convolutional layer, a pooling layer and an upsampling layer to obtain a mapping graph with a preset size; S5. Concatenate the hazy image of the target environment in S1 with the mapping graph, and obtain a combined parameter estimation value based on a convolutional layer; S6. Input the combined parameter estimation value into a combined parameter model to output a dehazed image.

2. The single image dehazing method based on neural network according to claim 1, characterized in that, During the convolution processing of the two parts of images, the non-linear activation function adopts the PRelu function.

3. The single image dehazing method based on neural network according to claim 1, characterized in that, In the S3, when extracting features from the preprocessed hazy image of the target environment through multiple scale convolution kernels, a pooling layer and a sampling layer are added after each convolutional layer.

4. The single image dehazing method based on neural network according to claim 1, characterized in that, The combined parameter model is expressed as: J(x) = K(x) * I(x) - K(x) + b where x represents an image; I(x) represents the hazy image of the target environment; J(x) represents the dehazed image; K(x) represents the combined parameter estimation value; b represents a fixed bias value.

5. The single image dehazing method based on neural network according to claim 1, characterized in that, The loss function of the combined parameter model is expressed as: Among them, J(x) represents the known dehazed image; represents the dehazed image output by the merged parameter model; represents the estimated value of the merged parameter; I(x) represents the foggy image of the target environment; b represents the fixed bias value.

6. A single image dehazing system based on neural network, characterized in that, Applying the single image dehazing method based on neural network according to any one of claims 1-5; the system includes: an input module, a preprocessing module, a backbone module, a postprocessing module and an output module; The input module is used to obtain a hazy image of the target environment; The preprocessing module is used to preprocess the hazy image of the target environment and extract features from the hazy image of the target environment during preprocessing to obtain a first feature; The preprocessing module specifically includes: performing scale conversion on the hazy image of the target environment; dividing the hazy image of the target environment after scale conversion into two partial images of the same size; respectively performing feature extraction on the two partial images to obtain the first feature; collecting convolution kernels during feature extraction to perform convolution processing on the two partial images; and combining the two partial images after convolution processing into a complete image; The backbone module is used to sequentially perform feature extraction on the preprocessed hazy image of the target environment through multiple-scale convolution kernels and an encoder-decoder to obtain the second feature; The postprocessing module is used to perform mapping processing on the first feature and the second feature to obtain a mapping graph; and concatenating the hazy image of the target environment with the mapping graph, and obtaining an estimated value of the merging parameter based on the convolutional layer; Among them, performing mapping processing on the first feature and the second feature to obtain a mapping graph; specifically includes: obtaining a feature graph based on the first feature and the second feature; performing mapping processing on the feature graph through a convolutional layer, a pooling layer, and an upsampling layer to obtain a mapping graph of a preset size; The output module is used to input the estimated value of the merging parameter into the merging parameter model and output a dehazed image.

Images