Saturation-based fast deep learning image defogging method

Abstract

The invention relates to the field of image processing, and discloses a saturation-based fast deep learning image defogging method, which comprises the following steps of: calculating intensity information and saturation information of an input fog image, performing feature extraction and saturation correction variable estimation by combining a deep learning network, and removing haze in the image through a restoration formula. Specifically, the input fog image and saturation information are spliced and then sent to a deep learning network, the image is processed through a feature extraction partial convolution module (FEPC) and a cross attention module (SCAB), a saturation correction variable is estimated, and then a clear defogged image is recovered by using a recovery formula. According to the method, an efficient and accurate defogging effect can be provided in a complex haze environment, the calculation efficiency is remarkably improved, and the method is suitable for a real-time image defogging task. Experimental results show that the method is superior to similar algorithms in a plurality of evaluation indexes, and particularly has outstanding performance in the aspects of detail recovery, fog removal and calculation speed.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a saturation-based fast deep learning image defogging method. Background Art

[0002] Defogging is a technique used in the field of image processing to improve image quality degraded by atmospheric conditions such as fog, smoke or haze. Under foggy conditions, images will become less contrasty, distorted in color and blurred in detail due to light scattering. Defogging technology aims to restore the clarity and true color of images and improve visual effects. Defogging algorithms are usually based on physical models, simulating the propagation of light in the atmosphere, estimating the impact of fog on images, and making corresponding corrections. Common defogging methods include methods based on physical atmospheric light scattering models and methods based on deep learning. In recent years, the development of deep learning technology has brought new breakthroughs in defogging technology and improved the accuracy and robustness of defogging effects. With the development of technology, defogging technology has been widely used in traffic monitoring, satellite image analysis, photography and other fields.

[0003] These methods can be divided into three categories: the first category is the traditional saturation-based algorithm, the second category focuses on combining the atmospheric light scattering model with deep learning, and the third category uses a neural network model with many parameters and high computational complexity to achieve end-to-end dehazing.

[0004] On the one hand, the traditional saturation-based algorithm [1] is relatively simple, does not require additional scene depth information or atmospheric light information, and only relies on the saturation information of the image itself. However, it is environmentally dependent. In non-uniform fog or special lighting conditions, the change in saturation may not be sufficient to accurately estimate the fog density, and a simple stretching function cannot adapt well to different areas of the entire image. In addition, the effect of the algorithm depends largely on the selection of parameters (such as the stretching function curve), and these parameters often need to be adjusted according to the specific fog conditions.

[0005] On the other hand, for the work combining the atmospheric light scattering model with deep learning, AODNet[2] first deformed the atmospheric light scattering model, merged the two key variables in the atmospheric light scattering model, atmospheric light A and transmission map t(x), into the same variable K(x) and directly used a neural network to estimate K(x). Later, some works such as Light-DehazeNet[3] and LFD-Net[4] continued to use this idea. However, since K(x) contains the two key variables of the atmospheric light scattering model, it is difficult to achieve accurate and ideal results by using a lightweight network to directly estimate K(x).

[0006] Finally, the end-to-end deep learning model with many parameters and large computational workload is often slow to compute, making it difficult to achieve real-time performance in edge device scenarios such as autonomous driving. Therefore, existing methods are usually difficult to apply in scenarios with high real-time requirements.

[0007] Although the traditional dehazing method using saturation in the prior art can improve the image quality to a certain extent, it still has the following significant disadvantages: 1. Insufficient adaptability to fog characteristics, and fog remains in some scenes.

[0008] Fog exhibits different characteristics in different environments and under different conditions, and may vary greatly in color, density, and distribution. Existing defogging algorithms are often designed based on specific fog models and may not be able to adapt to all types of fog characteristics, resulting in the inability to completely remove fog in some scenes. In addition to fog, images may also contain other factors that affect visual perception, such as shadows, lighting changes, etc. These factors may be confused with fog characteristics, making it difficult for the algorithm to distinguish and process them, which in turn affects the defogging effect.

[0009] In addition, it is not appropriate to use the same stretching curve for the saturation of the entire fog image. In [1], only a simple global stretching function is used, such as (in Indicates the saturation of the fog image. represents the estimated saturation of the haze-free image), this method cannot adapt well to the haze characteristics of different areas in the image.

[0010] 2. The difficulty of estimating K(x) for neural networks.

[0011] K(x) contains two key variables of the atmospheric light scattering model - atmospheric light and transmission map, which contains a large amount of information. However, lightweight networks usually mean fewer parameters and lower computational complexity, which limits the network's expressiveness and flexibility, making it difficult to directly and accurately estimate K(x), especially in the face of complex and changeable fog conditions. In addition, existing networks rarely have targeted designs for defogging acceleration, which makes it difficult for existing algorithms to be used in some scenarios with high real-time requirements.

[0012] In summary, it can be seen that the prior art still has significant deficiencies in accuracy and computational efficiency. Therefore, the present invention aims to overcome these problems and proposes an improved defogging method.

[0013] The prior art related to the present invention includes the following references: [1] Se Eun Kim, Tae Hee Park, and ll Kyu Eom, “Fast single imagedehazingusing saturation based transmission map estimation,” IEEETransactionson Image Processing,vol.29,pp.1985-1998, 2020. [2] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng“Aod-net: All-in-one dehazing network,” in 2017 lEEE InternationalConferenceon Computer Vision(ICCV),2017,pp.4780-4788. [3] Hayat Ullah, Khan Muhammad,Muhammad fan,Saeed Anwar,MuhammadSajjad, Ali Sharig Imran, and Victor Hugo C.de Albuquerque, “Light-dehazenet:A novel lightweight cnn architecture forsingle image dehazing,,lEEETransactions on Image Processing, vol.30,pp.8968-8982,2021. [4] Yizhu Jin, Jiaxing Chen, Feng Tian, and Kun Hu, “Lfd-net:Lightweightfeature-interaction dehazing network for real-time remote sensingtasks, IEEE Journal of Selected Topics in Applied Earth ObservationsandRemote Sensing, vol. 16, pp.9139-9153, 2023. [5] Kaiming He, Jian Sun, and Xiaoou Tang, “Single image haze removalusing dark channel prior,” IEEE Transactions on Pattern Analysis and MachineIntelligence, vol. 33, no. 12, pp. 2341–2353, 2011. [6] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao,“Dehazenet: An end-to-end system for single image haze removal,” IEEETransactions on Image Processing, vol. 25, no. 11, pp. 5187– 5198, 2016. [7] Shiyu Zhao, Lin Zhang, Ying Shen, and Yicong Zhou, “Refinednet: Aweakly supervised refinement framework for single image dehazing,” IEEETransactions on Image Processing, vol. 30, pp. 3391–3404, 2021. [8] Zeyuan Chen, Yangchao Wang, Yang Yang, and Dong Liu, “Psd:Principled synthetic-to-real dehazing guided by physical priors,” inProceedings of the IEEE / CVF Conference on Computer Vision and PatternRecognition (CVPR), June 2021, pp. 7180–7189. [9] Jinshan Pan, Jiangxin Dong, Yang Liu, Jiawei Zhang, Jimmy Ren,Jinhui Tang, Yu-Wing Tai, and Ming-Hsuan Yang, “Physics-based generativeadversarial models for image restoration and beyond,” IEEE Transactions onPattern Analysis and Machine Intelligence, vol. 43, no. 7, pp. 2449–2462,2021.

[10] Yudong Liang, Bin Wang, Wangmeng Zuo, Jiaying Liu, and WenqiRen, “Self-supervised learning and adaptation for single image dehazing,” 072022, pp. 1112–1118.

[11] Xudong Wang, Xi’ai Chen, Weihong Ren, Zhi Han, Huijie Fan,Yandong Tang, and Lianqing Liu, “Compensation atmospheric scattering modeland two-branch network for single image dehazing,” IEEE Transactions onEmerging Topics in Computational Intelligence, vol. 8, no. 4, pp. 2880–2896,2024.

[12] Boyi Li, Wengi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, WenjunZeng,and Zhangyang Wang,“Benchmarking single-image dehazing and beyond,,lEEETransactions on Image Processing, vol. 28, no. 1, pp.492-505,2019. Summary of the invention

[0014] In view of the shortcomings of the prior art, the present invention provides a saturation-based fast deep learning image defogging method, which solves the shortcomings of the existing image defogging methods in terms of adaptability to fog features, accuracy of defogging effects, and calculation speed.

[0015] To achieve the above objectives, the present invention is implemented by the following technical scheme: a saturation-based fast deep learning image defogging method, comprising the following steps: Calculate the intensity and saturation information of the input fog image; The fog image and its corresponding saturation information are concatenated as the input of the deep learning network; Saturation correction variables were estimated by deep learning networks; The defogged image is generated through a restoration formula using the saturation correction variable, the atmospheric light value and the fog map.

[0016] Preferably, the step of calculating the intensity information and saturation information of the input fog image comprises: Calculate the intensity information based on the RGB three-channel pixel values ​​of the input fog image; The saturation information is calculated based on the intensity information and the minimum value of the RGB three-channel pixel values.

[0017] Preferably, the deep learning network that inputs the fog image and its saturation information after splicing includes: The initial feature extraction module uses the convolutional layer to extract the initial feature map; Feature extraction partial convolution module, which is used to extract high-frequency features through partial channel convolution and Laplacian convolution and reduce computational complexity; The cross-attention module is used to calculate channel attention and spatial attention to enhance the feature expression of key areas.

[0018] Preferably, the feature extraction part convolution module includes: The input feature map is divided into three parts by channel, one part is subjected to 3×3 convolution, another part is subjected to Laplacian convolution to extract high-frequency information, and the rest is directly copied; The differently processed feature maps are concatenated to form the output feature map.

[0019] Preferably, the feature extraction partial convolution module adopts a dynamic channel segmentation strategy to optimize the feature extraction process.

[0020] Preferably, the cross attention module comprises: The input feature map is divided into a channel attention part and a spatial attention part; Generate a channel attention map through the channel attention mechanism and multiply it with the channel feature; Generate a spatial attention map through the spatial attention mechanism and multiply it with the spatial feature; The feature maps processed by the attention mechanism are concatenated and output.

[0021] Preferably, the cross-attention module is replaced by a multi-head attention mechanism or a self-attention mechanism to capture long-range dependencies in the feature map.

[0022] Preferably, the step of estimating the saturation correction variable by a deep learning network comprises: Estimation of saturation correction variables using deep learning networks , which is defined as:

[0023] in, Input fog map In Location The strength of Input fog map In Location The saturation of For fog-free image In Location The saturation of is estimated by the deep learning network through the training process.

[0024] Preferably, the restoration formula is used to generate a defogging image according to the correction variable and the atmospheric light value, specifically:

[0025] in, is the pixel value of the input fog image, is the saturation correction variable estimated by the deep learning network, It is the atmospheric light value, which is used to simulate the impact of ambient lighting on the image.

[0026] The present invention also provides a device for fast deep learning image defogging based on saturation, comprising: Saturation calculation module, used to calculate the intensity information and saturation information of the input fog image; A deep learning network module, used to receive the fog image and its saturation information, and estimate the saturation correction variable; The image restoration module is used to generate a defogged image through a restoration formula based on the saturation correction variable, the atmospheric light value and the input fog map.

[0027] The present invention provides a saturation-based fast deep learning image defogging method, which has the following beneficial effects: 1. By combining deep learning technology with saturation information, the present invention can adaptively adjust the saturation correction variable according to different haze intensities, thereby improving the accuracy of the defogging effect. This method can restore more image details in complex scenes and avoid excessive defogging or distortion. Through precise saturation correction, the color and brightness of the image are effectively restored, and the visual effect is more natural.

[0028] 2. The present invention significantly improves computational efficiency while maintaining high-quality defogging effects by optimizing the feature extraction partial convolution module (FEPC) and the cross-attention module (SCAB). Compared with traditional defogging methods, the present invention has excellent processing speed and is particularly suitable for real-time application scenarios. This method can achieve fast response in most practical applications and meet high requirements for computing speed and processing efficiency.

[0029] 3. The present invention adopts saturation-based feature extraction and deep learning estimation, which can adaptively process haze of different intensities and types. Through the dynamic adjustment of the network to the image, the method can maintain efficient defogging effect in different haze environments. Whether in outdoor complex scenes or synthetic test data, it shows strong robustness and flexibility.

[0030] 4. The present invention can effectively restore high-frequency details and structural information in images through an innovative defogging framework. This method not only improves the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the image, but also reduces the perceptual similarity (LPIPS) of the image, thereby better restoring the details of the original image. The quality of the defogging image is visually closer to the fog-free scene, enhancing the clarity and realism of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the method flow of the present invention; Figure 2 It is a schematic diagram of the defogging process framework of the present invention; Figure 3 Schematic diagram of the test results of the method of the present invention and other defogging methods on the data set Figure 4 It is a schematic diagram of the device structure of the present invention.

[0032] Among them, 10, saturation calculation module; 20, deep learning network module; 30, image restoration module. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Please refer to the attached Figure 1 -Attached Figure 3 The present invention provides a saturation-based fast deep learning image defogging method, which calculates the intensity information and saturation information of the input fog image, combines the saturation information with the fog image, and finally generates a defogged image through a restoration formula after processing by a deep learning network.

[0035] like Figure 1 As shown, the saturation-based fast deep learning image defogging method may include the following steps: S1, calculating the intensity information and saturation information of the input fog image; S2, concatenate the fog image and its corresponding saturation information as the input of the deep learning network; S3, estimation of saturation correction variables through deep learning networks; S4. Generate a defogged image using a restoration formula using the saturation correction variable, the atmospheric light value, and the fog map.

[0036] Each step of the method of the present invention is described in detail below.

[0037] For step S1, in this embodiment, the input fog image is first processed to calculate its intensity information and saturation information. This step is a crucial basic step in the image defogging process, and provides necessary feature information for subsequent saturation correction and image restoration. The specific technical implementation process is as follows: In the dehazing process, the fog map is input As the image data source, each pixel of the image is composed of the values ​​of its three RGB color channels. , , In order to effectively extract the key information in the fog image, we first need to calculate the intensity and saturation information of the image.

[0038] Intensity information calculation: Strength information Used to describe the brightness of an image, usually expressed as the mean of each pixel in the image in different channels. The calculation is performed using the following formula:

[0039] in, , , Represent the input fog image The red, green, and blue channels in the pixel The color value on .

[0040] This formula represents the average brightness value of each pixel in the image. , which reflects the overall brightness information of the pixel and is the basis for subsequent saturation calculation. Intensity information provides the overall brightness perception of the image and helps to understand the global brightness distribution of the image.

[0041] Saturation information calculation: Next, calculate the saturation information of the input fog image Saturation measures the purity of a color and usually indicates how concentrated the color is in an image. In this step, saturation is calculated using the following formula:

[0042] in, Represents the input fog image The minimum value of the three RGB channels of each pixel in the , indicating the color purity of the pixel; is the intensity value of the pixel, indicating the overall brightness of the pixel.

[0043] This calculation method determines the saturation by the ratio of the minimum channel value to the intensity. , the higher the saturation value, the purer the color of the pixel, and vice versa, the image color is relatively dark. In this way, the color information of the image can be extracted from the input fog image, further supporting the defogging process.

[0044] The intensity information calculated by this step and saturation information It will be used as the input feature of the subsequent deep learning model to provide basic data for the subsequent dehazing process. Specifically, the combination of intensity information and saturation information helps the deep learning network understand the overall brightness and color distribution of the image during the dehazing process, thereby achieving a more accurate dehazing effect.

[0045] Regarding step S2, in this embodiment, step S2 mainly converts the input fog image And its corresponding saturation information The images are concatenated into a new input feature map, and the feature map is passed to the deep learning network. The key to this step is how to effectively combine the color and saturation information of the image to provide the network with richer input features so that the network can learn the deep features in the haze image and perform dehazing.

[0046] In this embodiment, the input image It is an image containing three channels of RGB, and each pixel value can be expressed as , and , corresponding to the pixel values ​​of the red, green and blue channels respectively. Saturation information It is calculated through the aforementioned step S1 and reflects the color purity in the image.

[0047] According to the technical solution of this embodiment, the fog map and saturation map The concatenation is performed in the channel dimension. Specifically, the concatenation operation of the fog map and the saturation map can be performed in the following way:

[0048] in, Represents the concatenated input feature map, which contains the fog map and saturation information. By concatenating the two, the input feature map not only retains the original color information of the image, but also introduces additional saturation features, which helps the deep learning network better understand the color distribution and color purity of the image, thereby improving the dehazing effect.

[0049] The splicing operation combines the RGB three-channel value of each pixel of the fog image with the saturation value corresponding to the pixel to generate a new multi-channel feature map. This splicing operation enables the deep learning network to process the brightness, color and saturation information of the image at the same time, providing richer input data for subsequent feature extraction and dehazing processing.

[0050] In addition, the concatenated input feature map It will be passed to multiple convolutional layers, activation functions, and attention mechanism modules in the deep learning network for further feature learning and processing. In this process, the network will learn how to remove haze from the image and restore clear visual effects based on the local and global features of the image.

[0051] This step is implemented to enable the deep learning network to dynamically adjust the processing method according to the intensity and saturation characteristics of the image under different haze intensities, thereby improving the accuracy and stability of the dehazing effect. By splicing intensity and saturation information, the network can make more accurate judgments in complex image scenes, thereby better restoring the details and colors in the haze image.

[0052] In summary, step S2 provides rich input data for the subsequent deep learning model by splicing the input fog map and saturation information into a new input feature map, and provides necessary image feature support for dehazing processing.

[0053] For step S3, in this embodiment, the main goal of step S3 is to process the input fog image and saturation information through a deep learning network, and then estimate the saturation correction variable , this variable is used in the subsequent dehazing operation.

[0054] In step S2, the input fog image And the corresponding saturation information Concatenated into input feature map , as the input of the deep learning network. The input feature map contains the RGB information of the fog image and the saturation characteristics of the image, which enables the network to process the color distribution and color purity information of the image at the same time.

[0055] Next, the deep learning network learns and estimates the saturation correction variable through a series of feature extraction and attention mechanism modules. .like Figure 2 As shown, the saturation-based dehazing process framework of the present invention is demonstrated. (a) shows the process of sending the fog image into the framework proposed in this article, which mainly includes two traditional algorithm parts and a neural network part. The neural network part mainly includes two feature extraction partial convolution modules and a cross attention module; (b) is the detailed structure of the feature extraction partial convolution module; (c) is the detailed structure of the cross attention module.

[0056] Feature Extraction Partial Convolution Module (FEPC): In the first stage of the network, the input feature map Through the feature extraction partial convolution module (FEPC). The purpose of this module is to achieve efficient feature learning by reducing computational complexity and extracting high-frequency information in the image. The specific structure of the FEPC module is as follows: 1. Feature map segmentation: Input feature map Divided into three parts in the channel dimension .

[0057] 2. Convolution operation: right Using the regular 3×3 convolution operation, we get .

[0058] right Convolution operation using Laplacian feature extraction ,get ,in Convolution for Laplacian feature extraction, are learnable coefficients, and the Laplace convolution kernel is used to extract high-frequency information.

[0059] right Copies it to the output without any modification.

[0060] 3. Splicing output: , and Splicing to get the final feature map .

[0061] Through the above feature extraction process, the network can extract the basic features and detail information of the image, providing rich feature representation for subsequent dehazing processing.

[0062] Cross-Attention Module (SCAB): Based on the output of the feature extraction module, the input feature map It is sent to the Cross Attention Module (SCAB). The core function of this module is to enhance the network's ability to focus on important areas and improve the efficiency of feature extraction through channel and spatial attention mechanisms. The specific structure is as follows: 1. Feature map segmentation: Input feature map It is divided into two parts, namely and .

[0063] 2. Channel Attention Mechanism: right Processing, using Calculate the channel attention map. The specific process is: First, perform an average pooling operation, then pass a 1×1 convolution, a ReLU activation function, another 1×1 convolution, and a Sigmoid activation function to obtain the channel attention map.

[0064] Will After 1×1 convolution and multiplication with the channel attention map, we get .

[0065] 3. Spatial Attention Mechanism: right Processing, using Calculate the spatial attention map. The specific process is: First, perform 1×1 convolution, ReLU activation function, and then perform 1×1 convolution and Sigmoid activation function to obtain the spatial attention map.

[0066] Will After 1×1 convolution and multiplication with the spatial attention map, we get .

[0067] 4. Splicing output: and Splicing to get the final output .

[0068] Through the processing of the cross-attention module, the network can more accurately identify the important areas in the image and significantly improve the computational efficiency. This process helps the network better learn the key features in the image and correct the variables for saturation. The estimates of provide important characteristic support.

[0069] Estimation of the saturation correction variable: After feature extraction and attention mechanism processing, the network generates saturation correction variables , and its calculation formula is as follows:

[0070] in: Input fog map In Location The intensity information on is calculated in step S1; Input fog map In Location The saturation information on is calculated in step S1; For fog-free image In Location The saturation on is estimated by the network through training.

[0071] The correction variable It represents the difference between a foggy image and a haze-free image and is mainly used to correct the saturation in the foggy image to remove haze and restore the clarity of the image.

[0072] In step S3, the deep learning network effectively learns the deep features of the input image through the feature extraction part convolution module and the cross attention module, and estimates the saturation correction variable based on these features. This process not only improves computational efficiency, but also enhances the accuracy of the dehazing effect. By effectively extracting key features and introducing the attention mechanism, this step provides key intermediate variables for subsequent image restoration, thereby ensuring the efficiency and accuracy of the dehazing process.

[0073] Regarding step S4, in this embodiment, step S4 is mainly performed by using the estimated saturation correction variable , atmospheric light value And the input fog map , a clear image after defogging is generated through the restoration formula .

[0074] In the present invention, the core step of the defogging process is to restore the original details and colors of the image using the known fog map, saturation correction variables and atmospheric light values. The restoration formula is as follows:

[0075] in: is the restored dehazed image, which represents the clear image generated by the dehazing process; is the input fog map, which represents the image blurred by haze in the original image; is the saturation correction variable estimated by the deep learning network, which represents the saturation difference between the haze image and the haze-free image; It is the atmospheric light value, which is used to simulate the influence of ambient light. It is usually set to a constant value with a default value of 0.9, or it can be adaptively adjusted according to the actual scene.

[0076] The function of the restoration formula is to correct the input fog image, remove the haze in the image, and restore a clear fog-free image. Specifically, the saturation correction variable Adjusts the brightness and color information of an image based on the intensity and saturation differences of the image. Atmospheric light value Used to simulate the ambient light effect in images, it reflects the propagation effect of atmospheric light in a haze environment.

[0077] In the defogging process, by comparing the fog map and the estimated saturation correction variable , which can effectively remove haze and restore the real details of the image. Atmospheric light value It further helps adjust the brightness and color of the image to make it closer to the real effect in a fog-free environment.

[0078] In this embodiment, the image blur caused by haze can be effectively removed through this restoration method, so that the image can be restored to clarity and retain more details. The defogging method can handle different types of haze conditions, and does not rely on additional depth information or complex physical models. It relies on saturation information and the adaptive ability of deep learning networks to achieve efficient and accurate defogging effects.

[0079] The application of the restoration formula, combined with the correction variables generated by the deep learning network, ensures that the dehazing process not only removes the haze, but also restores the true color and brightness of the image, making the dehazed image more natural and clear. Through this process, the final dehazed image is Not only does it restore clarity visually, but it also effectively avoids over-processing or distortion of the image.

[0080] In summary, step S4 combines the input fog image with the saturation correction variable and the atmospheric light value through the restoration formula, thereby achieving defogging of the fog image, restoring the clarity and real details of the image, and ensuring a significant improvement in image quality.

[0081] In general, the present invention calculates the intensity and saturation information of the input fog image, combines this information with the deep learning network, estimates the saturation correction variable and removes the haze in the image through the restoration formula. The method first calculates the intensity and saturation information of the image, then extracts and estimates the features through the deep learning network, and finally uses the defogging formula to restore the clarity of the image. By introducing deep learning technology and saturation information, the present invention can effectively remove haze and restore image details, and has high processing efficiency. It is suitable for complex haze environments, especially in the fields of autonomous driving and monitoring. It has broad application prospects.

[0082] In some embodiments, different attention mechanisms can be used as alternatives for the cross-attention module (SCAB) in the present invention. Specifically, in addition to channel attention and spatial attention, self-attention or multi-head attention mechanisms can also be introduced. These mechanisms can capture more complex long-distance dependencies in feature maps and provide richer contextual information for the model. Compared with traditional attention mechanisms, self-attention and multi-head attention have enhanced context capture capabilities, and are particularly suitable for processing data with long-distance dependencies, thereby improving the performance and flexibility of the model. In addition, these mechanisms can be flexibly adjusted according to the different requirements of the task, adapt to different data features, and provide stronger feature representation capabilities.

[0083] In some embodiments, the selection and configuration of the attention mechanism can be customized according to specific needs, for example, by adjusting parameters such as the number of heads and scaling factors to optimize network performance and meet the needs of different tasks. At the same time, as a modular design, these attention mechanisms can be easily integrated into the existing cross-attention framework to ensure good interaction with other modules. In addition, by introducing self-attention or multi-head attention mechanisms, the performance of the model in complex data processing can be significantly improved.

[0084] In some embodiments, the feature map segmentation strategy can dynamically adjust the segmentation method according to the different characteristics of the input feature map to optimize the network's computing resource usage and improve the adaptability of the model. Specifically, the feature map can not only be evenly divided, but also a dynamic segmentation method based on specific rules such as frequency or importance can be used to achieve more efficient feature extraction. The dynamic segmentation strategy helps to better capture the key information in the feature map, reduce redundant calculations, and improve resource utilization efficiency.

[0085] In some embodiments, the adjustment mechanism of the dynamic segmentation strategy can be flexibly changed according to the different features of the input image, further improving the computational efficiency and model adaptability. Through intelligent segmentation, irrelevant parts of the feature map can be removed, thereby reducing computational overhead and improving overall performance.

[0086] In some embodiments, the choice of pooling strategy can also replace global average pooling. In addition to traditional global average pooling, pooling strategies such as maximum pooling and adaptive pooling can be considered. These pooling strategies can capture different information according to different features of the image and enhance the feature expression ability of the model. Specifically, maximum pooling can extract the most significant features, and adaptive pooling can dynamically adjust the size of the pooling window according to the needs of specific tasks, further improving the flexibility and efficiency of the pooling strategy.

[0087] In some embodiments, the selection of pooling strategies and integration methods are optimized according to the needs of specific tasks to ensure that the pooling operation improves model performance. Different pooling strategies may bring significant performance improvements under different tasks and data characteristics. Therefore, using different pooling strategies can provide more diverse feature expressions for tasks, while improving the effect of the dehazing model in specific tasks.

[0088] Through these alternatives, the present invention can further improve the adaptability and computational efficiency of the dehazing model, while enhancing the performance of the model under different data characteristics and task requirements.

[0089] In order to verify the effectiveness and superiority of the method of the present invention, the present invention has conducted comparative tests on experimental results on multiple data sets. In the experiment, the present invention has compared the performance with other mainstream dehazing methods (such as DCP method, AODNet method, DehazeNet method, etc.), covering three main evaluation indicators: peak signal-to-noise ratio (PSNR), structural similarity (SSIM), image perceptual similarity (LPIPS) and calculation speed (average time). The experimental results show that the performance of the present invention in restoring details and removing haze is significantly better than that of similar algorithms.

[0090] Dataset and testing method The present invention uses two public datasets for testing: the SOTS (Outdoor) dataset and the HSTS (Synthetic) dataset

[12] . These two datasets have different scene and data features and are used to verify the dehazing effect of the present invention method in real environments (outdoor scenes) and synthetic images (artificially generated images).

[0091] During the test, different dehazing methods were evaluated, and the following evaluation indicators were used: PSNR (Peak Signal-to-Noise Ratio): It is used to measure the quality of the dehazed image. The higher the PSNR, the better the image quality.

[0092] SSIM (structural similarity): used to measure the similarity of image structures. The closer the SSIM value is to 1, the better the dehazing effect is.

[0093] LPIPS (Perceptual Image Similarity): This metric is used to measure the perceptual similarity of images. The lower the LPIPS value, the closer the dehazed image is to the real image in terms of perception.

[0094] Computational speed: measures the processing speed of each method, expressed in milliseconds per image.

[0095] The test results are shown in the following table and Figure 3 shown.

[0096] As can be seen from the table above, in the test results on the SOTS and HSTS datasets, the method of the present invention performs well in the three indicators of PSNR, SSIM and LPIPS. In particular, in terms of PSNR and SSIM, the method of the present invention reaches 25.5383 and 0.9099 on the SOTS dataset, and 27.2990 and 0.9539 on the HSTS dataset, which are higher than other similar methods, showing higher image quality and structural similarity.

[0097] In terms of image perceptual similarity (LPIPS), the performance of the method of the present invention is also significantly better than other methods, which are 0.0613 and 0.0286 on the SOTS dataset and HSTS dataset, respectively, which are much lower than the values ​​of other methods, indicating that the effect of the present invention in restoring details and removing haze is closer to the real haze-free image.

[0098] In addition, the method of the present invention performs outstandingly in terms of computing speed, especially in terms of average time, which is 4.329ms on the SOTS dataset and 4.272ms on the HSTS dataset, significantly better than other methods. This shows that the method of the present invention can not only provide high-quality defogging effects, but also play an important role in application scenarios with high real-time requirements (such as autonomous driving, monitoring, etc.).

[0099] from Figure 3 It can be seen that the performance of the present invention on some data sets is better than that of similar DCP methods, AODNet methods, etc., and is better than similar algorithms in terms of recovering details and removing fog.

[0100] By comparing the test results, it can be seen that the method of the present invention is superior to similar defogging algorithms in multiple evaluation indicators, especially in restoring image details, removing fog and computing speed. This performance improvement is due to the deep learning technology and feature extraction strategy introduced in the present invention, which enables the method to provide more accurate and efficient defogging effects in complex haze environments.

[0101] The device for fast deep learning image defogging based on saturation described below and the method for fast deep learning image defogging based on saturation described above can refer to each other.

[0102] Please refer to the attached Figure 4 The present invention also provides a device for fast deep learning image defogging based on saturation, comprising: The saturation calculation module 10 is used to calculate the intensity information and saturation information of the input fog image; A deep learning network module 20 is used to receive the fog image and its saturation information and estimate the saturation correction variable; The image restoration module 30 is used to generate a defogged image through a restoration formula according to the saturation correction variable, the atmospheric light value and the input fog map.

[0103] The device of this embodiment can be used to execute the above method embodiment, and its principles and technical effects are similar, which will not be repeated here.

[0104] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fast deep learning image defogging method based on saturation, characterized in that: The following steps are involved: Calculate the intensity and saturation information of the input fog image; The fog image and its corresponding saturation information are concatenated as the input of the deep learning network; Saturation correction variables were estimated by deep learning networks; The defogged image is generated through a restoration formula using the saturation correction variable, the atmospheric light value and the fog map.

2. The saturation-based fast deep learning image defogging method according to claim 1, characterized in that: The step of calculating the intensity information and saturation information of the input fog image comprises: Calculate the intensity information based on the RGB three-channel pixel values ​​of the input fog image; The saturation information is calculated based on the intensity information and the minimum value of the RGB three-channel pixel values.

3. The saturation-based fast deep learning image defogging method according to claim 1, characterized in that: The deep learning network that inputs the fog image and its saturation information after splicing includes: The initial feature extraction module uses the convolutional layer to extract the initial feature map; Feature extraction partial convolution module, which is used to extract high-frequency features through partial channel convolution and Laplacian convolution and reduce computational complexity; The cross-attention module is used to calculate channel attention and spatial attention to enhance the feature expression of key areas.

4. The saturation-based fast deep learning image defogging method according to claim 3, characterized in that: The feature extraction part convolution module includes: The input feature map is divided into three parts by channel, one part is subjected to 3×3 convolution, another part is subjected to Laplacian convolution to extract high-frequency information, and the rest is directly copied; The differently processed feature maps are concatenated to form the output feature map.

5. The saturation-based fast deep learning image defogging method according to claim 4, characterized in that: The feature extraction partial convolution module adopts a dynamic channel segmentation strategy to optimize the feature extraction process.

6. The saturation-based fast deep learning image defogging method according to claim 3, characterized in that: The cross-attention module includes: The input feature map is divided into a channel attention part and a spatial attention part; Generate a channel attention map through the channel attention mechanism and multiply it with the channel feature; Generate a spatial attention map through the spatial attention mechanism and multiply it with the spatial feature; The feature maps processed by the attention mechanism are concatenated and output.

7. The saturation-based fast deep learning image defogging method according to claim 6, characterized in that: The cross-attention module is replaced by a multi-head attention mechanism or a self-attention mechanism to capture long-range dependencies in feature maps.

8. The saturation-based fast deep learning image defogging method according to claim 1, characterized in that: The step of estimating the saturation correction variable by the deep learning network comprises: Estimation of saturation correction variables using deep learning networks , which is defined as: in, Input fog map In Location The strength of Input fog map In Location The saturation of For fog-free image In Location The saturation of is estimated by the deep learning network through the training process.

9. The saturation-based fast deep learning image defogging method according to claim 1, characterized in that: The restoration formula is used to generate a defogging image according to the correction variable and the atmospheric light value, specifically: in, is the pixel value of the input fog image, is the saturation correction variable estimated by the deep learning network, It is the atmospheric light value, which is used to simulate the impact of ambient lighting on the image.

10. A device for fast deep learning image defogging based on saturation, used to execute the fast deep learning image defogging method based on saturation according to any one of claims 1 to 9, characterized in that: include: Saturation calculation module, used to calculate the intensity information and saturation information of the input fog image; A deep learning network module, used to receive the fog image and its saturation information, and estimate the saturation correction variable; The image restoration module is used to generate a defogged image through a restoration formula based on the saturation correction variable, the atmospheric light value and the input fog map.

Images