A hierarchical image rain removal method based on enhanced rain streak perception

Abstract

The application discloses a hierarchical image rain removal method based on enhanced rain streak perception. In view of the image quality degradation problem in rainy environment, an enhanced rain streak perception feature enhancement module E-RAFEM is designed, which integrates a learnable direction filter and a rain streak texture modeling two core components, and accurately models the directionality and linear texture features of rain streaks. A Transformer-based encoder-decoder backbone network is used, and the E-RAFEM module is embedded in the first three encoding levels to form a hierarchical rain streak processing mechanism. An incremental multi-scale fusion mechanism is introduced, multi-scale features are captured through different expansion rate convolution kernels, and an incremental method is used to gradually integrate to avoid information loss. The enhanced rain streak perception network ERA-Net proposed by the application realizes high-quality recovery of images under various complex rainy conditions through accurate rain streak feature modeling and hierarchical processing strategy.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, specifically to a hierarchical image deraining method based on deep learning to enhance rain pattern perception, which is particularly suitable for processing image deraining problems under various complex rainy conditions. Background Technology

[0002] Image rain descratching, as an important branch of low-level vision tasks, plays a significant role in practical applications such as autonomous driving, video surveillance, and intelligent transportation. The presence of rain patterns not only affects the visual quality of images but also severely degrades the performance of high-level vision tasks such as object detection and semantic segmentation. With the rapid development of deep learning technology, numerous rain descratching methods based on CNNs and Transformers have emerged, but technical challenges remain when dealing with complex rain pattern patterns.

[0003] Early rain pattern removal methods mainly relied on traditional image processing techniques, such as morphological operations and frequency domain filtering, to remove rain patterns. These methods are often based on specific prior assumptions and perform reasonably well in simple scenes, but they are difficult to handle the complex and varied rain pattern patterns in real-world environments.

[0004] In recent years, researchers have explored solutions to the rain pattern removal problem from different perspectives. Early deep learning methods mainly performed rain pattern detection and removal simultaneously through multi-task learning; subsequent works utilized high-frequency detail information for rain pattern removal. Recent research has proposed an efficient rain pattern removal architecture based on Transformer, demonstrating the advantages of self-attention mechanisms in long-range dependency modeling. Other works have explored the application of the Mamba architecture in image restoration, achieving excellent rain pattern removal performance by combining Fourier transform and a state-space model.

[0005] However, existing methods still have the following shortcomings in rain strife feature modeling:

[0006] (1) Insufficient directional modeling: The angle and direction of rain streaks vary in real scenes. Traditional fixed-direction filters cannot effectively adapt to this variety of changes, resulting in poor processing of oblique rain streaks.

[0007] (2) Inaccurate texture feature extraction: Rain patterns have obvious linear texture features, but existing methods lack a special design for the linear texture features of rain patterns and often use general feature extraction methods, which cannot accurately capture the unique texture pattern of rain patterns.

[0008] (3) Imperfect multi-scale information fusion: Rain patterns in images present different scale features, from small raindrops to large rain lines. Existing methods are prone to information loss during multi-scale feature fusion, which affects the final restoration effect.

[0009] (4) Lack of hierarchical processing strategy: Existing methods usually adopt single-layer or simple feature processing strategies, lacking a hierarchical rain pattern perception mechanism from details to the whole. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a hierarchical image deraining method based on enhanced rain pattern perception. This method can accurately model the directionality and texture features of rain patterns. Through hierarchical processing strategies and progressive multi-scale fusion mechanisms, it can achieve high-quality restoration of images under various complex rainy conditions, and has good rain pattern removal effect and generalization ability.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a hierarchical image deraining method based on enhanced rain pattern perception, comprising the following steps:

[0012] Collect images containing rain streaks, perform preprocessing and data augmentation, and create a dataset of rain streak degraded images.

[0013] An enhanced rain-slip perception image restoration network model, ERA-Net, is established, comprising: constructing an enhanced rain-slip perception feature enhancement module (E-RAFEM) and embedding it into a Transformer-based encoder-decoder backbone network; the enhanced rain-slip perception feature enhancement module (E-RAFEM) includes a rain-slip perception feature extraction unit and a multi-scale adaptive enhancement unit. The rain-slip perception feature extraction unit is used to adaptively extract and model the directional and linear texture features of raindrops in rain-containing images, accurately capturing different raindrop patterns; the multi-scale adaptive enhancement unit is used for progressive multi-scale fusion of extracted features, enhancing raindrop texture information and avoiding information loss during feature fusion; the encoder-decoder backbone network adopts a four-level hierarchical encoder structure, embedding the E-RAFEM module in the first three encoding levels to form a hierarchical rain-slip processing mechanism. The encoder extracts the latent features of the input rain-slip image I_rain through a self-attention mechanism, and the decoder outputs a clear image I_clean using a symmetrical upsampling structure; the network is trained using a dataset, and the parameters are adjusted by backpropagation using a loss function to obtain the ideal ERA-Net model;

[0014] The system acquires rain-textured images and preprocesses them before inputting them into the ERA-Net ideal model, which automatically outputs the corresponding clear restored images.

[0015] The preprocessing includes, but is not limited to, random cropping, flipping, rotation, brightness and contrast adjustment, and size normalization; the rain pattern degradation image pairs are complex background images of rainy days in multiple scenes and their corresponding clear images.

[0016] The rain pattern perception feature extraction unit integrates two core components: a learnable directional filter and rain pattern texture modeling. The learnable directional filter adaptively captures the rain pattern directional feature F_dir through a learnable multi-directional filter bank composed of multiple independent convolutional layers. The rain pattern texture modeling uses an asymmetric convolutional kernel design to model the rain pattern linear texture feature F_texture.

[0017] The multi-scale adaptive enhancement unit adopts a progressive multi-scale fusion mechanism, which captures multi-scale features through convolutional kernels with different dilation rates, integrates them step by step in a progressive manner to avoid information loss, and introduces a gating mechanism to control the residual connection strength.

[0018] The learnable direction filter includes:

[0019] (1) Multi-directional filter bank: It consists of 8 independent convolutional layers, each of which extracts the input image X at a set angle. Directional features F i =Conv i (X), i = 1, 2, ..., 8,

[0020] (2) Orientation weight learning: The importance of each directional feature is dynamically adjusted through the learned orientation weight network, and the weight W is calculated;

[0021] W = Softmax(Conv) weight (X));

[0022] Where Softmax represents the function, Conv weight This represents the convolutional layer used to calculate the weights;

[0023] (3) Feature fusion: The weighted multi-directional features are integrated to output the directional enhancement feature F. dir ;

[0024]

[0025] Among them, Conv fusion It is a convolutional layer used to effectively fuse features from different directions; Concat represents the concatenation operation.

[0026] The rain pattern texture modeling includes:

[0027] (1) Asymmetric convolution design: 7×1 and 1×7 convolution kernels are used to extract linear textures in the horizontal and vertical directions, respectively; F linear =Conv 7×1 (X),F vertical =Conv 1×7 (X)

[0028] (2) Texture enhancement processing: The extracted linear features are concatenated and then enhanced through the TextureEnhance network;

[0029] F texture =TextureEnhance(Concat([F linear F vertical ]))

[0030] (3) Texture attention mechanism: Attention weights A are calculated using global average pooling (GAP) and multilayer perceptron (MLP). texture Output the final texture enhancement feature F out ;

[0031] A texture =Sigmoid(MLP(GAP(F) texture )))

[0032] F out =X⊙(1+A) texture ⊙F texture )

[0033] Here, Sigmoid represents a function.

[0034] The progressive multi-scale fusion mechanism includes:

[0035] For the obtained F out and F dir Preliminary fusion yields F fusion The formula is F fusion =Concat(F dir F out );

[0036] The extracted feature maps are enhanced, optimized, and fused at multiple scales through a feature refinement module, as shown below:

[0037] F refined =Refinement(MultiScaleFusion(F fusion ))

[0038] F final =X+Gate(F refined )⊙F refined

[0039] MultiScaleFusion captures multi-scale features using convolutional kernels with different dilation rates, and Gate is a gate control mechanism to control the strength of residual connections; Refinement is specifically Conv2(BatchNorm(ReLU(Conv1(X)))), where Conv1 and Conv2 are defined as two convolutional layers, Frefined This represents the output feature map obtained after refinement. Gate represents the weight output of the gating mechanism, used to dynamically adjust the residual connection strength of the feature map, specifically σ(Conv). g (F refined Convg: Convolutional layer used to generate gated weights; ⊙ represents element-wise multiplication (dot product), which means applying weights to each feature element; F final This represents the final feature map after gating is applied.

[0040] The hierarchical rain pattern perception strategy includes:

[0041] (1) Encoder hierarchical structure: A four-level hierarchical structure is adopted, with feature dimensions of 48, 96, 192 and 384 respectively. The first three coding levels embed E-RAFEM modules respectively.

[0042] (2) Processing from details to the whole: forming a hierarchical processing mechanism of "details → local → global", with the first level processing detailed features, the second level processing local features, and the third level processing global features;

[0043] (3) Latent space protection: Level 4 is used as a latent space and does not use E-RAFEM to avoid loss of semantic information due to overprocessing.

[0044] The loss function used is L1 loss, which is employed for network training optimization.

[0045] The method is applicable to the following types of rain streak degradation: light rain streak, moderate rain streak, heavy rain streak, rain lines at different angles, and dense rain curtain scenes.

[0046] The present invention has the following beneficial effects and advantages:

[0047] 1. Precise Rain Vein Feature Modeling: This invention achieves comprehensive modeling of rain vein directionality and texture features through two core components of the E-RAFEM module. The learnable direction filter, compared to a fixed direction filter, has stronger adaptability and can dynamically adjust the feature extraction strategy according to the actual directional distribution of rain veins in different scenes, effectively solving the problem of poor handling of oblique rain vein lines by traditional methods.

[0048] 2. Specialized texture modeling design: By using asymmetric convolution kernel design to specifically capture the linear features of rain patterns, and combining it with a texture attention mechanism to adaptively adjust feature weights, it has significant advantages over traditional general feature extraction methods in linear texture modeling of rain patterns, and can more accurately identify and process the unique texture patterns of rain patterns.

[0049] 3. Hierarchical processing strategy: The hierarchical rain pattern perception strategy of "details → local → global" is adopted. E-RAFEM modules are embedded in different levels of the encoder, which makes full use of the feature advantages of different levels of the encoder and forms a complete processing system from local details to global semantics, which is superior to single-layer feature processing methods.

[0050] 4. Progressive multi-scale fusion: By progressively integrating features at different scales, the problem of information loss that may be caused by direct fusion is effectively avoided. At the same time, a gating mechanism is introduced to control the residual connection strength, which significantly improves the image restoration quality while maintaining computational efficiency.

[0051] 5. Excellent performance: Experimental results show that ERA-Net achieves stable performance improvements in PSNR and SSIM metrics compared to the best existing methods on multiple benchmark datasets. Specifically, on the Rain100H dataset, PSNR and SSIM metrics are improved by 2.3dB and 0.045, respectively, which fully verifies the effectiveness of the method.

[0052] 6. Excellent generalization ability: Through precise feature modeling and hierarchical processing strategies, ERA-Net demonstrates excellent generalization ability in rain ripple scenarios of different types and intensities, including various complex scenarios such as light rain ripples, heavy rain ripples, and rain lines at different angles. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall architecture of ERA-Net according to the method of the present invention.

[0054] Figure 2 This is a visual comparison of the ERA-Net of the present invention with other advanced methods. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] I. Constructing a Network Model

[0057] Figure 1 This is a schematic diagram of the overall architecture of ERA-Net in this invention. The hierarchical image deraining method for enhanced rain ripple perception of this invention includes the following steps:

[0058] Step 1: Construct a Transformer-based encoder-decoder backbone network

[0059] A four-level hierarchical encoder structure is adopted, with feature dimensions of 48, 96, 192, and 384 respectively. The encoder consists of multiple Transformer blocks, each containing a multi-head self-attention layer and a feedforward network. It utilizes the self-attention mechanism to capture long-range dependencies and extract latent features from the input rainy image. The decoder adopts a symmetrical upsampling structure, fusing encoder features through skip connections to progressively restore image resolution and output a clear image.

[0060] Step 2: Constructing an Enhanced Rain Stroke Perception Feature Enhancement Module (E-RAFEM)

[0061] The E-RAFEM module integrates two core components: a learnable directional filter and rain ripple texture modeling, specifically designed for the directional and linear texture features of rain ripples. This module is embedded in the first three levels of the encoder, forming a hierarchical rain ripple processing mechanism.

[0062] Step 3: Design a hierarchical integration strategy

[0063] E-RAFEM modules are embedded in the encoder's first, second, and third levels to form a hierarchical processing mechanism of "detail → local → global". The first level is dedicated to processing detailed features, such as small-scale rain streaks; the second level processes local features, such as medium-scale rain lines; and the third level processes global features, such as large-scale rain curtains. The fourth level, as the latent space, does not use E-RAFEM to avoid semantic information loss due to overprocessing.

[0064] Step 4: Construct a progressive multi-scale fusion mechanism

[0065] Multi-scale features are captured by using convolutional kernels with different dilation rates, and integrated gradually in a progressive manner to effectively avoid information loss that may result from direct fusion. A gating mechanism is introduced to control the strength of residual connections, enabling the network to adaptively determine how much original feature information to retain.

[0066] Step 5: Multi-stage output optimization

[0067] The network produces results in two stages: coarse output and fine output. Through a progressive optimization strategy, preliminary de-wiring is performed in the coarse stage, and then detailed optimization is performed in the fine stage to improve the final de-wiring effect.

[0068] II. Detailed Design of Core Modules

[0069] Module 1: Design of Learnable Directional Filters

[0070] This component consists of eight independent convolutional layers, each dedicated to feature extraction in a specific direction. For the input features...

[0071]

[0072] Parallel extraction of directional features by filters in each direction:

[0073] F i =Conv i (X), i = 1, 2, ..., 8

[0074] Each convolutional kernel is initialized to the corresponding angle. A linear filter is used, and the importance of features in each direction is dynamically adjusted through a learned directional weight network:

[0075] W = Softmax(Conv) weigh (X))

[0076]

[0077] Where Softmax represents the function, Conv weight This refers to the convolutional layer used to calculate the weights, Conv fusion It is a convolutional layer used to effectively fuse features from different directions.

[0078] This design allows the network to adaptively focus on the most important rain ripple directions in the current scene, resulting in stronger expressive power compared to fixed-direction filters.

[0079] Module Two: Rain Pattern Texture Modeling and Design

[0080] This method employs an asymmetric convolution kernel design to specifically capture the linear features of rain patterns. Rain patterns primarily exhibit linear textures in images; therefore, 7×1 and 1×7 convolution kernels are used. 7×1 Conv 1×7 Extract the linear textures in the horizontal and vertical directions separately:

[0081] The formula is as follows:

[0082] F linear =Conv 7×1 (X)

[0083] F vertical =Conv 1×7 (X)

[0084] The extracted linear features are concatenated and then enhanced using the TextureEnhance network to strengthen their representation.

[0085] F texture=TextureEnhance(Concat([F linear F vertical A texture attention mechanism is introduced to adaptively adjust feature weights, enabling the network to focus on rain-texture regions. The attention weight A is calculated. texture Output the final texture enhancement feature F out A texture =σ(MLP(GAP(F) texture )))

[0086] F out =X⊙(1+A) texture ⊙F texture )

[0087] Wherein, GAP represents global average pooling, and MLP represents multilayer perceptron;

[0088] Module 3: Progressive Multi-Scale Fusion Mechanism

[0089] For the obtained F out and F dir Preliminary fusion yields F fusion The formula is F fusion =Concat(F dir F out Then, the extracted feature maps are enhanced and optimized through the feature refinement module, and a gating mechanism is introduced to control the residual connection strength.

[0090] F refined =Refinement(MultiScaleFusion(F fusion ))

[0091] F final =X+Gate(F refined )⊙F refined

[0092] MultiScaleFusion captures multi-scale features using convolutional kernels with different dilation rates, and Gate is a gate control mechanism to control the strength of residual connections; Refinement is specifically Conv2(BatchNorm(ReLU(Conv1(X)))), where Conv1 and Conv2 are defined as two convolutional layers, F refined This represents the output feature map obtained after refinement. Gate represents the weight output of the gating mechanism, used to dynamically adjust the residual connection strength of the feature map, specifically σ(Conv). g (F refined Convg: Convolutional layer used to generate gated weights; ⊙ represents element-wise multiplication (dot product), which means applying weights to each feature element; Ffinal This represents the final feature map after gating is applied.

[0093] III. Loss Function Design

[0094] The L1 loss method is used to calculate the pixel-level differences between the restored image and the real image.

[0095]

[0096] Where N is the total number of pixels, and I_clean and I_gt represent the restored image and the original image, respectively.

[0097] IV. Training Implementation Details

[0098] 1. Data preprocessing:

[0099] This invention performs the following preprocessing operations on the original rainy day images to standardize their size and expand the dataset:

[0100] Randomly crop to 256×256 size; randomly flip and rotate horizontally (±15 degrees); randomly adjust brightness / contrast (±0.2); normalize to scale pixel values ​​to the range [0,1].

[0101] 2. Training strategy:

[0102] The Adam optimizer (β1 = 0.9, β2 = 0.999) was used, with an initial learning rate of 2e-4. A cosine learning rate scheduling strategy was employed, decaying to 0.5 times the original rate every 50 epochs. The batch size was 8, the image patch size was 256×256, and training lasted for 200 epochs. Gradient clipping (maximum norm 1.0) was applied to prevent gradient explosion, and mixed precision training was used to accelerate the training process.

[0103] 3. Dataset configuration:

[0104] Training datasets: Rain100H, containing 1800 pairs of training images; Rain100L, containing 200 pairs of training images;

[0105] Test datasets: Test sets for Rain100H and Rain100L, each containing 100 pairs of images, used for performance evaluation;

[0106] Real-world scenario testing: Collected real rainy day scene images to verify the actual application effect of the method.

[0107] V. Performance Verification and Analysis

[0108] 1. Quantitative comparative analysis

[0109] As shown in Table 1, ERA-Net achieved the best performance on all test datasets. A detailed comparison with existing state-of-the-art methods is shown in Table 1 below:

[0110] Table 1

[0111]

[0112]

[0113] On the Rain100H dataset: ERA-Net achieves a PSNR of 31.21 dB and an SSIM of 0.9103, which are 0.27 dB PSNR and 0.007 SSIM improvements respectively compared to the suboptimal method DRSformer;

[0114] On the Rain100L dataset: ERA-Net achieves a PSNR of 37.94dB and an SSIM of 0.9804, which are 0.46dB higher PSNR and 0.002 higher SSIM than the suboptimal method DRSformer.

[0115] 2. Qualitative comparative analysis

[0116] like Figure 2 As shown, the visual comparison results between ERA-Net and other state-of-the-art methods demonstrate that:

[0117] ERA-Net can remove rain streaks more thoroughly while better preserving background details and texture information;

[0118] In complex texture areas, ERA-Net avoids over-smoothing and maintains better detail clarity;

[0119] In dense rain pattern scenarios, ERA-Net demonstrates stronger processing capabilities, effectively handling overlapping and intersecting rain lines.

[0120] VI. Practical Application Scenarios

[0121] The ERA-Net method of this invention can be widely applied to the following scenarios:

[0122] Intelligent Transportation Systems: Improve the image quality of traffic monitoring cameras in rainy conditions, and enhance the accuracy of vehicle recognition and behavior analysis;

[0123] Autonomous driving: Providing clear visual perception for autonomous vehicles, improving safety and reliability when driving in rainy weather;

[0124] Security monitoring: Improve the image quality of the monitoring system in rainy conditions, and enhance the effectiveness of target detection and recognition;

[0125] Mobile devices: Integrated into smartphones and other mobile devices to improve the quality of photos taken in the rain.

[0126] The above descriptions are merely specific embodiments of this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A hierarchical image deraining method based on enhanced rain pattern perception, characterized in that, Includes the following steps: Collect images containing rain streaks, perform preprocessing and data augmentation, and create a dataset of rain streak degraded images. An enhanced rain-slip perception image restoration network model, ERA-Net, is established, comprising: constructing an enhanced rain-slip perception feature enhancement module (E-RAFEM) and embedding it into a Transformer-based encoder-decoder backbone network; the enhanced rain-slip perception feature enhancement module (E-RAFEM) includes a rain-slip perception feature extraction unit and a multi-scale adaptive enhancement unit; the rain-slip perception feature extraction unit is used to adaptively extract and model the directional and linear texture features of raindrops in rain-containing images, accurately capturing different raindrop patterns; the multi-scale adaptive enhancement unit is used for progressive multi-scale fusion of extracted features, enhancing raindrop texture information and avoiding information loss during feature fusion; the encoder-decoder backbone network adopts a four-level hierarchical encoder structure, embedding the E-RAFEM module in the first three encoding levels to form a hierarchical rain-slip processing mechanism; the encoder extracts the latent features of the input rain-slip image I_rain through a self-attention mechanism, and the decoder outputs a clear image I_clean using a symmetrical upsampling structure; the network is trained using a dataset, and the parameters are adjusted by backpropagation using a loss function to obtain the ideal ERA-Net model; The system acquires rain-textured images to be processed, preprocesses the data, and inputs them into the ERA-Net ideal model, which automatically outputs the corresponding clear restored images. The rain pattern perception feature extraction unit integrates two core components: a learnable directional filter and rain pattern texture modeling. The learnable directional filter adaptively captures rain pattern directional features through a learnable multi-directional filter bank composed of multiple independent convolutional layers. The rain pattern texture modeling employs an asymmetric convolution kernel design to model the linear texture features of the rain pattern. ; The multi-scale adaptive enhancement unit adopts a progressive multi-scale fusion mechanism, which captures multi-scale features through convolutional kernels with different dilation rates, integrates them step by step in a progressive manner to avoid information loss, and introduces a gating mechanism to control the residual connection strength. The rain pattern texture modeling includes: (1) Asymmetric convolution design: 7×1 and 1×7 convolution kernels are used to extract linear textures in the horizontal and vertical directions, respectively; ; in, Indicates the input image; (2) Texture enhancement processing: The extracted linear features are concatenated and then processed through a texture enhancement network. Strengthen the expression; ； (3) Texture attention mechanism: through global average pooling and multilayer perceptron Calculate attention weights Output the final texture enhancement features ; ； ； in, Represents a function; The progressive multi-scale fusion mechanism includes: The obtained and Preliminary fusion was achieved The formula is ; The extracted feature maps are enhanced, optimized, and fused at multiple scales through a feature refinement module, as shown below: ； Among them, MultiScaleFusion captures multi-scale features through convolutional kernels with different dilation rates; Specifically Conv1 and Conv2 are defined as two convolutional layers, respectively. This represents the output feature map obtained after refinement. The weight output of the gating mechanism is used to dynamically adjust the residual connection strength of the feature map. Convg: A convolutional layer used to generate gated weights. This represents the element-wise multiplication of the dot product, which means applying weights to each feature element. This represents the final feature map after gating is applied.

2. The hierarchical image deraining method based on enhanced rain pattern perception according to claim 1, characterized in that, The preprocessing includes random cropping, flipping, rotation, brightness and contrast adjustment, and size normalization; The rain streak degraded image pairs consist of complex background images of rainy days in multiple scenes and their corresponding clear images.

3. The hierarchical image deraining method based on enhanced rain pattern perception according to claim 1, characterized in that, The learnable direction filter includes: (1) Multidirectional filter bank: It consists of 8 independent convolutional layers, each of which extracts the input image. Set angle directional features , ; (2) Directional weight learning: The importance of features in each direction is dynamically adjusted through the learned directional weight network, and the weights are calculated. ; ； in, Represents a function, This represents the convolutional layer used to calculate the weights; (3) Feature fusion: The weighted multi-directional features are integrated to output directional enhancement features. ; ； in, It is a convolutional layer used to effectively fuse features from different directions. This indicates a splicing operation.

4. The hierarchical image deraining method based on enhanced rain pattern perception according to claim 1, characterized in that, The hierarchical rain pattern processing mechanism includes: (1) Encoder hierarchical structure: A four-level hierarchical structure is adopted, with feature dimensions of 48, 96, 192 and 384 respectively. The first three coding levels are embedded with E-RAFEM modules respectively; (2) Processing from details to the whole: forming a hierarchical processing mechanism of "details → local → global", with the first level processing detailed features, the second level processing local features, and the third level processing global features; (3) Latent space protection: Level 4 is used as a latent space and does not use E-RAFEM to avoid loss of semantic information due to overprocessing.

5. A hierarchical image deraining method based on enhanced rain pattern perception according to claim 1, characterized in that, The loss function used is L1 loss, which is used for network training optimization.

6. A hierarchical image deraining method based on enhanced rain pattern perception according to any one of claims 1-5, characterized in that, The method is applicable to the following types of rain streak degradation: light rain streak, moderate rain streak, heavy rain streak, rain lines at different angles, and dense rain curtain scenes.

Images