A raw image reconstruction method, raw image compression method and system based on a third-order attention module

Abstract

The application provides a RAW image reconstruction method, a RAW image compression method and a system based on a third-order attention module, and belongs to the technical field of image data processing. The method comprises the following steps: constructing an image reconstruction model; inputting an RGB image into the image reconstruction model; processing the image in the image reconstruction model in two stages of encoding and decoding; performing inverse sub-pixel convolution operation on the decoded image data to obtain reconstructed image data and output the reconstructed image data; and performing compression operation on any RAW image data by inversely executing the previous operation to obtain an RGB image. The application specifically proposes an image reconstruction model comprising multiple levels, each level having different resolutions and a combination of multiple RISPBlock, so that the model has strong feature extraction, screening and reconstruction capabilities, and the image reconstruction model can be directly trained end to end, which is more suitable for actual scenes.

Description

Technical Field

[0001] This invention belongs to the technical field of image data processing, and in particular relates to a RAW image reconstruction method, a RAW image compression method and system based on a third-order attention module. Background Technology

[0002] As input to the camera's in-camera image signal processing (ISP) module, RAW images play an immeasurable role in the entire imaging system. As data directly derived from the camera's image sensor (CMOS), RAW data exhibits a linear correlation with scene irradiance.

[0003] The adjustments to RAW reflected in the output RGB image can be seen as a direct adjustment to external environmental factors, avoiding the blurring effect of nonlinear operations within the camera's image signal processing module (ISP). Therefore, image data processing operations such as HDR and noise reduction are all performed in the RAW domain.

[0004] In the absence of specific requirements, existing devices such as mobile phones often choose to retain only the final output color image (sRGB). Therefore, in this case, the non-linear operation in the camera's image signal processing module (ISP) makes subsequent image processing more challenging. Summary of the Invention

[0005] Objective: To propose a RAW image reconstruction method, a RAW image compression method, and a system based on a third-order attention module, in order to solve the aforementioned problems existing in the prior art. Through a simple yet effective inverse camera image signal processing module network, end-to-end training is possible to recover original camera image data from camera color image data. Furthermore, the third-order attention module achieves a complex trade-off between recovering spatial details and recovering high-level contextual information, thereby improving reconstruction accuracy.

[0006] Technical Solution: Firstly, a RAW image reconstruction method based on a third-order attention module is proposed, which specifically includes the following steps:

[0007] Step 1: Construct an image reconstruction model. In the process of image reconstruction, the beginning and end of the model processing are independent convolution operations, while the middle part is a symmetrical encoding and decoding operation.

[0008] Step 2: Transmit the RGB image as input image data to the image reconstruction model;

[0009] Step 3: The image reconstruction model processes the image by dividing it into two stages: encoding and decoding. In the encoding and decoding stages, the data processing is divided into 5 processing levels. The output data of the encoding layer at the same level is added to the input of the decoding layer through skip connections.

[0010] Step 4: Perform inverse subpixel convolution on the decoded image data to obtain the reconstructed image data;

[0011] Step 5: Output the reconstructed image data.

[0012] For an input RGB image of size [H, W, 3] (i.e., the height, width, and number of channels are H, W, and 3 respectively), an image reconstruction model is used to reconstruct the RAW image, outputting a camera-original image of size [4H, 4W, 1]. The values ​​of each pixel range from [0, 2^10-1]. The RISPNet image reconstruction model uses the RISPBlock module as the basic module at each layer in both the encoding and decoding stages. Decoding outputs an image of size [4H, 4W, 1], followed by pixel-unshuffle operations to output an image of size [H, W, 4]. The four channels correspond to the R, G, G, and B channels of the reconstructed RAW image, respectively. This image is equivalent to a camera-original image (RAW).

[0013] In some possible implementations of the first aspect, during the image processing in both the encoding and decoding stages, after the image data is decoded using the RISPBlock module, the reconstructed RAW image is output through inverse subpixel convolution. The RISPBlock module contains two residual skip connections, which are element-wise added to the output feature map of the network module before the residual connections.

[0014] The first-level residuals of the two residual skip connections in the RISPBlock module, in sequence, include: layer normalization, 1 1 convolution, 3 3 separable convolutions, third-order attention modules, 1 1. Convolution. The third-order attention module includes: a distributed attention module, a channel attention module, a spatial attention module, and an element-wise multiplication attention operation. The element-wise multiplication attention operation combines the feature map output by the distributed attention module with the next element in the sequence. The feature map output from the convolution operation is then multiplied element-wise. The distributed attention module contains a split operation and an element-wise multiplication operation. First, the input feature map is divided into two equal parts according to channels, and then the two parts are multiplied element-wise to output the processed feature map; subsequently, it is processed sequentially through 1... 1 convolution, parameter correction linear unit, 1x1 convolution operation. The middle part of the channel attention module contains, in sequence: global average pooling, 1x1 convolution, Gaussian error linear unit, 1... 1. Convolution and sigmoid activation function; multiply the output of the sigmoid activation function by the input of the channel attention module to obtain the output feature map of the channel attention module. The middle part of the spatial attention module sequentially includes: a combination of parallel global average pooling and global max pooling operations, then the outputs are merged, and then sequentially passed through a 5-step process. Five convolutions and a minimum-maximum normalization operation are applied to multiply the output feature map with the input of the spatial attention module to obtain the output feature map of the spatial attention module.

[0015] The second-level residuals, in sequence, include: layer normalization, simple attention gating module, and 1... 1. Convolution. The upper branch sequence of the simple attention gating module includes: 1 1 convolution, 3 3 separable convolutional, Gaussian error linear units. The next branch sequence includes: 1 1 convolution, 3 3. Separable convolution. Subsequently, based on the processing results of the upper and lower branches, the two output feature maps are multiplied element-wise and then output.

[0016] Among the possible implementations of the first aspect, RISPBlock mainly includes layer normalization (LayerNorm), 1 1 convolution (Conv), 3 3. Separable Convolution (DWConv), Parameter Corrected Linear Unit (PReLU), 5. The system employs basic deep learning operations such as Convolution (Conv), Gaussian Error Linear Unit (GELU), Global Average Pooling, Global Max Pooling, Element-wise Addition, Split, Concatenation, Element-wise Multiplication, and Sigmoid activation function. Each operation maintains the same input and output size ([H, W]), and the number of channels is C except for the Split operation, which halves the channel count. Overall, RISPBlock is a two-stage residual skip connection structure, with the final stage connected via an element-wise addition operation, i.e., adding element-wise to the output of the preceding network module. The first-stage residual sequence includes Layer Normalization (LayerNorm), 1... 1 convolution (Conv), 3 3 separable convolutions (DWConv), third-order attention (TOA) module, 1 1. Convolution (Conv), etc. The second-level residual sequence includes layer normalization (LayerNorm), Simple Attention Gate (SAG) module, 1 1. Convolution (Conv), etc.

[0017] Secondly, a RAW image reconstruction system based on a third-order attention module is proposed to implement the RAW image reconstruction method mentioned in the first aspect. The system specifically includes the following modules: model building module, data analysis module, image reconstruction module, and data output module.

[0018] The model building module is set up to build a RAW image reconstruction model;

[0019] The data analysis module is configured to input the original RGB image into the RAW image reconstruction model for data analysis;

[0020] The image reconstruction module is configured to reconstruct the RAW image from the received RGB image using a RAW image reconstruction model;

[0021] The data output module is set to output the RAW image obtained by the image reconstruction module.

[0022] Thirdly, a RAW image compression method based on a third-order attention module is proposed to realize the inverse process of the RAW image reconstruction method mentioned in the first aspect, that is, to compress any RAW image and output the corresponding RGB image. The system specifically includes the following modules: model building module, data analysis module, RAW image compression module, and data output module.

[0023] The model building module is set to be the inverse transform model of the RAW image reconstruction model mentioned in the second aspect;

[0024] The data analysis module is configured to input any RAW image into the inverse transform model of the RAW image reconstruction model for data analysis;

[0025] The RAW image compression module is configured to reconstruct RGB images from RAW images received using the inverse transform model of the RAW image reconstruction model.

[0026] The data output module is set to output RAW image compression module to obtain RGB image.

[0027] Beneficial effects: This invention proposes a RAW image reconstruction method, a RAW image compression method and system based on a third-order attention module. It specifically proposes a multi-level image reconstruction model RISPNet, with each level having different resolutions and combinations of multiple RISPBlocks. Thus, the model designed in this patent has strong feature extraction, filtering and reconstruction capabilities, and the RISPNet model can be directly trained end-to-end, making it more suitable for real-world scenarios.

[0028] RISPBlock contains a rich set of attention modules and employs a two-level residual structure. The first-level residual structure performs initial feature screening through an attention mechanism, while the second-level residual structure further enhances useful features through a gating mechanism, thereby improving the effectiveness of the features.

[0029] In addition, the third-order attention module in RISPNet achieves a balance between filtering detailed information and high-level contextual information in the network through multiple attention operations such as distributed attention module, channel attention, spatial attention, and element-wise multiplication. This effectively improves the efficiency of feature extraction and filtering at different levels, thereby improving the reconstruction performance of the RISPNet network. Attached Figure Description

[0030] Figure 1 This is a flowchart of the data processing of the RISPNet image reconstruction model of the present invention.

[0031] Figure 2 This is a schematic diagram of the RISPBlcok structure of the present invention. Detailed Implementation

[0032] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0033] The applicant argues that image data processing involves two aspects: camera raw image (RAW) restoration and camera color image (sRGB) restoration. The goal of camera color image restoration is to recover high-quality images from low-quality input images. In recent years, numerous deep learning-based methods such as Restormer, NAFNet, MultiStageIR, Uformer, PPIR, HINet, and RDN have been proposed, outperforming traditional methods like CIR1 and CIR2. Most of these applications are based on the encoder-decoder UNet architecture, capable of hierarchically extracting multi-scale representations, with skip connections typically used to focus on learning residual signals. Simultaneously, attention mechanisms popular in Natural Language Processing (NLP) have been incorporated into image restoration tasks to selectively focus on relevant information (Restormer, Uformer, IRA1, IRA2, IRA3, IRA4). Recently, NAFNet, utilizing a single-stage UNet architecture with enhancement blocks, has achieved good results in image deblurring and denoising tasks.

[0034] Camera Raw Image (RAW) Restoration Techniques: Methods such as UPI, MBISP, InvISP, and CycleISP aim to estimate the mapping from camera color images (sRGB) to camera raw images (RAW). UPI, after a detailed analysis of traditional in-camera ISPs, models and inverses each step of a general-purpose camera image signal processing module (ISP) pipeline using carefully crafted functions and their inverses. However, these function parameters are manually set based on camera priors and cannot be generalized to new devices. MBISP, a model-based camera image signal processing module (ISP), builds upon UPI, designing an explicit hierarchical and reversible architecture to approximate the bidirectional mapping between the RAW and RGB domains. Instead of manually fine-tuning transformation parameters like UPI, MBISP learns them directly from the data. Similar to MBISP, InvISP also establishes a learnable and reversible camera image signal processing module (ISP) framework, consisting of a set of bijective functions used to generate RGB images in the forward pass and recover RAW data in the reverse pass. The difference lies in that InvISP models the camera image signal processing module (ISP) pipeline as a whole, while MBISP explicitly simulates each stage of the camera image signal processing module (ISP) pipeline and strings them together to form the entire camera image signal processing module (ISP) pipeline. Unlike the single but reversible network used in MBISP and InvISP, CycleISP employs two independent network branches to convert RGB images to RAW data and back to RGB images. Its RGB2RAW network branch utilizes successive recursive residual sets to learn the mapping from RGB images to RAW data, containing multiple dual attention blocks named after parallel spatial attention and channel attention modules.

[0035] In practical applications of image data processing, although numerous image data processing technologies have emerged, the reverse process of camera image signal processing (RISP), which simulates the camera image signal processing (ISP) process, still suffers from poor processing performance. The main reasons are: 1. The camera image signal processing (ISP) process typically converts 12-bit or 14-bit RAW data into 8-bit RGB images. This conversion inevitably sacrifices some information, making it inherently irreversible. Therefore, simple mathematical models cannot adequately model the inverse process. 2. The camera image signal processing (ISP) process involves numerous sub-processes and various mathematical models. Simulating the ISP process requires extensive parameter tuning, consuming significant manpower. 3. Existing deep learning-based methods have relatively simple structures and cannot effectively handle the large number and diverse types of images.

[0036] To address practical problems, this application proposes a RAW image reconstruction method, a RAW image compression method, and a system based on a third-order attention module, for reconstructing camera color image (sRGB) signals into camera raw images (RAW). Through an end-to-end trainable inverse camera image signal processing (RISP) network, camera raw image (RAW) data is recovered from camera color image (sRGB) data. Simultaneously, the proposed third-order attention (TOA) module improves reconstruction accuracy by balancing the complex trade-off between restoring spatial details and high-level contextual information during the image reconstruction process.

[0037] Example 1

[0038] In one embodiment, a RAW image reconstruction method based on a third-order attention module is proposed to reconstruct the camera's color image signal into the camera's original image. The method specifically includes the following steps:

[0039] Step 1: Construct the image reconstruction model RISPNet;

[0040] Specifically, the image reconstruction model RISPNet, such as Figure 1 As shown, both ends of the model are separate 3 The convolution operation of 3 includes two stages: encoding and decoding. The middle part is divided into 5 layers from top to bottom, where the output of the encoding layer at the same level is added to the input of the decoding layer through skip connections.

[0041] Step 2: Transmit the RGB image as input image data to the image reconstruction model;

[0042] Step 3: The image reconstruction model processes the image by dividing it into two stages: encoding and decoding.

[0043] Specifically, during the encoding stage, the size of the feature maps at different levels decreases progressively, with each decrease being by a factor of 2. The resolution of the feature maps is reduced by setting the stride of the last convolution in that layer to 2, while the stride of other convolutional layers is set to 1 to ensure that the resolution remains unchanged.

[0044] The size of the feature map in the decoding layer increases incrementally, with each increment being a factor of 2. This increase is achieved through a pixel-shuffle upsampling operation.

[0045] Step 4: Perform inverse subpixel convolution on the decoded image data to obtain the reconstructed image data;

[0046] Step 5: Output the reconstructed image data.

[0047] In a further embodiment, deep learning-based methods perform well in tasks such as denoising, super-resolution, and low-light enhancement in the RAW image domain. However, they require a large amount of paired RAW and sRGB image data as training data. In practical applications, it is often necessary to retain only the final output sRGB image data. This is because RAW image processing involves specialized domain knowledge and has low universality, and because RAW images occupy a large amount of device space.

[0048] To address the problems existing in the prior art, this embodiment employs a RAW image reconstruction method based on a third-order attention module to reconstruct the original camera image (RAW), thereby easily constructing a large number of paired camera images.<RAW,sRGB> Data pairs are used for tasks such as denoising, super-resolution, and low-light enhancement in the camera's raw image (RAW) domain.

[0049] In a preferred embodiment, for cloud service products containing GPU computing power, the user specifies the requirements, model, training parameters, etc., and also provides the required camera color images (sRGB), with the specific data volume determined by the user. This cloud service product can use the image reconstruction model RISPNet proposed in this embodiment to reconstruct the corresponding original camera images (RAW) as input, thereby constructing a large number of... <raw-srgb>The data pairs, combined with user-provided models, training parameters, and other information, provide a user interface for using this cloud service product.

[0050] The proposed image reconstruction model RISPNet effectively fuses multi-level resolution features to extract features at different resolution levels, thereby improving the reconstruction performance from camera color image (sRGB) signals to camera raw image (RAW) signals. Furthermore, the reconstruction from camera color image (sRGB) signals to camera raw image (RAW) signals can construct a large number of camera pairs at a very low cost.<RAW,sRGB> The data pairs are used for training models for tasks such as denoising, super-resolution, and low-light enhancement in the camera's raw image (RAW) domain. Meanwhile, the RISPNet model proposed in this embodiment can effectively improve the PSNR index and visualization effect of reconstructing camera color image (sRGB) signals from camera raw images (RAW).

[0051] Example 2

[0052] In a further embodiment based on Embodiment 1, based on the proposed RAW image reconstruction method, this embodiment uses an image reconstruction model to reconstruct a RAW image with an input size of [H, W, 3], i.e., the height, width, and number of channels of the image are H, W, and 3 respectively, and outputs a camera original image with a size of [4H, 4W, 1]. The value range of each pixel is [0, 2^10-1].

[0053] The RISPNet image reconstruction model used employs a RISPBlock module as the basic module at each layer of the encoding and decoding stages. The decoding output is an image of size [4H, 4W, 1]. Then, through inverse subpixel convolution (pixel-unshuffle) operation, an image of size [H, W, 4] is output. The four channels correspond to the R, G, G, and B channels of the reconstructed RAW image, respectively. This image is equivalent to a camera's original image (RAW).

[0054] like Figure 2 As shown, RISPBlock contains two residual skip connections, which are connected at the end by an element-wise addition operation, that is, element-wise addition with the output feature map of the network module before the residual connection.

[0055] Specifically, the first-level residual order includes Layer Normalization (LayerNorm), 1 1 convolution (Conv), 3 3 separable convolutions (DWConv), third-order attention (TOA) module, 1 The system includes four basic operations: convolution (Conv), etc. The third-order attention module comprises multiple attention operations such as split attention, channel attention (CA), spatial attention (SA), and element-wise multiplication. The element-wise multiplication attention module combines the feature map output from the split attention module with the aforementioned last... 1. The feature maps output by the convolution (Conv) operation are multiplied element-wise.

[0056] The second-level residual ordering includes LayerNorm, SimpleAttentionGate (SAG), and 1 1. Convolution (Conv), etc. Among them, the branching order on the Simple AttentionGate (SAG) module includes 1... 1 convolution (Conv), 3 3. Separable convolution (DWConv), Gaussian error linear unit (GELU), and other three operations, the order of the down branches includes 1. 1 convolution (Conv), 3 3. Two operations, such as separable convolution (DWConv), are performed, and then the two output feature maps are multiplied element-wise.

[0057] In a further embodiment, the third-order attention (TOA) module includes three attention operations. Based on the mechanism of human vision perceiving multiple targets simultaneously, the first is a split attention module, which includes a split operation and an element-wise multiplication operation. Specifically, the input feature map is first divided into two equal parts by channels, with a size of [H, W, C / 2]. Then, the two parts are multiplied element-wise to output a feature map of [H, W, C / 2]. This is then processed sequentially through 1... The operations include 1x1 convolution (Conv), parameter-corrected linear unit (PReLU), and 1x1 convolution (Conv). The first 1x1 convolution (Conv) restores the output feature map size to [H, W, C], and the second 1x1 convolution (Conv)... 1. Convolution (Conv) does not change the size of the feature map, which remains [H, W, C].

[0058] The second attention operation comprises two parallel sub-modules: Channel Attention (CA) and Spatial Attention (SA). For example... Figure 2 As shown, the last 1 above The first convolution (Conv) operation outputs a feature map of size [H, W, C], which is then input into the CA and SA sub-modules respectively. The outputs two feature maps of size [H, W, C]. After a concatenation operation, a single feature map of size [H, W, C] is output. This is then processed by a third convolution operation. After the first convolution (Conv) operation, a feature map of size [H, W, C] is output. The third attention submodule is a skip element-wise multiplication operation, which combines the feature map output from the split attention module with the feature map from the last convolution operation. The feature maps output by the first convolution (Conv) operation are multiplied element-wise, resulting in an output size of [H, W, C / 2]. Thus, the input size of the third-order attention (TOA) module is [H, W, C], and the output size is [H, W, C / 2].

[0059] In a further embodiment, the channel attention (CA) submodule also includes a skip element-wise multiplication operation, with the intermediate part sequentially including global average pooling, 1x1 convolution (Conv), Gaussian error linear unit (GELU), and 1... Operations such as convolution (Conv) and sigmoid activation function are performed, with each operation having an input and output size of [H, W, C]. Finally, the output of the sigmoid activation function is multiplied by the input of the channel attention (CA) submodule, and the result is the output feature map of the channel attention (CA) submodule.

[0060] In a further embodiment, the SA submodule also includes a skip element-wise multiplication operation, with the middle part sequentially including a combination of parallel global average pooling and global max pooling operations, then concatenating the output results, and then sequentially passing them through a 5-bit multiplication table. Five convolutional (Conv) operations and a min-max normalization (Min-Max Normalization) operation are performed to output a feature map of [H, W, C]. Finally, this feature map is multiplied with the input of the SA submodule, and the result is the output feature map of the SA submodule.

[0061] In a further embodiment, the Simple Attention Gate (SAG) included in the second-level residual module of the RISPBlock module is a two-branch structure, with the upper branch sequence including 1 1 convolution (Conv), 3 3. Separable convolution (DWConv), Gaussian error linear unit (GELU), and other three operations, the order of the down branches includes 1. 1 convolution (Conv), 3 3. Two operations can be separated, such as convolution (DWConv). The input of the Simple Attention Gate (SAG) goes through two branches, and then the two output feature maps are multiplied element-wise to obtain the output of the Simple Attention Gate (SAG).

[0062] This example further proposes a RISPBlock module for image data reconstruction based on the existing image reconstruction model, and performs continuous feature extraction and network architecture fusion based on the RISPBlock module. Simultaneously, the third-order attention module (TOA) in the RISPBlock module enhances the network's balance in filtering detailed information and high-level contextual information, thereby effectively improving the efficiency of feature extraction and filtering at different levels. Furthermore, this embodiment verifies the effectiveness of the Simple Attention Gate (SAG) module in RISPBlock for inverse ISP RAW image reconstruction, and its combination with the third-order attention module (TOA) further improves the network's reconstruction performance.

[0063] Example 3

[0064] In a further embodiment based on Embodiment 1, a RAW image compression method based on a third-order attention module is proposed to implement arbitrary RAW image compression processes, such as... Figure 3 As shown, the method specifically includes the following steps:

[0065] Step 1: Receive any RAW image data;

[0066] Step 2: Perform inverse transformation on the RAW image data using the image reconstruction model disclosed in Example 1;

[0067] Step 3: Output the inverse-transformed RGB image data to obtain the compressed RAW image data.

[0068] In a further embodiment, the data flow for implementing the inverse transformation of the RAW image reconstruction model first involves rearranging (packing) the acquired RAW image data to obtain RAW image data with half the resolution and 4 channels; then, the feature map with reduced resolution is subjected to the inverse transformation of the RAW image reconstruction process; finally, the RAW image data is compressed.

[0069] Example 4

[0070] In one embodiment, a RAW image reconstruction system based on a third-order attention module is proposed to implement a RAW image reconstruction method based on a third-order attention module. The system specifically includes: a model building module, a data analysis module, an image reconstruction module, and a data output module.

[0071] In a further embodiment, based on the reconstruction requirements of the RAW image, the model building module constructs an image reconstruction model; subsequently, the data analysis module uses the RGB image as input image data and transmits it to the image reconstruction model for data analysis; during the image reconstruction process, the image reconstruction model processes the image by dividing it into two stages: encoding and decoding, and performs inverse sub-pixel convolution on the decoded image data to obtain the reconstructed image data; finally, the data output module outputs the reconstructed image data.

[0072] The image reconstruction model constructed by the model building module includes separate convolutional layers at both ends of the model, and encoding and decoding modules placed in the middle of the model in a symmetrical distribution. The encoding and decoding modules are divided into 5 levels from top to bottom, with the output of the encoding module at the same level being added to the input of the decoding module through skip connections.

[0073] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A RAW image reconstruction method based on a third-order attention module, characterized in that, Includes the following steps: Step 1: Construct an image reconstruction model; Step 2: Transmit the RGB image as input image data to the image reconstruction model; Step 3: The image reconstruction model processes the image through two stages: encoding and decoding. During the image processing in the two stages of encoding and decoding, the RISPBlock module is used to decode the image data, and then the reconstructed RAW image is output through inverse subpixel convolution operation. The RISPBlock module contains two residual skip connections, which are added element-wise to the output feature map of the network module before the residual connection. The first-level residuals of the two residual jump connections in the RISPBlock module, in sequence, include: Layer normalization convolution, Separable convolution, third-order attention module, convolution; The third-order attention module includes: a distributed attention module, a channel attention module, a spatial attention module, and an element-wise multiplication attention operation; The element-wise multiplication attention operation combines the feature map output by the distraction attention module with the next element in the sequence. The feature maps output by the convolution operation are multiplied element-wise. The distraction attention module includes a Split operation and an element-wise multiplication operation. First, the input feature map is divided into two equal parts by channel, and then the two parts are multiplied element-wise to output the processed feature map. Subsequently, it passes through... Convolution, parameter-corrected linear unit, Convolution operation; The middle part of the channel attention module sequentially includes: global average pooling, Convolution, Gaussian error linear unit, Convolution and sigmoid activation function; multiply the output of the sigmoid activation function with the input of the channel attention module to obtain the output feature map of the channel attention module; The middle part of the spatial attention module sequentially includes: a combination of parallel global average pooling and global max pooling operations, then merging the outputs, and then sequentially passing them through a... A convolution and a minimum-maximum normalization operation are applied to multiply the output feature map with the input of the spatial attention module to obtain the output feature map of the spatial attention module. Step 4: Perform inverse subpixel convolution on the decoded image data to obtain the reconstructed image data; Step 5: Output the reconstructed image data.

2. The RAW image reconstruction method based on a third-order attention module according to claim 1, characterized in that, In the process of image reconstruction, the image reconstruction model performs independent convolution operations at the beginning and end, and the middle part consists of encoding and decoding operations with a symmetrical structure.

3. The RAW image reconstruction method based on a third-order attention module according to claim 1, characterized in that, In the encoding and decoding stages, the data processing is divided into 5 processing levels, where the output data of the encoding layer at the same level is added to the input of the decoding layer through skip connections.

4. The RAW image reconstruction method based on a third-order attention module according to claim 1, characterized in that, The second-level residuals of the two residual jump connections in the RISPBlock module, in sequence, include: Layer normalization, simple attention gating module, convolution.

5. The RAW image reconstruction method based on a third-order attention module according to claim 4, characterized in that, The branching sequence of the simple attention gating module includes: convolution, Separable convolutional and Gaussian error linear units; The order of the next branch includes: convolution, Separable convolution; Based on the processing results of the upper and lower branches, the two output feature maps are multiplied element-wise and then output.

6. A RAW image compression method based on a third-order attention module, used to model the inverse process of the RAW image reconstruction process as described in claim 1, that is, to perform compression operations on arbitrary RAW image data to convert it into an RGB image, characterized in that, The method includes: Step 1: Receive any RAW image data; Step 2: Perform inverse transformation on the RAW image data using the RAW image reconstruction model as described in claim 1; Step 3: Output the inverse-transformed RGB image data to obtain the compressed RAW image data.

7. A RAW image reconstruction system based on a third-order attention module, used to implement the RAW image reconstruction method as described in any one of claims 1 to 5, characterized in that, Specifically, it includes the following modules: The model building module is configured to build image reconstruction models; The data analysis module is configured to input the original RGB image into the image reconstruction model for data analysis; The image reconstruction module is configured to reconstruct the RAW image from the received RGB image using an image reconstruction model; The data output module is set to output the RAW image obtained by the image reconstruction module.

Images