Image compression system fusing edge perception and density guidance

Abstract

The invention discloses an image compression system fusing edge perception and density guidance. The image compression system comprises an image preprocessing module used for carrying out channel separation and size standardization on an RGB image to be compressed; the density mapping module is used for mapping the channel separated image into a physical density space diagram; the edge sensing module is used for extracting an edge image from the RGB image with the standardized size; the compression encoder is used for forming a feature tensor by the RGB image with the standardized size, the physical density space graph and the edge graph and extracting high-semantic image compression features from the feature tensor; the low-dimensional encoder output module is used for carrying out dimensionality reduction on the image compression features to obtain dimensionality-reduced features; and the compression decoding module is used for reconstructing a compressed image from the dimension reduction features. According to the method, the edge features and the density features are reserved in the image compression process, and the recognizable precision of the compressed image is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of image compression technology, specifically, it relates to an image compression system that integrates edge perception and density guidance. Background Technology

[0002] In recent years, with the widespread application of artificial intelligence in fields such as security inspection, medical care, and industrial inspection, the ability to acquire X-ray and CT images has been significantly enhanced, with their resolution and frame rate continuously improving. In security inspection scenarios, an X-ray machine acquires images at a rate of 20 frames per second, with each frame exceeding 1MB in size, generating tens of gigabytes of raw data per hour. CT image data is even more massive; a single scan sequence after 3D reconstruction can reach hundreds of megabytes or even several gigabytes. Due to the enormous data volume of X-ray and CT images, real-time response is difficult when transmitting data between terminal devices and central servers due to network bandwidth and latency limitations. Furthermore, the storage and processing of high-resolution images rely on GPU resources, significantly increasing operational costs.

[0003] Traditional image compression methods with visual optimization, such as JPEG and JPEG2000, cannot effectively preserve the structural edges and physical density features of images, resulting in compressed images that are unsuitable for intelligent recognition. With the development of deep learning, deep learning-based image compression models exist; however, their complex network structures and large number of model parameters lead to slow inference speeds, poor robustness, and difficulty in deployment on real-world terminals. Furthermore, deep learning-based image compression models neglect the physical semantic information of images during compression, resulting in a significant decrease in the recognition performance of the compressed images. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an image compression system that integrates edge perception and density guidance, which preserves edge features and density features during the image compression process, greatly improving the identifiability of the compressed image.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution: An image compression system integrating edge awareness and density guidance includes: an image preprocessing module, a density mapping module, an edge awareness module, a compression encoder, a low-dimensional encoder output module, and a compression decoder; The image preprocessing module is used to perform channel separation and size normalization on the RGB image to be compressed; The density mapping module is used to map channel-separated images into physical density space maps. The edge sensing module is used to extract edge maps from size-normalized RGB images; The size-normalized RGB image, physical density space map, and edge map are combined to form a feature tensor, and the compression encoder is used to extract high-semantic image compression features from the feature tensor; The low-dimensional encoder output module is used to reduce the dimensionality of the image compression features to obtain the dimensionality-reduced features. The compression decoding module is used to reconstruct the compressed image from the dimensionality reduction features.

[0006] Furthermore, the density mapping module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer; The first convolutional layer passes through The convolution kernel extracts local spatial features on the channel-separated image, expands the number of channels, and performs nonlinear modeling through the ReLU activation function to generate a channel-expanded local spatial feature map; The second convolutional layer passes through The convolutional kernel extracts high-level features from the local spatial feature map with channel expansion, normalizes the features of each channel of the high-level features through batch normalization, and performs nonlinear modeling through the ReLU activation function to generate a high-level feature map that preserves the channels. The third convolutional layer passes through The convolutional kernel compresses the number of channels in the high-level feature map to 1, generating a single-channel initial density map; The output layer maps the initial density map to a physical density space map.

[0007] Furthermore, the compression encoder consists of multi-level cascaded convolutional feature extraction modules, each of which includes: a fourth convolutional layer, a fifth convolutional layer, and a max pooling layer; The fourth convolutional layer is used to extract local feature maps of the input features; The fifth convolutional layer is used to extract higher-level feature maps from local feature maps; The max pooling layer downsamples the feature maps at higher levels.

[0008] Furthermore, the compression decoding module is composed of convolutional feature reconstruction modules cascaded at the same level as the compression encoder, and each convolutional feature reconstruction module includes: an upsampling layer and a sixth convolutional layer; The upsampling layer is used to upsample the input features; The sixth convolutional layer is used to combine the upsampled features with higher-level feature maps extracted from the corresponding layers in the compression decoding module to reconstruct the reconstructed compressed feature map.

[0009] Furthermore, a fusion loss function based on pixel loss, density loss, edge loss, and structure loss is constructed. The image compression system is trained by collecting historical images and corresponding compressed images until the fusion loss function converges, thus completing the training of the image compression system.

[0010] Furthermore, the process of constructing the fusion loss function is as follows:

[0011] in, Represents the fusion loss function. Indicates pixel loss, express Weighting coefficients; Indicates density loss, express Weighting coefficients; Indicates structural loss. express Weighting coefficients; Indicates marginal loss. express The weighting coefficients.

[0012] Furthermore, the pixel loss is used to compare the differences in pixel values ​​between the reconstructed compressed image and the corresponding real compressed image:

[0013] in, Represents the number of pixels in an image. Represents pixel coordinates. This indicates that the actual compressed image has pixel coordinates of . Pixel value at that location, The reconstructed compressed image is represented by pixel coordinates as follows: The pixel value at that location.

[0014] Furthermore, the density loss is used to compare the difference in density values ​​between the reconstructed compressed image and the corresponding real image:

[0015] in, Represents pixel coordinates. This indicates that the real image is located at pixel coordinates. Density value at that location, The reconstructed compressed image is represented by pixel coordinates as follows: Density value at that location, express Norm.

[0016] Furthermore, the structural loss is used to compare the differences in pixel values ​​between the compressed feature map reconstructed by the compression decoder and the real image:

[0017] in, Represents a real image. This indicates the number of convolutional feature reconstruction modules in the compression decoder of an image compression system. Indicates the first The compressed feature map reconstructed by each convolutional feature reconstruction module. This represents the L2 norm.

[0018] Furthermore, the edge loss is used to compare the differences in edge gradients between the reconstructed compressed image and the corresponding ground truth image:

[0019] in, Represents pixel coordinates. This indicates that the real image is located at pixel coordinates. Edge operators at the location, The reconstructed compressed image is represented by pixel coordinates as follows: Edge operators at the location, This represents the L2 norm.

[0020] Compared with the prior art, the present invention has the following beneficial effects: (1) The image compression system of the present invention integrates edge perception and density guidance. It uses a size-standardized RGB image, a physical density space map and an edge map to form a feature tensor. The RGB image provides rich color and texture information, which can reflect the appearance features of objects in the image. The physical density space map describes the physical attribute information of objects in the image and supplements the semantic information of the image. The edge map can highlight the contour and boundary of objects in the image and enhance the perception of shape and structure. By constructing a feature tensor with the three, it can simultaneously capture the appearance, physical attributes and edge structure information of objects in the image, form a comprehensive feature representation, provide rich image features for subsequent image compression, and greatly improve the recognition accuracy of compressed images. (2) The image compression system of the present invention, which integrates edge perception and density guidance, adopts a symmetrical encoder-decoder network structure. The compression encoder gradually extracts image features and compresses data dimensions through convolutional layers and max pooling layers. The compression decoder gradually restores image details and reconstructs compressed images through upsampling and convolutional layers. Since the number of layers and channels of the compression encoder and the compression decoder are symmetrically designed, the computational complexity of the network is reduced, which is suitable for terminal deployment. At the same time, the object outline is highlighted by the edge map and the compression encoder is guided to focus on key areas by the physical density space map, thereby improving the visual quality of the reconstructed compressed image and enhancing the robustness of the image compression system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the image compression system that integrates edge sensing and density guidance according to the present invention; Figure 2 This is a schematic diagram of the density mapping module in this invention; Figure 3 This is a schematic diagram showing the connection between the compression encoder, the low-dimensional encoder output module, and the compression decoder in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be further explained and described below with reference to the accompanying drawings.

[0023] like Figure 1 This is a schematic diagram of the image compression system integrating edge awareness and density guidance of the present invention. The image compression system includes: an image preprocessing module, a density mapping module, an edge awareness module, a compression encoder, a low-dimensional encoder output module, and a compression decoder. The edge map and physical density space map are acquired in parallel by the edge awareness module and the density mapping module to reduce coupling. At the same time, a symmetrical encoder-decoder network structure is adopted. The compression encoder gradually extracts image features and compresses data dimensions through convolutional layers and max pooling layers. The compression decoder gradually restores image details and reconstructs the compressed image through upsampling and convolutional layers. Since the number of layers and channels of the compression encoder and compression decoder are symmetrically designed, the computational complexity of the network is reduced, making it suitable for terminal deployment.

[0024] The image preprocessing module is used to separate channels and normalize the size of the RGB image to be compressed.

[0025] The density mapping module is used to map channel-separated images into physical density space maps, thereby improving the image compression system's ability to model physical semantic information. For example... Figure 2 The density mapping module in this invention includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer; The first convolutional layer passes through Convolutional kernels extract local spatial features from channel-separated images, expand the number of channels to enhance feature representation, and perform nonlinear modeling through the ReLU activation function to generate channel-expanded local spatial feature maps, providing rich detail information; The second convolutional layer passes through The convolutional kernel extracts high-level features from the local spatial feature map with channel expansion. Batch normalization is used to normalize the features of each channel of the high-level features to alleviate the covariate shift on the channel. Nonlinear modeling is performed through the ReLU activation function to generate a high-level feature map that preserves the channel, thereby enhancing the robustness of the features and avoiding information loss. The third convolutional layer passes through The convolutional kernel compresses the number of channels in the high-level feature map to 1, generating a single-channel initial density map; The output layer maps the initial density map to a physical density space map.

[0026] The edge-aware module is used to extract edge maps from size-normalized RGB images, thereby enhancing the ability to recognize image edge structures.

[0027] A feature tensor is constructed by combining a size-normalized RGB image, a physical density space map, and an edge map. The RGB image provides rich color and texture information, reflecting the appearance features of objects in the image. The physical density space map describes the physical properties of objects in the image, supplementing the semantic information of the image. The edge map can highlight the contours and boundaries of objects in the image, enhancing the perception of shape and structure. By constructing a feature tensor using these three elements, the appearance, physical properties, and edge structure information of objects in the image can be captured simultaneously, forming a comprehensive feature representation. This provides rich image features for subsequent image compression, greatly improving the identifiability of compressed images.

[0028] A compression encoder is used to progressively extract high-semantic image compression features from feature tensors. It highlights object contours through edge maps, enabling the encoder to preserve these contours during encoding. Furthermore, it guides the encoder to focus on key regions using physical density space maps, thereby improving the visual quality of the reconstructed compressed image and enhancing the robustness of the image compression system. For example... Figure 3 In this invention, the compression encoder consists of multiple cascaded convolutional feature extraction modules. Each convolutional feature extraction module includes a fourth convolutional layer, a fifth convolutional layer, and a max pooling layer. The fourth convolutional layer is used to extract local feature maps of the input features. The fifth convolutional layer is used to extract higher-level feature maps from the local feature maps. The max pooling layer halves the resolution and doubles the number of channels by downsampling the higher-level feature maps.

[0029] The low-dimensional encoder output module is used to reduce the dimensionality of image compression features to obtain dimensionality-reduced features.

[0030] The compression decoding module is used to reconstruct the compressed image from the dimensionality-reduced features. The compression decoding module consists of convolutional feature reconstruction modules cascaded at the same level as the compression encoder. Each convolutional feature reconstruction module includes an upsampling layer and a sixth convolutional layer. The upsampling layer is used to upsample the input features, which doubles the resolution, halves the number of channels, and preserves global semantic information. The sixth convolutional layer is used to combine the upsampled features with higher-level feature maps extracted from the corresponding level in the compression decoding module to reconstruct the compressed feature map.

[0031] In one technical solution of the present invention, a fusion loss function based on pixel loss, density loss, edge loss and structure loss is constructed. The image compression system is trained by collecting historical images and corresponding compressed images until the fusion loss function converges, thus completing the training of the image compression system.

[0032] The process of constructing the fusion loss function in this invention is as follows:

[0033] in, Represents the fusion loss function. Indicates pixel loss, express Weighting coefficients; Indicates density loss, express Weighting coefficients; Indicates structural loss. express Weighting coefficients; Indicates marginal loss. express The weighting coefficients.

[0034] Pixel loss is used to compare the differences in pixel values ​​between the reconstructed compressed image and the corresponding real compressed image, ensuring that the reconstructed compressed image is close to the real image in overall brightness and color distribution, avoiding severe distortion. The calculation process of pixel loss is as follows:

[0035] in, Represents the number of pixels in an image. Represents pixel coordinates. This indicates that the actual compressed image has pixel coordinates of . Pixel value at that location, The reconstructed compressed image is represented by pixel coordinates as follows: The pixel value at that location.

[0036] Density loss is used to compare the density values ​​of the reconstructed compressed image and the corresponding real image, ensuring that the density distribution of the reconstructed compressed image is consistent with that of the real image, thereby improving the semantic understanding of the scene by the image compression system. The calculation process of density loss is as follows:

[0037] in, Represents pixel coordinates. This indicates that the real image is located at pixel coordinates. Density value at that location, The reconstructed compressed image is represented by pixel coordinates as follows: Density value at that location, express Norm.

[0038] Structural loss is used to compare the differences in pixel values ​​between the compressed feature map reconstructed by the compression decoder and the real image, guiding the reconstructed compressed feature map to be perceptually closer to the real image. The calculation process of structural loss is as follows:

[0039] in, Represents a real image. This indicates the number of convolutional feature reconstruction modules in the compression decoder of an image compression system. Indicates the first The compressed feature map reconstructed by each convolutional feature reconstruction module. This represents the L2 norm.

[0040] Edge loss is used to compare the differences in edge gradients between the reconstructed compressed image and the corresponding ground truth image. It preserves the spatial structure of the compressed image through edge alignment. The calculation process for edge loss is as follows:

[0041] in, Represents pixel coordinates. This indicates that the real image is located at pixel coordinates. Edge operators at the location, The reconstructed compressed image is represented by pixel coordinates as follows: Edge operators at the location, This represents the L2 norm.

[0042] This invention integrates edge-aware and density-guided image compression systems, which improve the recognizability and compression rate of compressed images, resulting in images with smaller feature dimensions and faster transmission speeds. It is particularly suitable for transmitting large batches of security inspection and CT images in wireless, low-bandwidth network environments.

[0043] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. An image compression system integrating edge awareness and density guidance, characterized in that, include: Image preprocessing module, density mapping module, edge sensing module, compression encoder, low-dimensional encoder output module, and compression decoder; The image preprocessing module is used to perform channel separation and size normalization on the RGB image to be compressed; The density mapping module is used to map channel-separated images into physical density space maps. The edge sensing module is used to extract edge maps from size-normalized RGB images; The size-normalized RGB image, physical density space map, and edge map are combined to form a feature tensor, and the compression encoder is used to extract high-semantic image compression features from the feature tensor; The low-dimensional encoder output module is used to reduce the dimensionality of the image compression features to obtain the dimensionality-reduced features. The compression decoding module is used to reconstruct the compressed image from the dimensionality reduction features.

2. The image compression system integrating edge awareness and density guidance according to claim 1, characterized in that, The density mapping module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer; The first convolutional layer passes through The convolution kernel extracts local spatial features on the channel-separated image, expands the number of channels, and performs nonlinear modeling through the ReLU activation function to generate a channel-expanded local spatial feature map; The second convolutional layer passes through The convolutional kernel extracts high-level features from the local spatial feature map with channel expansion, normalizes the features of each channel of the high-level features through batch normalization, and performs nonlinear modeling through the ReLU activation function to generate a high-level feature map that preserves the channels. The third convolutional layer passes through The convolutional kernel compresses the number of channels in the high-level feature map to 1, generating a single-channel initial density map; The output layer maps the initial density map to a physical density space map.

3. The image compression system integrating edge awareness and density guidance according to claim 1, characterized in that, The compression encoder consists of multiple cascaded convolutional feature extraction modules, each of which includes a fourth convolutional layer, a fifth convolutional layer, and a max pooling layer. The fourth convolutional layer is used to extract local feature maps of the input features; The fifth convolutional layer is used to extract higher-level feature maps from local feature maps; The max pooling layer downsamples the feature maps at higher levels.

4. The image compression system integrating edge awareness and density guidance according to claim 3, characterized in that, The compression decoding module is composed of convolutional feature reconstruction modules cascaded at the same level as the compression encoder. Each convolutional feature reconstruction module includes an upsampling layer and a sixth convolutional layer. The upsampling layer is used to upsample the input features; The sixth convolutional layer is used to combine the upsampled features with higher-level feature maps extracted from the corresponding layers in the compression decoding module to reconstruct the reconstructed compressed feature map.

5. The image compression system integrating edge awareness and density guidance according to claim 1, characterized in that, A fusion loss function based on pixel loss, density loss, edge loss, and structure loss is constructed. The image compression system is trained by collecting historical images and corresponding compressed images until the fusion loss function converges, thus completing the training of the image compression system.

6. The image compression system integrating edge awareness and density guidance according to claim 5, characterized in that, The process of constructing the fusion loss function is as follows: in, Represents the fusion loss function. Indicates pixel loss, express Weighting coefficients; Indicates density loss, express Weighting coefficients; Indicates structural loss. express Weighting coefficients; Indicates marginal loss. express The weighting coefficients.

7. The image compression system integrating edge awareness and density guidance according to claim 6, characterized in that, The pixel loss is used to compare the differences in pixel values ​​between the reconstructed compressed image and the corresponding real compressed image: in, Represents the number of pixels in an image. Represents pixel coordinates. This indicates that the actual compressed image has pixel coordinates of . Pixel value at that location, The reconstructed compressed image is represented by pixel coordinates as follows: The pixel value at that location.

8. The image compression system integrating edge awareness and density guidance according to claim 6, characterized in that, The density loss is used to compare the difference in density values ​​between the reconstructed compressed image and the corresponding real image: in, Represents pixel coordinates. This indicates that the real image is located at pixel coordinates. Density value at that location, The reconstructed compressed image is represented by pixel coordinates as follows: Density value at that location, express Norm.

9. The image compression system integrating edge awareness and density guidance according to claim 6, characterized in that, The structural loss is used to compare the differences in pixel values ​​between the compressed feature map reconstructed by the compression decoder and the real image: in, Represents a real image. This indicates the number of convolutional feature reconstruction modules in the compression decoder of an image compression system. Indicates the first The compressed feature map reconstructed by each convolutional feature reconstruction module. This represents the L2 norm.

10. An image compression system integrating edge awareness and density guidance according to claim 6, characterized in that, The edge loss is used to compare the difference in edge gradients between the reconstructed compressed image and the corresponding real image: in, Represents pixel coordinates. This indicates that the real image is located at pixel coordinates. Edge operators at the location, The reconstructed compressed image is represented by pixel coordinates as follows: Edge operators at the location, This represents the L2 norm.

Images