Image compression method based on traffic scene

Abstract

The invention provides an image compression method based on a traffic scene, and relates to the technical field of image processing, and the method comprises the steps: collecting video image data in the traffic scene, and constructing a traffic video image data set; defining an overall framework of an image compression model, deeply compressing a potential codec and a de-noising UNet network, and constructing a model loss function at the same time; and performing model training on the image compression model based on the traffic video image data set, constructing a traffic image compression model, and performing image compression processing through the traffic image compression model. The technical problems that in an existing image compression technology, calculation is complex, compression performance is poor, and consistency of image reconstruction cannot be guaranteed are solved. The technical effects of improving the compression performance and realizing high-quality and high-fidelity image reconstruction while ensuring low calculation complexity are achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an image compression method based on traffic scenes. Background Technology

[0002] In recent years, high-position video road monitoring technology has developed rapidly. By installing high-position video cameras at roadside locations and intersections to collect data, and utilizing visual algorithms to perform comprehensive traffic scene analysis, it facilitates more refined and accurate management of different traffic scenarios. Furthermore, it can contribute to various aspects such as intelligent transportation systems, traffic safety, data analysis, emergency response, smart city construction, and public transportation optimization, thus playing a positive role in urban traffic management and driving safety. However, this process presents significant challenges to the storage and transmission of data captured by high-position cameras. Image compression technology, by reducing redundant information in image data, enables more efficient storage and transmission of data.

[0003] Traditional image data compression methods mainly include JPEG, JPEG2000, and BPG. In recent years, with the development of deep learning, deep learning-based image compression methods have also emerged. Deep learning-based image compression methods have explored various model architectures, such as convolutional neural networks and transformers, to model the distribution of image content, thereby achieving compression effects. However, achieving good image compression performance while maintaining low computational complexity remains challenging, and the generated images often deviate from the original images, failing to guarantee the consistency of image reconstruction.

[0004] Existing image compression technologies suffer from technical problems such as computational complexity, poor compression performance, and inability to guarantee the consistency of image reconstruction. Summary of the Invention

[0005] The purpose of this application is to provide an image compression method based on traffic scenes, which solves the technical problems of existing image compression technologies, such as computational complexity, poor compression performance, and inability to guarantee the consistency of image reconstruction.

[0006] In view of the above problems, this application provides an image compression method based on traffic scenes. The method includes: collecting video image data in traffic scenes to construct a traffic video image dataset; defining the overall framework of an image compression model, a deep compression latent codec, and a denoising UNet network, while constructing a model loss function; training the image compression model based on the traffic video image dataset to construct a traffic image compression model, and performing image compression processing through the traffic image compression model.

[0007] Optionally, the overall framework of the image compression model is based on SD-Turbo, which includes a variable division encoder (ESD), a variable division decoder (DSD), and a denoising network (UNet). SD-Turbo is a real-time generation version based on the Stable Diffusion diffusion model architecture.

[0008] Optionally, the overall framework of the image compression model further includes a dual-branch encoding / decoding structure, wherein the dual-branch encoding / decoding structure includes an auxiliary encoder E. Aux Auxiliary decoder D Aux .

[0009] Optionally, the deep compression latent codec of the image compression model includes a deep compression encoder ge, a deep compression decoder gd, and an autoregressive entropy model, wherein the deep compression decoder gd is composed of multiple convolutional modules and upsampling layers, and the autoregressive entropy model adopts a hybrid entropy model that combines super-prior and autoregressive.

[0010] Optionally, the input to the deep compression decoder gd is the variable sub-encoder E. SD and the auxiliary encoder E Aux The latent space representation of the output is specifically represented as: F l =concat(E SD (x),E Aux (x)), where F l Let E represent the latent space representation, x represent the input image, and E represent the latent space representation. SD (x) represents the variable encoder E SD The output characteristics, E Aux (x) represents the auxiliary encoder E Aux The output features, concat() indicates feature aggregation along the channel dimension.

[0011] Alternatively, the deep compression latent encoding / decoding process can be represented as: y = g e (F l ), Among them, g e (F1) represents the deep compression encoding process of the latent feature representation F1. Represents the latent features after quantization. Perform deep compression and decoding process. This represents the quantization process, which converts a continuous signal into a discrete representation.

[0012] Optionally, the denoising network UNet consists of an encoder-decoder and skip connections, and includes convolutional layers, pooling layers, and activation function layers, specifically represented as: l0 = UNet(l T ), where lT The features are deeply compressed, UNet() represents the UNet network, and l0 represents the denoised output features.

[0013] Optionally, the denoised output features are fed into a decoder for image feature reconstruction. The decoder includes a variable-segment decoder (DSD) and an auxiliary decoder (D). Aux The specific process is as follows: in, This indicates that the auxiliary decoder decodes the quantized features. This represents the final reconstructed image output.

[0014] Optionally, the model loss function is a rate-distortion function, specifically: L = λ*D + R. Where D represents image distortion calculation, and x represents input features. Let d represent the reconstructed features, d be the image quality evaluation function, and R represent the overall frame compression bitrate, expressed using cross-entropy.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0016] The method provided in this application constructs a traffic video image dataset by collecting video image data in traffic scenes; defines the overall framework of an image compression model, a deep compression latent codec, and a denoising UNet network, while constructing a model loss function; trains the image compression model based on the traffic video image dataset to construct a traffic image compression model, and performs image compression processing through the traffic image compression model. This achieves the technical effect of improving compression performance while maintaining low computational complexity, and realizing high-quality, high-fidelity image reconstruction.

[0017] The above description is merely an overview of the technical solution of this application. To enable a clearer understanding of the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an image compression method based on a traffic scene provided in this application.

[0020] Figure 2 This is a schematic diagram of the overall framework of the image compression model in the image compression method based on traffic scenes provided in this application. Detailed Implementation

[0021] This application provides an image compression method based on traffic scenes, which addresses the technical problems of existing image compression technologies, such as computational complexity, poor compression performance, and inability to guarantee the consistency of image reconstruction. It achieves the technical effect of improving compression performance while maintaining low computational complexity, and realizing high-quality, high-fidelity image reconstruction.

[0022] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0023] like Figure 1 As shown, this application provides an image compression method based on traffic scenes, the image compression method based on traffic scenes includes:

[0024] Collect video image data in traffic scenarios and construct a traffic video image dataset.

[0025] Specifically, visible light cameras are installed at various traffic locations, including but not limited to major urban roads, intersections, roadside parking areas, and parking areas in key areas such as schools and hospitals. These locations encompass different traffic scenarios. For example, at an urban intersection, multiple visible light cameras are installed, each pointing in a different direction, to comprehensively record the flow of vehicles and pedestrians from all directions. These cameras are used to collect video images of the dynamic changes of vehicles, pedestrians, and other targets in the traffic scene over different time periods, creating a traffic video image dataset.

[0026] We define the overall framework of the image compression model, the deep compression latent codec, and the denoising UNet network, and construct the model loss function.

[0027] In one embodiment, such as Figure 2 As shown, the overall framework of the image compression model is based on SD-Turbo, which includes a variable encoder ESD, a variable decoder DSD, and a denoising network UNet. SD-Turbo is a real-time generation version based on the StableDiffusion diffusion model architecture.

[0028] Specifically, the overall framework of the image compression model is defined, and it is built upon SD-Turbo. SD-Turbo is a real-time generation version based on the Stable Diffusion model architecture. The diffusion model belongs to the generation model; it gradually adds noise to the data and then learns to recover the original data from the noisy data to generate new data samples. By using the SD-Turbo architecture as the foundation of the image compression model, the image compression processing speed can be improved while ensuring the generation quality, enabling fast real-time image generation.

[0029] The overall framework of the image compression model includes a variable-division encoder (ESD), a variable-division decoder (DSD), and a denoising network (UNet). The ESD encodes the input traffic video image into a representation in a latent space, converting high-dimensional image data into low-dimensional latent feature vectors. The latent space is a low-dimensional representation of the original data; similar data points cluster together in the latent space, better capturing the data's inherent structure and features. The DSD, in contrast to the ESD, decodes the features in the latent space to restore the image data, achieving image reconstruction. The UNet denoising network, the core network of the diffusion model, removes and repairs noise and distortion generated during compression and decoding, improving the quality of the reconstructed image.

[0030] The deep compression latent codec includes a deep compression latent encoder and a deep compression latent decoder. The deep compression latent encoder is used to further compress the latent spatial features output by the variable-resolution encoder (ESD) to achieve a lower image compression bit rate. The deep compression latent encoder employs a multi-layer convolutional neural network structure, extracting features by sliding convolutional kernels across the image and automatically learning local features. Through multi-layer convolution and pooling operations, including max pooling and average pooling, the feature map is downsampled, reducing data volume and computational cost, while simultaneously enhancing the translation invariance of the image compression model, gradually reducing the dimensionality of the feature map, and achieving a lower image compression bit rate.

[0031] The deep compression decoder corresponds to the deep compression encoder and is used to restore the deep-compressed features to a representation close to the original latent space. The deep compression decoder consists of multiple convolutional layers and upsampling layers. The upsampling layer acts as the opposite of the pooling layer, increasing the dimensionality of the feature map through interpolation and other operations, gradually restoring the image's detailed information. During the decoding process, the convolutional layers further process and refine the features, and combined with the upsampling layers, gradually reconstruct features similar to the original latent space.

[0032] The model loss function is an indicator used to measure the difference between the output of an image compression model and the actual result. By minimizing the loss function, the image compression model can optimize its own parameters and improve image compression performance.

[0033] By defining an overall framework to provide the structure for the image compression model, the model can perform encoding, compression, decoding, and denoising operations on traffic video images, thus achieving image compression. The deep compression latent codec achieves efficient image compression by extracting key image features through multi-layer convolution and pooling operations and performing deep compression. Furthermore, the denoising UNet network effectively removes noise and distortion generated during compression and decoding, improving the quality of the reconstructed image and enhancing the accuracy and clarity of the reconstructed traffic image. Simultaneously, by constructing a model loss function, the optimization of the image compression model is given a clear direction and objective, guiding the model to continuously adjust parameters to achieve the best image compression effect.

[0034] The image compression model is trained based on the traffic video image dataset to construct a traffic image compression model, and then the image compression is performed using the traffic image compression model.

[0035] Specifically, the image compression model is trained using a traffic video image dataset, which is divided into training, validation, and test sets. During training, an end-to-end approach is adopted. Traffic video images are input into the image compression model, where the variable decomposer (ESD) encodes the input traffic video images, converting them into a latent representation in the latent space. A deep compression encoder further compresses this latent representation to reduce the data volume, and a deep compression decoder decodes the compressed features to reconstruct features close to the original latent space representation. The variable decomposer (DSD) decodes the reconstructed features into a reconstructed image. Simultaneously, a denoising UNet network denoises the reconstructed image, improving image quality.

[0036] The image compression model calculates the difference between the output and the real image based on the model's loss function, and uses the backpropagation algorithm to propagate the error back to each layer of the image compression model, updating the model's parameters. The backpropagation algorithm is a gradient calculation method based on the chain rule, which can efficiently calculate the gradient of the loss function with respect to the model parameters, guiding the direction of parameter updates. The gradient represents the rate of change of the function at a certain point; by updating the parameters in the opposite direction of the gradient, the loss function gradually decreases.

[0037] During the training of the image compression model, the above process is iterated continuously until the performance of the image compression model on the validation set stabilizes, i.e., the model loss function value no longer decreases significantly, indicating that the image compression model has converged. After training, the trained traffic image compression model is evaluated using a test set. Evaluation metrics include peak signal-to-noise ratio (PSNR) and structural similarity index. PSNR measures the image reconstruction quality. It is calculated by comparing the mean squared error between the reconstructed image and the original image with the maximum possible signal power of the image. A higher PSNR value indicates better reconstructed image quality. The structural similarity index measures the similarity between two images in terms of brightness, contrast, and structure, with values ​​ranging from 0 to 1. Values ​​closer to 1 indicate greater similarity between the two images.

[0038] Based on the evaluation results of the test set, the image compression model is further optimized, and its output performance is validated using the validation set to improve its generalization ability. Through training and optimization, a high-performance traffic image compression model is obtained. This model is then used to compress new traffic images, achieving efficient image compression and high-quality image reconstruction, thus improving the accuracy and reliability of the compressed images.

[0039] Furthermore, the overall framework of the image compression model also includes a dual-branch encoding / decoding structure, wherein the dual-branch encoding / decoding structure includes an auxiliary encoder EAux Auxiliary decoder D Aux .

[0040] Specifically, the overall framework of the image compression model also includes a dual-branch encoding / decoding structure, and the dual-branch decoding structure includes an auxiliary encoder E. Aux Auxiliary decoder D Aux By employing auxiliary encoders and decoders, traffic image features can be extracted and reconstructed from different perspectives, improving the fidelity of the reconstructed image. The auxiliary encoder, using a convolutional neural network structure, performs auxiliary feature extraction on the input traffic image. It consists of multiple convolutional and pooling layers. The convolutional layers slide their kernels across the image, performing convolution operations to extract features at different levels. For example, lower-level convolutional layers extract basic features such as edges and colors, while higher-level convolutional layers extract more complex semantic features, such as vehicle shapes and pedestrian poses. The convolutional kernels slide across the image and perform dot products with local regions to extract features from those regions. The pooling layers downsample the feature maps output by the convolutional layers, reducing the size and computational cost of the feature maps. This also enhances the translation invariance of the image compression model. Translation invariance means that even when an object's position in the image changes slightly, the enhanced image compression model can still recognize the object. Pooling operations are either max pooling or average pooling. Max pooling extracts the maximum value within a local region as the output, while average pooling takes the average value within the local region as the output. An auxiliary encoder is used to process the data, obtaining auxiliary feature representations that complement the features extracted by the variable encoder.

[0041] The auxiliary decoder, corresponding to the auxiliary encoder, is used to decode the auxiliary features extracted by the auxiliary encoder. The auxiliary decoder consists of multiple deconvolutional layers and upsampling layers. The deconvolutional layers are transposed convolutional layers used to upsample the input feature map, increasing its size. Upsampling transforms a low-resolution feature map into a high-resolution one, thereby restoring image details. The deconvolutional layers learn transposed convolutional kernels and perform convolution operations between the input feature map and the kernels to achieve upsampling. The upsampling layers further expand the size of the feature map, gradually bringing it closer to the original image size.

[0042] By introducing a dual-branch encoding and decoding structure, the image compression model's ability to extract and reconstruct complex features of traffic images is improved, key information in the image is better preserved, and the quality and fidelity of the reconstructed image are enhanced.

[0043] Furthermore, the deep compression latent codec of the image compression model includes a deep compression encoder ge, a deep compression decoder gd, and an autoregressive entropy model, wherein the deep compression decoder gd is composed of multiple convolutional modules and upsampling layers, and the autoregressive entropy model adopts a hybrid entropy model that combines super-prior and autoregressive.

[0044] Specifically, the deep compression latent codec of the image compression model includes a deep compression encoder (ge), a deep compression decoder (gd), and an autoregressive entropy model. The deep compression encoder automatically learns local and global features of the image through convolutional and pooling layers. In the deep compression encoder, the convolutional layers use multiple convolutional kernels to perform convolution operations on the input image. Each convolutional kernel can extract a specific feature of the image, such as edges, texture, or color. Pooling layers are used to downsample the feature maps output by the convolutional layers, reducing the size and computational cost of the feature maps while enhancing the translation invariance of the image compression model. After multiple layers of convolution and pooling operations, the deep compression encoder compresses the input traffic scene image into a low-dimensional latent representation, which contains the key information of the image.

[0045] A deep compression decoder is used to reconstruct an approximate image of the input traffic scene image from the latent representation. The deep compression decoder is constructed from multiple convolutional modules and upsampling layers. The convolutional modules employ a convolutional neural network structure to perform feature transformations and refinement on the latent representation, gradually recovering the detailed information of the image. Each convolutional module contains multiple convolutional layers, batch normalization layers, and activation function layers. The batch normalization layer normalizes each batch of data, accelerating the training process of the image compression model and improving its stability. Activation functions, such as ReLU and Sigmoid, introduce nonlinear transformations to enhance the expressive power of the image compression model. Upsampling layers progressively upsample the low-resolution latent representation to a high-resolution image. Upsampling can be achieved using transposed convolution; by learning a set of transposed convolution kernels, the input feature map is convolved with the kernels to achieve upsampling.

[0046] The autoregressive entropy model employs a hybrid entropy model combining super-prior and autoregression to accurately and efficiently model the entropy of latent representations in compression, achieving more efficient lossless compression. The super-prior models the global statistical features of the latent representation. By introducing additional encoders and decoders, it learns the global statistical information of the latent representation. The super-prior encoder further compresses the latent representation, extracts its global features, and encodes them as a super-prior. The super-prior decoder reconstructs the prior distribution parameters of the latent representation, such as mean and variance, based on the super-prior. Autoregression captures the dependencies between spatially adjacent pixels, modeling the latent representation pixel by pixel. The autoregressive part uses autoregressive models such as pixel convolution or pixel recurrent neural networks to model the conditional probability of the latent representation. The hybrid entropy model combining super-prior and autoregression can fully utilize the global and local information of the latent representation, providing a more accurate model of the probability distribution of the latent representation. Compared to a single entropy model, the hybrid entropy model provides more accurate probability estimates, reduces the number of bits required for encoding, and improves the efficiency of image compression.

[0047] The deep compression encoder, deep compression decoder, and autoregressive entropy model in the deep compression latent codec work together to enable the image compression model to achieve more efficient lossless compression while ensuring image quality.

[0048] Furthermore, the input to the deep compression decoder gd is the variable sub-encoder E. SD and the auxiliary encoder E Auc The latent space representation of the output is specifically represented as: F l =concat(E SD (x),E Aux (x)), where F l Let E represent the latent space representation, x represent the input image, and E represent the latent space representation. SD (x) represents the variable encoder E SD The output characteristics, E Aux (x) represents the auxiliary encoder E Aux The output features, concat() indicates feature aggregation along the channel dimension.

[0049] Specifically, the input to the deep compression decoder gd is the variable sub-encoder E. SD and the auxiliary encoder E Auc The latent space representation of the output is specifically represented as: F l =concat(E SD (x),E Aux (x)), where F l Let E represent the latent space representation, x represent the input image, and E represent the latent space representation. SD (x) represents the variable encoder E SD The output characteristics, E Aux (x) represents the auxiliary encoder E Aux The output features, concat() indicates feature aggregation along the channel dimension.

[0050] For example, E SD (x) is a feature map of shape (C1, H, W), where C1 is the number of channels, H is the height, W is the width, and E is the height. Aux (x) is the feature map of shape (C2, H, W). Using the concat() function, E... SD (x) and E Aux (x) The two feature maps are concatenated along the channel dimension to obtain a new feature map F with shape (C1+C2, H, W). lThe latent space representation is used as a representation of the image. In a convolutional neural network, the feature dimension represents different feature channels, each capturing different features of the image. Through a concatenation operation, different types of features extracted by different encoders are integrated, making the latent space representation contain richer and more comprehensive image features. This improves the quality and fidelity of the reconstructed image and reduces information loss during compression.

[0051] Furthermore, the deep compression latent encoding / decoding process can be represented as: y = g e (F l ), Among them, g e (F l ) represents the latent feature representation F l Perform deep compression encoding process. Represents the latent features after quantization. Perform deep compression and decoding process. This represents the quantization process, which converts a continuous signal into a discrete representation.

[0052] Specifically, a deep compression encoder is used to represent the latent space F of the input. l Further compression processing, the deep compression potential encoding and decoding process can be represented as: y = g e (F l ), Among them, g e (F l ) represents the latent feature representation F l Perform deep compression encoding process. Represents the latent features after quantization. Perform deep compression and decoding process. This represents the quantization process, which converts a continuous signal y into a discrete representation. In image compression, continuous latent representations occupy a large amount of storage space and bandwidth, making direct and efficient transmission impossible. Quantization maps continuous values ​​to a finite set of discrete values, significantly reducing the data volume. Quantization methods include scalar quantization and vector quantization. Scalar quantization quantizes each latent value individually, dividing the continuous value range into several intervals, each interval corresponding to a discrete value. For example, uniform scalar quantization divides continuous values ​​into multiple equal-width intervals and maps the latent value to its corresponding discrete value based on the interval it belongs to. Vector quantization involves combining multiple latent values ​​into a vector, quantizing the combined vector, and dividing the vector space into several regions, each region corresponding to a discrete vector.

[0053] By employing a deep compression latent encoding and decoding process, efficient image compression and high-quality reconstruction are achieved, ensuring the visual quality of the images.

[0054] Furthermore, the denoising network UNet consists of an encoder-decoder and skip connections, and includes convolutional layers, pooling layers, and activation function layers, specifically represented as: l0 = UNet(l T ), where l T The features are deeply compressed, UNet() represents the UNet network, and l0 represents the denoised output features.

[0055] Specifically, the UNet denoising network is used to denoise and reconstruct a clear image during the diffusion process. The UNet consists of an encoder-decoder and skip connections. The encoder progressively extracts semantic features from the image, reducing the size of the feature map through downsampling to capture information at different levels. The decoder uses the semantic features extracted by the encoder to progressively recover the image's details, increasing the size of the feature map through upsampling to reconstruct an output with the same size as the original image. Skip connections directly pass features from different levels in the encoder to the corresponding layers in the decoder. By concatenating the output features of each convolutional layer in the encoder with the upsampled feature map in the decoder, the original structure and texture of the image are better restored. The UNet includes convolutional layers, pooling layers, and activation function layers. The convolutional layers perform convolution operations on the input feature map using convolution kernels. The convolution kernels are learnable parameters that are continuously adjusted during training to extract different features from the image. The UNet uses multiple convolution kernels of different sizes to extract multi-scale feature information, such as 3×3, 5×5, and 7×7. Pooling layers are used to downsample the feature maps output by convolutional layers, reducing their size. Max pooling is used to better preserve information in the image. Activation function layers can use the ReLU activation function to activate the output of the convolutional layers. The denoising network UNet is specifically represented as: l0 = UNet(l T ), where l T The features are deeply compressed, UNet() represents the UNet network, and l0 represents the denoised output features.

[0056] Single-step denoising is employed to achieve a single-step denoising diffusion process for the latent process, directly from the deeply compressed and noisy latent features. T Starting from this point, the denoising operation is completed in one step using the denoising network Unet, resulting in the denoised output feature l0. This greatly reduces the number of diffusion denoising steps, computational complexity, and inference time, while maintaining good denoising performance.

[0057] Furthermore, the denoised output features are fed into the decoder for image feature reconstruction. The decoder includes a variable-resolution decoder (DSD) and an auxiliary decoder (D). AuxThe specific process is as follows: in, This indicates that the auxiliary decoder decodes the quantized features. D represents the final reconstructed output image. SD This indicates a variable-resolution decoder.

[0058] Specifically, the denoised output features are input into the decoder for image feature reconstruction. The decoder includes a variable-segment decoder (DSD) and an auxiliary decoder (D). Aux The specific process is as follows: in, This indicates that the auxiliary decoder decodes the quantized features. D represents the final reconstructed output image. SD This indicates a variable-resolution decoder.

[0059] The denoised output features are input into the decoder for image feature reconstruction. The decoder works in concert with a variable divider decoder and an auxiliary decoder to improve the accuracy and quality of the reconstructed image.

[0060] Furthermore, the model loss function is a rate-distortion function, specifically: L = λ*D + R. Where D represents image distortion calculation, and x represents input features. Let d represent the reconstructed features, d be the image quality evaluation function, and R represent the overall frame compression bitrate, expressed using cross-entropy.

[0061] Specifically, a model loss function is constructed for model training. This model loss function is a rate-distortion function, comprising two parts: the distortion of the original image and the reconstructed image, and the overall frame compression bitrate. The bitrate can be selected and controlled by adjusting the coefficient parameters. A larger parameter results in less distortion in the reconstructed image of the trained image compression model, while a smaller parameter results in more distortion. By adjusting the parameters, reconstructed images with different compression ratios can be obtained. The model loss function is specifically L = λ*D + R, where D represents the image distortion calculation, specifically expressed as... Where x represents the input feature, Let represent the reconstructed features, and d be the image quality assessment function. Different image quality assessment functions such as MSE, PSNR, SSIM, and MS-SSIM can be used. A pixel-wise image quality assessment function is used to calculate the distortion of the original and reconstructed images. R represents the overall frame compression bitrate, expressed using cross-entropy.

[0062] By using the rate-distortion function as the loss function of the image compression model, and by comprehensively considering image distortion and compression bitrate, a clear optimization goal and comprehensive guidance are provided for the training, optimization, and performance evaluation of the image compression model. This enables the image compression model to reduce the compression bitrate as much as possible while ensuring image quality, improving transmission efficiency, and maintaining the consistency of image reconstruction.

[0063] In summary, the image compression method based on traffic scenes provided in this application has the following technical effects:

[0064] 1. SD-Turbo, a fast generation model based on the Stable Diffusion architecture, is optimized for real-time image generation, enabling the generation of high-quality images with minimal inference steps. Through the design of a deep compression latent codec and an autoregressive entropy model, traffic scene images can be efficiently compressed while maintaining a certain level of image quality, reducing the storage space and bandwidth required for image data storage and transmission.

[0065] 2. Add an auxiliary encoder-decoder to enable the image compression model to encode and decode image features from different angles, making full use of the various feature information of the image. The dual-branch encoding and decoding structure improves the image compression model's processing capability and adaptability for traffic scene images, further enhancing the fidelity of image reconstruction and the consistency of the reconstructed image.

[0066] 3. The introduction of the UNet denoising network to denoise the deeply compressed features can effectively reduce the noise and distortion generated during the compression process, improve the quality of the reconstructed image, and make the reconstructed traffic scene image clearer and more accurate.

[0067] The above description of the disclosed embodiments enables those skilled in the art to make or use 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 disclosed herein.

[0068] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. An image compression method based on traffic scenes, characterized in that, The method includes: Collect video and image data from traffic scenarios to construct a traffic video and image dataset; Define the overall framework of the image compression model, the deep compression latent codec, and the denoising UNet network, and construct the model loss function; The image compression model is trained based on the traffic video image dataset to construct a traffic image compression model, and then the image compression is performed using the traffic image compression model.

2. The image compression method based on traffic scenes as described in claim 1, characterized in that, The overall framework of the image compression model is based on SD-Turbo, which includes a variable division encoder (ESD), a variable division decoder (DSD), and a denoising network (UNet). SD-Turbo is a real-time generation version based on the Stable Diffusion diffusion model architecture.

3. The image compression method based on traffic scenes as described in claim 2, characterized in that, The overall framework of the image compression model also includes a dual-branch encoding / decoding structure, wherein the dual-branch encoding / decoding structure includes an auxiliary encoder E. Aux Auxiliary decoder D Aux .

4. The image compression method based on traffic scenes as described in claim 3, characterized in that, The deep compression latent codec of the image compression model includes a deep compression encoder ge, a deep compression decoder gd, and an autoregressive entropy model. The deep compression decoder gd is composed of multiple convolutional modules and upsampling layers, and the autoregressive entropy model adopts a hybrid entropy model that combines super-prior and autoregressive.

5. The image compression method based on traffic scenes as described in claim 4, characterized in that, The input to the deep compression decoder gd is the variable sub-encoder E. SD and the auxiliary encoder E Aux The latent space representation of the output is as follows: F l =concat(E SD (x),E Aux (x)), where F l Let E represent the latent space representation, x represent the input image, and E represent the latent space representation. SD (x) represents the variable encoder E SD The output characteristics, E Aux (x) represents the auxiliary encoder E Aux The output features, concat() indicates feature aggregation along the channel dimension.

6. The image compression method based on traffic scenes as described in claim 5, characterized in that, The deep compression underlying encoding and decoding process can be represented as: Among them, g e (F l ) represents the latent feature representation F l Perform deep compression encoding process. Represents the latent features after quantization. Perform deep compression and decoding process. This represents the quantization process, which converts a continuous signal into a discrete representation.

7. The image compression method based on traffic scenes as described in claim 6, characterized in that, The denoising network UNet consists of an encoder-decoder and skip connections, and includes convolutional layers, pooling layers, and activation function layers, specifically represented as: l0 = UNet(l T ), where l T The features are deeply compressed, UNet() represents the UNet network, and l0 represents the denoised output features.

8. The image compression method based on traffic scenes as described in claim 7, characterized in that, The method further includes: The denoised output features are fed into the decoder for image feature reconstruction. The decoder includes a variable-segment decoder (DSD) and an auxiliary decoder (D). Aux The specific process is as follows: in, This indicates that the auxiliary decoder decodes the quantized features. This represents the final reconstructed image output.

9. The image compression method based on traffic scenes as described in claim 1, characterized in that, The model loss function is a rate-distortion function, specifically: Where D represents image distortion calculation, and x represents input features. Let d represent the reconstructed features, d be the image quality evaluation function, and R represent the overall frame compression bitrate, expressed using cross-entropy.

Images