Image compression method based on traffic scene
By constructing a content-adaptive fusion module and a context-aware entropy module, combined with a uniform quantizer and a decoder, the problem of low image compression efficiency in traffic scenes in existing technologies is solved, achieving efficient and accurate image compression and reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-21
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies suffer from low image compression efficiency when processing large-scale, continuous traffic video surveillance data, increasing storage and transmission burdens, and failing to fully consider dynamic characteristics of traffic scenarios such as traffic flow and vehicle movement.
A content-adaptive fusion module and a context-aware entropy module are constructed. By combining a uniform quantizer and a decoder, the image compression process is optimized through feature extraction, quantization, and reconstruction. A state-space model and a convolutional neural network are used for feature extraction and reconstruction, and a loss function is constructed for model training.
It improves the efficiency and quality of image compression, ensures that the compressed image is highly consistent with the original image, reduces quantization errors, achieves efficient image storage and transmission, and maintains good reconstruction results.
Smart Images
Figure CN121661160A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image compression 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 and intersection locations to collect data, and using visual algorithms to analyze the collected data comprehensively, 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 has brought significant challenges to the storage and transmission of data captured by high-position cameras.
[0003] Image compression technology reduces redundant information in image data, thus enabling more efficient storage and transmission of data. Traditional image data compression methods mainly include JPEG, JPEG2000, and BPG. These methods primarily focus on compressing single images. However, when processing large-scale, continuous video surveillance data, they often suffer from low compression efficiency due to redundant information. Specifically, traditional methods cannot fully consider the dynamic characteristics of traffic scenarios, such as changes in traffic flow and vehicle movement, resulting in a still large data storage requirement after compression and low transmission efficiency, thereby increasing the burden of storage and transmission. Summary of the Invention
[0004] This application provides an image compression method based on traffic scenes, aiming to solve the technical problem that existing traffic scene image compression algorithms are usually based on fixed image feature extraction and compression strategies, which result in low data compression efficiency and increased storage and transmission burden when processing large-scale, continuous traffic video surveillance data.
[0005] This application discloses an image compression method based on traffic scenes. The method includes: acquiring video image data in traffic scenes to construct an original image dataset; constructing a content-adaptive fusion module, inputting the original image dataset into the content-adaptive fusion module for feature extraction and fusion, and outputting a fused feature representation; constructing a context-aware entropy module, inputting the fused feature representation into the context-aware entropy module for latent feature entropy modeling, and outputting a feature representation vector; inputting the feature representation vector into a uniform quantizer to obtain a quantized feature representation; inputting the quantized feature representation into a decoder for image feature reconstruction; constructing a loss function, and training a model based on the image feature reconstruction results and the original image dataset.
[0006] One or more technical solutions provided in this application have at least the following beneficial effects:
[0007] By acquiring video image data specifically for traffic scenarios, the constructed raw image dataset realistically reflects various dynamic changes and complex backgrounds in traffic scenes, providing high-quality input data for subsequent feature extraction and compression, ensuring the applicability and accuracy of the image compression method in real-world scenarios. A content-adaptive fusion module is constructed to adaptively extract and fuse features based on different image content, effectively preserving key visual information. This module combines global and local feature extraction, improving the reconstruction quality after image compression and ensuring a high degree of visual consistency between the compressed and original images. A context-aware entropy module is introduced, which models the latent feature entropy of the fused feature representation, enabling more accurate capture of the statistical distribution of various important features in the image. This module, through refined feature modeling, ensures that the compression process considers not only image... The compression process can be optimized based on the contextual information of the image, effectively reducing entropy and improving compression efficiency. The use of a uniform quantizer, combined with additive uniform noise quantization, reduces quantization errors and improves the quality of compressed data. Through quantized feature representation, image data is discretized, making it suitable for further compression and transmission while maintaining good reconstructibility of the original image. In the decoding stage, the quantized features are reconstructed by the decoder. The joint processing of the state-space model and convolutional neural network ensures that the details and structural information of the image are effectively restored, ensuring that too much information is not lost during the compression process, thus achieving high-quality image restoration. By constructing a loss function and combining it with the image feature reconstruction results, the model is trained. The optimization of the loss function during training can achieve an efficient balance between compression ratio and reconstructed image quality, adapting to different compression requirements.
[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description
[0009] Figure 1 This application provides a schematic flowchart of an image compression method based on a traffic scene.
[0010] Figure 2 This application provides a schematic diagram of the process for obtaining fused feature representation in an image compression method based on a traffic scene. Detailed Implementation
[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0012] like Figure 1 As shown in the figure, this application provides an image compression method based on traffic scenes, the method including:
[0013] Collect video and image data in traffic scenarios and construct the original image dataset.
[0014] In different traffic scenarios, appropriate traffic locations are selected for data collection, including high-risk or critical traffic areas such as urban roads, intersections, parking lots, schools, and hospitals. Each area has different traffic flow, road conditions, and time-of-day variations, ensuring the diversity and representativeness of the dataset. Visible light cameras or video surveillance equipment are installed to collect video image data in real time in these traffic scenarios. This video data includes the dynamic changes of targets such as vehicles and pedestrians, and is collected multiple times at different time periods, such as daytime and nighttime traffic conditions, and changes between weekdays and holidays. The collected image data is preprocessed to remove noise, calibrate lens distortion, and segment or extract frames from the images. Finally, the processed data is organized into an original image dataset according to different traffic scenarios, time periods, and other characteristics for subsequent model training and image compression.
[0015] A content-adaptive fusion module is constructed, and the original image dataset is input into the content-adaptive fusion module for feature extraction and fusion, and the fused feature representation is output.
[0016] The content-adaptive fusion module extracts global and local features from the original image data and combines their advantages to form image compression features adapted to traffic scenarios. This module includes two branches: a first branch and a second branch, which then fuse the features output from both. The first branch uses a state-space model to capture the global features of the image. Specifically, it uses visual state-space blocks and a 2D selective scanning layer to traverse the input image. The 2D selective scanning layer scans the image in four directions, generating an ordered sequence of blocks, each containing spatial domain information. This sequence is then processed by the state-space model to capture the global contextual information of the image. The second branch uses a convolutional neural network, such as ResNet, to extract fine-grained local features from the image. ResNet uses residual connections to enable the network to extract more detailed information at a deeper level, thereby capturing important local feature information in image compression. The outputs of the first and second branches are then fused through the fusion module. The fusion operation involves adding the two features pixel by pixel to combine the information from both features and obtain the final fused feature representation.
[0017] A context-aware entropy module is constructed. The fused feature representation is input into the context-aware entropy module to perform latent feature entropy modeling and output a feature representation vector.
[0018] The main function of the context-aware entropy module is to perform efficient entropy modeling by modeling the spatial and channel dependencies of image features. The core task of this module is to model the entropy of latent image features, thereby providing an effective representation and information content measure for image compression. In this module, the fused feature representation is first split along the channel dimension into several groups, each representing a sub-feature of the image. These groups are further processed through a state-space model. Specifically, a super-prior network is used to generate super-prior features, which are then connected with the already decoded group features to form new feature combinations for modeling. During the entropy modeling of latent feature representations, Gaussian distribution parameters are generated through the state-space model for calculating the cumulative distribution function of arithmetic encoding, thereby optimizing the entropy of the compression process and reducing the representation of redundant information.
[0019] The feature representation vector is input into a uniform quantizer to obtain a quantized feature representation.
[0020] In image compression, quantization is the process of converting continuous values into finite discrete values. The quantization step reduces the number of bits required for image representation and facilitates data storage and transmission during compression. A uniform quantizer is used to quantize the feature representation vector, ensuring that the output quantized feature representation has sufficient precision and effectively reduces redundant information during compression. The feature representation vector is input into the uniform quantizer, which discretizes the feature values using additive uniform noise. Specifically, the quantizer maps each value in the feature vector to a discrete quantization level or bucket, thereby reducing data precision and the number of bits required for storage. After uniform quantization, the quantized feature representation is obtained. This representation contains key information from the input data, but its redundancy is reduced by decreasing precision, providing a more efficient representation for subsequent image compression and transmission.
[0021] The quantized feature representation is input into the decoder for image feature reconstruction.
[0022] The decoder is used to recover image information lost during compression. It consists of multiple modules, including a state-space model and a convolutional neural network. The decoder's main task is to take the quantized feature representation as input, recover the original image features as closely as possible, and generate a reconstructed image that closely resembles the original. The obtained quantized feature representation is input to the decoder, which receives these features and recovers the latent features of the original image through reverse processing. Through these operations, the decoder restores the image's feature representation from the quantized feature representation and outputs the reconstructed image. Theoretically, the reconstructed image should be as similar as possible to the original image; however, due to quantization and information loss during compression, some distortion may exist.
[0023] A loss function is constructed, and the model is trained based on the image feature reconstruction results and the original image dataset.
[0024] The loss function is a tool used during training to evaluate model performance. Its main function is to measure the difference between the reconstructed image and the original image. By optimizing this loss function, model parameters are adjusted to minimize the trade-off between reconstruction distortion and compression bitrate. Model training uses the loss function, and minimizing the loss function adjusts network parameters to optimize image compression. During training, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and gradient descent is used to update the parameters, thereby gradually optimizing the compression effect and image reconstruction quality. The trained model can effectively achieve efficient image compression in traffic scenes while maintaining high image reconstruction quality as much as possible.
[0025] Furthermore, such as Figure 2 As shown, the content adaptive fusion module includes a first branch, a second branch, and a fusion module; the original image dataset is input into the first branch for global feature extraction to obtain a first output feature representation; the original image dataset is input into the second branch for local feature extraction to obtain a second output feature representation; the first output feature representation and the second output feature representation are input into the fusion module, and feature fusion is performed by pixel addition to obtain the fused feature representation.
[0026] The content-adaptive fusion module is used to extract feature information at different levels from the original image dataset and fuse this information to provide a more accurate and richer image representation. It includes the first branch, the second branch, and the fusion module.
[0027] The first branch is primarily responsible for capturing global features in the image. Global feature extraction focuses on the overall structure and macroscopic information of the image, such as the background and the overall outline of objects. This information is crucial for image compression and reconstruction. The original image dataset is input into the first branch, which employs a state-space model to capture global texture and color information. Specifically, the 2D selective scan layer of the state-space model traverses the spatial domain of the image, extracting global features by scanning and processing different regions. After processing, the first branch outputs a global feature representation, a high-dimensional feature vector or feature map representing the overall information of the image; this feature representation is called the first output feature representation.
[0028] The second branch focuses on capturing local features in the image. Local features refer to the details in the image, such as object edges, textures, and local structures. These local features are used for image restoration in image compression. The original image dataset is input into the second branch, which employs a convolutional neural network. Convolutional neural networks excel at extracting local information from images, performing multi-level processing through convolution operations to progressively extract local features from low to high levels. The feature representation output by the second branch is the local information in the image, called the second output feature representation.
[0029] The task of the feature fusion module is to combine the outputs from the first and second branches to obtain a more complete and detailed image representation, thereby enhancing the model's ability to understand image content. The first and second output feature representations are input into the fusion module. Pixel addition is a commonly used feature fusion method, which achieves fusion by summing the corresponding pixel values in the two feature representations. This method is simple and effective, effectively combining global and local information. The fused feature representation contains both overall and detailed information of the image, thus providing richer feature representations for subsequent image reconstruction and compression.
[0030] Furthermore, the original image dataset is input into the first branch for global feature extraction to obtain a first output feature representation, including:
[0031] The first branch is a state space model, which includes a visual state space block. The visual state space block traverses the spatial domain of the original image dataset through a 2D selective scanning layer to perform an ordered transformation of the non-causal visual image. The 2D selective scanning layer expands the feature block along four directions to generate four feature block sequences. The four feature block sequences are processed by the state space model to generate the first output feature representation.
[0032] The first branch employs a state-space model, where a visual state-space block is used to extract global features from the raw image data. The state-space model can process visual data and capture long-range dependencies in images without sacrificing the receptive field. This facilitates the extraction of global information during image compression and its use in subsequent feature fusion and reconstruction. The visual state-space block, by learning global information in the spatial domain, helps the network better understand the overall structure of the image.
[0033] Within the visual state space blocks, a 2D selective scanning layer traverses the spatial domain of the image and performs ordered transformations of the visual image. This process is crucial for extracting global information from the image, especially when processing visual data, where the ability to effectively scan and transform image content is essential. 2D selective scanning is a special scanning method that traverses the spatial region of the image along different directions and transforms it into ordered block sequences. This allows the network to scan the image from different angles and capture a variety of different visual information. Non-causal transformation means that this scanning process does not rely on traditional causal models but instead uses a more flexible approach to data stream processing, resulting in richer and more comprehensive image feature extraction. During image scanning, the goal of the selective scanning layer is to transform the image into multiple ordered block sequences, which are then further processed to extract visual features from the image.
[0034] In the 2D selective scanning layer, image data is scanned along four different directions, such as top, bottom, left, and right. Each scan generates a separate sequence of feature blocks, comprehensively covering different regions of the image and thus better extracting global features. These feature block sequences from the four directions are processed through a state-space model. The state-space model's role is to fuse information from different directions, capture the dependencies between different regions in the image, and generate a comprehensive feature representation. The final output is the first output feature representation, which contains global structural information of the image, aiding in further image compression and reconstruction.
[0035] Furthermore, the 2D selective scanning layer is represented as follows:
[0036] ;
[0037] ;
[0038] ;
[0039] ;
[0040] in, , All of these are learnable model parameters, and LN() represents the layer normalization function. () represents the activation function, specifically the SiLU activation function; A represents the feature representation generated by the activation function. This represents the element-wise multiplication operation. Indicates input, Indicates the output; () indicates a 2D selective scan operation, specifically represented as:
[0041] ;
[0042] ;
[0043] ;
[0044] Where V = {1, 2, 3, 4} is a set of four different scanning directions. Indicates the current specific scanning direction , Indicates direction Above input Perform scan expansion and output The operation is passed to the state space model for execution. Indicates to Perform state-space model calculation operations. For output, This means merging the outputs from all directions. This indicates the result of the merge.
[0045] It is the original image input data, processed... After a linear transformation of the weight matrix, the layer normalization function LN() and the activation function are applied. () is used for processing, and at the same time, input data is used. It also undergoes a second linear transformation And it is processed through layer normalization and activation functions. Next, element-wise multiplication is performed to... (The output from the first step) is fused with the feature representation A generated by the activation function. The fused features are then combined with... We perform a weighted sum to obtain .at last, After layer normalization and weighting, the final output features are obtained. This processing flow helps state-space models better capture global structural information in images and provides richer feature representations for image compression and reconstruction tasks.
[0046] () is a key operation used to perform spatial domain scanning processing on image data in a 2D selective scanning layer to generate an ordered feature representation. This operation uses four different scanning directions and fuses the scanning results from each direction.
[0047] In each specific scan direction v∈V, where V={1,2,3,4}, for the input features Scanning expansion is performed to facilitate local processing of spatial information in the image and ensure that information in each direction is fully extracted. Expanded features Then it is passed to the state space model This model is used to capture the global structural features of an image, and further processing of these features is performed through model calculations. Finally, the scan results in all four directions are obtained. Combined together, they form the final output features. The merge operation is performed through Implemented by this function, the operation fuses information from different scanning directions to ensure that the global features of the image are fully represented. This process effectively extracts complex structural information from the image by performing multi-directional scanning processing on the spatial domain of the image and combining state-space model calculations with feature fusion.
[0048] Furthermore, the second branch is a convolutional neural network, and the convolutional module uses a ResNet network.
[0049] The second branch uses a convolutional neural network for local feature extraction. The convolutional module uses a ResNet network. By introducing a residual learning mechanism, ResNet solves the common gradient vanishing and overfitting problems in deep networks, thereby improving the feature extraction effect. By stacking multiple residual blocks, ResNet can effectively capture detailed features in the image.
[0050] Furthermore, the fusion module is represented as:
[0051] ;
[0052] in, For feature representation, This is the first output feature representation. This represents the second output feature.
[0053] The fusion module combines the first and second output feature representations to generate the final fused feature representation. Specifically, fusion is performed by pixel addition, meaning the feature values at corresponding positions in both representations are added together to generate the fused feature representation. This fused feature representation combines global and local information and serves as input for subsequent steps.
[0054] Furthermore, inputting the fusion feature representation into the context-aware entropy module for latent feature entropy modeling and outputting a feature representation vector includes:
[0055] Defining the fusion feature representation as a latent feature representation, splitting the latent feature representation along the channel dimension into S groups, denoted as { ,..., }; processing through a state space model, denoted as:
[0056] ;
[0057] ;
[0058] = + ;
[0059] where is a learnable parameter in the connection process, [ ] represents the channel connection operation, includes two parts: a hyperprior encoder and a hyperprior decoder, represents the features of the decoded groups, for s < i, i represents the currently decoded group, s < i means the decoded groups among these S groups; where represents performing the state space model calculation operation on , represents the combination of the hyperprior and the decoded features, represents the calculation result of the state space model; where , is .
[0060] Defining the fusion feature representation as a latent feature representation, and then splitting it according to the channel dimension, denoted as { ,..., }, this splitting method helps to better process each feature group in subsequent steps and is also convenient for parallel processing or multi-stage decoding.
[0061] The split feature groups are processed through a state space model, and at the same time, a hyperprior network is combined to enhance the effect of entropy modeling. The specific steps are as follows: Processing the split features through the hyperprior network, which includes two parts: a hyperprior encoder and a hyperprior decoder. First, the hyperprior network generates a hyperprior , and then combines the decoded feature group with the hyperprior through the channel connection operation [ ] to form .
[0062] Will The input is given to the state-space model for computation, and the result is... , This represents the computational process of the state-space model, which is related to... Perform the operation and return the result, while also returning the original input. Adding it to the output as a residual connection, similar to the skip connection in ResNet, helps maintain the integrity of information and enhances the stability of the model.
[0063] The output of the state-space model Input to feedforward neural network In the process of generating Gaussian distribution parameters Used to describe the potential distribution of data. It is a learnable parameter of a feedforward neural network. The representation layer normalization operation ultimately yields These are Gaussian distribution parameters used to model latent features.
[0064] The super-prior network is used to generate Gaussian distribution parameters for latent features, helping to capture latent structural information in the image. The super-prior network consists of a super-prior encoder and a super-prior decoder. The encoder generates the super-prior parameters, and the decoder uses these parameters to adjust the image reconstruction process. The state-space model is used to capture the dynamic relationships and spatiotemporal structure in the image data. In this method, the state-space model enhances the expressive power of image features through processing input features and residual connections. Gaussian distribution parameters. Used to describe the probability distribution of image data, it helps generate smoother and more ordered latent features, thereby optimizing compression performance.
[0065] Furthermore, the uniform quantizer uses additive uniform noise for quantization.
[0066] The function of a uniform quantizer is to quantize input features, converting them into discrete quantized values. Specifically, because quantization is non-differentiable, resulting in a piecewise function with a derivative of 0 at non-boundary points (and truncating backpropagation), and having no derivative at boundary points, additive uniform noise is used instead of quantization during model training. Specifically, uniform noise is added during training to approximate differentiability for backpropagation optimization. During inference, rounding quantization is performed. This operation helps control quantization errors and data discreteness during compression. The introduction of additive uniform noise balances quantization errors and compression ratios, making the quantized features more suitable for subsequent image reconstruction while reducing negative impacts on image quality.
[0067] Furthermore, the decoder is implemented using a state-space model and a convolutional neural network.
[0068] The decoder is responsible for reconstructing the quantized features. This reconstruction process combines a state-space model and a convolutional neural network to achieve more efficient and accurate image restoration. Specifically, the decoder extracts local features of the image through a convolutional neural network, while the state-space model is responsible for processing global information and spatiotemporal relationships in the image, thereby achieving more detailed and accurate image reconstruction.
[0069] Furthermore, this includes:
[0070] The loss function is expressed as:
[0071] ;
[0072] ;
[0073] Where L represents the loss function, This represents an adjustable parameter that controls the weight between image distortion and compression bitrate, where D represents image distortion and x represents the input features. The reconstructed features are represented by d(), which is the image quality evaluation function calculated pixel by pixel, and R represents the overall frame compression bitrate, expressed using cross-entropy.
[0074] The loss function is part of the image compression model used to train the image compression model. It aims to balance the two major goals in image compression: the distortion of image reconstruction and the bit rate of the compressed image. The image distortion D measures the difference between the input image and the reconstructed image. The goal is to make the reconstructed image as close as possible to the original image and reduce the distortion and loss of image details caused by compression. The compression bit rate R measures the number of bits or the amount of data required in the compression process. In the context of image compression, a lower compression bit rate means a better compression effect, but it may lead to a decrease in image quality. Therefore, the goal is to reduce the bit rate as much as possible while maintaining sufficient image quality.
[0075] It is an adjustable hyperparameter used to balance the weights between image distortion D and compression bitrate R. The value of can control the model's preference for distortion and compression bitrate during training. If the image size is large, the model tends to reduce distortion and prioritize maintaining image quality; if... If the bit rate is lower, the model will prioritize the compression bit rate, thus accepting a higher level of distortion to reduce the amount of data after compression.
[0076] This loss function combines two objectives: image quality and compression ratio. Therefore, it is a typical multi-objective optimization problem. In practical applications, it is necessary to adjust the balance between these two objectives through alternating training to meet different compression and image restoration requirements, thereby achieving more efficient and higher-quality image compression.
[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. An image compression method based on traffic scenes, characterized in that, The method includes: Collect video and image data in traffic scenarios and construct a raw image dataset; A content-adaptive fusion module is constructed, and the original image dataset is input into the content-adaptive fusion module for feature extraction and fusion, and the fused feature representation is output. Construct a context-aware entropy module, input the fused feature representation into the context-aware entropy module to perform latent feature entropy modeling, and output a feature representation vector; The feature representation vector is input into a uniform quantizer to obtain the quantized feature representation; The quantized feature representation is input into the decoder for image feature reconstruction. A loss function is constructed, and the model is trained based on the image feature reconstruction results and the original image dataset.
2. The image compression method based on traffic scenes as described in claim 1, characterized in that, The content adaptive fusion module includes a first branch, a second branch, and a fusion module; The original image dataset is input into the first branch for global feature extraction to obtain the first output feature representation; The original image dataset is input into the second branch for local feature extraction to obtain the second output feature representation; The first output feature representation and the second output feature representation are input into the fusion module, and feature fusion is performed by adding pixels to obtain the fused feature representation.
3. The image compression method based on traffic scenes as described in claim 2, characterized in that, The original image dataset is input into the first branch for global feature extraction to obtain a first output feature representation, including: The first branch is a state-space model, which contains visual state-space blocks; The visual state space block traverses the spatial domain of the original image dataset through a 2D selective scanning layer to perform ordered transformation of non-causal visual images; The 2D selective scanning layer expands feature blocks in four directions to generate four feature block sequences. These four feature block sequences are processed by the state space model to generate the first output feature representation.
4. The image compression method based on traffic scenes as described in claim 3, characterized in that, The 2D selective scanning layer is represented as follows: ； ； ； ； in, , All of these are learnable model parameters, and LN() represents the layer normalization function. () represents the activation function, specifically the SiLU activation function; A represents the feature representation generated by the activation function. This represents the element-wise multiplication operation. Indicates input, Indicates the output; () indicates a 2D selective scan operation, specifically represented as: ； ； ； Where V = {1, 2, 3, 4} is a set of four different scanning directions. Indicates the current specific scanning direction , Indicates direction Above input Perform scan expansion and output The operation is passed to the state space model for execution. Indicates to Perform state-space model calculation operations. For output, This means merging the outputs from all directions. This indicates the result of the merge.
5. The image compression method based on traffic scenes as described in claim 2, characterized in that, The second branch is a convolutional neural network, and the convolutional module uses a ResNet network.
6. The image compression method based on traffic scenes as described in claim 2, characterized in that, The fusion module is represented as follows: ； in, For feature representation, This is the first output feature representation. This represents the second output feature.
7. The image compression method based on traffic scenes as described in claim 1, characterized in that, The fused feature representation is input into the context-aware entropy module for latent feature entropy modeling, and the output feature representation vector includes: The fused feature representation is defined as a latent feature representation, and the latent feature representation is split along the channel dimension into S groups, represented as { ,..., }; Processed using a state-space model, it can be represented as: ； ； = + ； Among them, is a learnable parameter during the connection process, [ ] represents the channel connection operation, and the hyperprior is generated by using a hyperprior network, and the hyperprior network includes two parts: a hyperprior encoder and a hyperprior decoder. represents the features of the decoded groups. For s < i, where i represents the currently decoded group, s < i means the groups that have been decoded among these S groups; in, Indicates to Perform state-space model calculation operations. This represents a combination of prior and decoded features. This represents the computational results of the state-space model; in, , yes .
8. The image compression method based on traffic scenes as described in claim 1, characterized in that, The uniform quantizer uses additive uniform noise for quantization.
9. The image compression method based on traffic scenes as described in claim 1, characterized in that, The decoder is implemented using a state-space model and a convolutional neural network.
10. The image compression method based on traffic scenes as described in claim 1, characterized in that, include: The loss function is expressed as: ； ； Where L represents the loss function, This represents an adjustable parameter that controls the weight between image distortion and compression bitrate, where D represents image distortion and x represents the input features. The reconstructed features are represented by d(), which is the image quality evaluation function calculated pixel by pixel, and R represents the overall frame compression bitrate, expressed using cross-entropy.