A compression method of video data in a vehicle networking scene

Abstract

The application provides a video data compression method in a vehicle networking scene, and relates to the technical field of image compression. A double-branch structure is designed in an image encoder, that is, a branch of a general feature extraction model and a branch of a vehicle networking feature extraction model, so as to solve the problem that a model trained by an existing learning type image compression algorithm does not match the characteristics of video data in the vehicle networking field, and the quality of decompressed video data is significantly improved. The application uses a lightweight fine-tuning technology, fine-tunes parameters of the general feature extraction model based on an existing vehicle networking public data set, so that a model suitable for the characteristics of vehicle networking video data can be quickly trained, the amount of parameters to be learned is greatly reduced, and the model training efficiency is improved. The hybrid architecture designed by the application can capture local texture features and global context dependency of an image, significantly reduces the average bit rate of pixels, and improves the storage efficiency of vehicle networking video data.

Description

Technical Field

[0001] This invention relates to the field of image compression technology, specifically a method for compressing video data in a vehicle networking scenario. Background Technology

[0002] With the rapid development of intelligent connected vehicle technology, modern vehicles are equipped with thousands of sensors. These sensors generate massive amounts of signal data in real time, covering multiple aspects such as vehicle status, driving behavior, and environmental perception. The types of sensor data are complex and diverse, and the data acquisition frequency has gradually increased from seconds to milliseconds, leading to a significant increase in data generation rate. This large-scale generation of data poses a severe challenge to enterprises' data processing and storage capabilities, especially in vehicle-to-everything (IoV) systems, where vehicle communication and data storage systems need to be able to handle petabytes of data. This data comes from a wide range of sources, including in-vehicle infotainment systems, vehicle hardware, user behavior data, and third-party partners, and is both massive in volume and has high real-time requirements.

[0003] Taking a six-lens in-vehicle camera as an example, this device captures 30 high-resolution images per second, with each image having a resolution of 1920x1080, resulting in a data volume of approximately 1.12 GB / s. In connected vehicle applications, the real-time transmission and storage of such high-resolution video data places extremely high demands on bandwidth and storage space, especially in application scenarios such as autonomous driving, traffic flow analysis, and road monitoring, where large amounts of image data need to be stored for extended periods to support intelligent analysis and decision-making. Therefore, effectively compressing this image data to reduce storage and transmission costs has become a significant challenge for database systems.

[0004] Currently, traditional video compression standards such as the H.26x series are relatively mature in terms of compression efficiency. They employ a hybrid video coding framework, significantly reducing video data redundancy through techniques such as intra-frame and inter-frame prediction, integer discrete cosine transform (DCT), adaptive entropy coding (CABAC and CAVLC), and deblocking filters. While maintaining high visual quality, the H.26x series significantly reduces bitrate requirements in video storage and transmission, and is widely used in streaming media, high-definition video, and video conferencing. In recent years, neural network-based image compression technology has shown great potential. Learning-based image compression methods, represented by Deep ImagePrior and Neural Image Compression, model the statistical distribution of image data using deep learning models, enabling the compression algorithm to adapt to image features and retain more detail information while achieving higher compression ratios. These methods can achieve higher quality image restoration with limited data transmission bandwidth, effectively reducing storage space requirements.

[0005] However, the two methods mentioned above are general-purpose video compression methods, and their video compression effect is not significant in the field of vehicle networking. Therefore, developing a compression scheme suitable for high-resolution video data in vehicle networking is of great significance, thereby effectively reducing the storage and transmission costs of in-vehicle camera image data.

[0006] Traditional general-purpose compression algorithms (such as H.264 / H.265), while widely used and highly efficient in video compression, exhibit significant limitations in Internet of Vehicles (IoV) scenarios. These traditional algorithms primarily achieve compression through techniques such as inter-frame prediction, intra-frame prediction, discrete cosine transform (DCT), and entropy coding. However, in high-resolution and large-data-volume scenarios, especially with continuous high-frame-rate video data generated by in-vehicle cameras, traditional compression techniques often struggle to further improve compression rates while maintaining video quality. This results in excessively high bandwidth and storage requirements for real-time transmission and storage, failing to meet the demands of practical applications.

[0007] Neural network methods utilize deep learning models to model the statistical properties of image data, maintaining high image quality at higher compression ratios. However, most current neural network compression models are trained on general image datasets (such as ImageNet), which contain a wide range of scenes and content, but differ significantly from the characteristics of image data in connected vehicle (V2V) scenarios. Because V2V data has a different distribution and entropy encoding than traditional data, directly using neural network compression models trained on general datasets like ImageNet often fails to achieve optimal compression results for V2V video data. These models lack adaptability to the specific data distribution in the V2V environment, preventing them from fully utilizing the characteristics of in-vehicle camera images and making it difficult to strike a balance between efficient compression and high-quality reconstruction. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to propose a method for compressing video data in a vehicle-to-everything (V2X) scenario, comprising:

[0009] Step 1: The original image data is encoded using an image encoder to obtain the feature representation of the original image data;

[0010] Step 2: Add a uniformly distributed noise to the feature representation of the original image data to obtain the feature representation after noise addition. By introducing the cumulative distribution function CDF, the feature representation after noise addition is quantized to obtain the target quantized feature representation. Then, based on the target quantized feature representation, the probability distribution after quantization is calculated.

[0011] Step 3: Perform entropy encoding on each quantized feature representation in the target quantized feature representation to obtain the encoded feature representation, and then calculate the probability distribution after entropy encoding;

[0012] Step 4: Perform entropy decoding on the encoded feature representation to obtain the recovered quantized feature representation, and then perform dequantization on the recovered quantized feature representation to obtain the recovered continuous feature representation;

[0013] Step 5: Decode the recovered continuous feature representation through CNN deconvolution and attention mechanism in the image decoder to obtain reconstructed image data. Then, based on the reconstructed image data, calculate the probability that the original image data is successfully reconstructed given the target quantized feature representation.

[0014] Step 6: Calculate the optimization objective function based on the quantized probability distribution, the probability of successful reconstruction of the original image data given the target quantized feature representation, and the probability distribution after entropy encoding;

[0015] Step 7: Update the parameters of the image encoder and the image decoder according to the optimization objective function;

[0016] Step 8: Acquire multiple raw image data, repeat steps 1 to 7 until the number of repetitions of steps 1 to 7 reaches the preset number of rounds, and obtain the trained image encoder and image decoder. Use the trained image encoder to compress the image through the processes of image encoding, quantization and entropy encoding.

[0017] Optionally, step 1 specifically includes:

[0018] The image encoder encodes the original image data using multiple basic blocks. Specifically, in the first basic block, the original image data is used as input. A general feature extraction model processes the original image data to obtain general image feature information, while a vehicle-to-everything (V2X) feature extraction model processes it to obtain V2X image feature information. The general image feature information and the V2X image feature information are then fused to obtain a combined feature representation of the first basic block. This combined feature representation is used as input to the next basic block. Similarly, the general feature extraction model and the V2X feature extraction model process the combined feature representation to obtain combined general image feature information and combined V2X image feature information. These combined feature representations are then fused to obtain the combined feature representation of the basic block. This combined feature representation is used as input to the next basic block. This process of processing and feature fusion using the general feature extraction model and the V2X feature extraction model is repeated until the last basic block. The combined feature output of the last basic block is the feature representation of the original image data.

[0019] Optionally, the general feature extraction model is obtained in the following way:

[0020] Image data from the general dataset ImageNet is input into the hybrid Transformer-CNN architecture model to obtain general feature information of the predicted images. Based on the general feature information of the predicted images and the general feature information of the real images, the loss value is calculated. The model parameters of the hybrid Transformer-CNN architecture model are updated based on the loss value until the loss value is less than a preset threshold, thus obtaining the general feature extraction model.

[0021] The vehicle-to-everything (V2X) feature extraction model is obtained by fine-tuning the general feature extraction model based on existing publicly available V2X datasets using lightweight fine-tuning techniques.

[0022] Optionally, in step 2, the probability distribution after quantization is calculated based on the target quantized feature representation, specifically through the following formula:

[0023]

[0024] Where x represents the original image data, This represents the quantitative feature representation of the target. This represents the quantized probability distribution. These represent the parameters of the image encoder, where U represents a uniform distribution, and y... iThis represents the i-th feature representation after adding noise. Indicates y i Quantized feature representation.

[0025] Optionally, in step 3, the probability distribution after entropy encoding is calculated, specifically using the following formula:

[0026]

[0027] in, This represents the quantitative feature representation of the target. This represents the probability distribution after entropy encoding. The parameters represent the entropy encoder. for The i-th component, Indicates that in a given Under the condition of y i The probability distribution, The i-th feature in the feature representation after adding noise is the quantized feature representation.

[0028] Optionally, in step 5, based on the reconstructed image data, the probability that the original image data can be successfully reconstructed given the target quantized feature representation is calculated, specifically through the following formula:

[0029]

[0030] Where x represents the original image data, This represents the quantitative feature representation of the target. θ represents the probability that the original image data can be successfully reconstructed given the target quantized feature representation. g The parameters represent the image decoder. Let λ represent the reconstructed image data, where λ is a hyperparameter used to measure the significance of the mean square error between the original and reconstructed image data. This represents the mean square error between the original image data and the reconstructed image data.

[0031] Optionally, step 6 can be implemented using the following formula:

[0032]

[0033] Where x represents the original image data, Let q represent the target quantization feature representation, and p represent the vehicle-to-everything (V2X) feature extraction model. x Represents the distribution of the original image data. This represents the true probability distribution after quantization. To measure q and The KL divergence of the differences between them This represents the quantized probability distribution. This represents the probability distribution after entropy encoding. This represents the probability that the original image data can be successfully reconstructed given the target quantized feature representation, and const represents the pre-set residual.

[0034] The beneficial effects of adopting the above technical solution are as follows:

[0035] This invention designs a dual-branch structure in the image encoder, namely a branch for a general feature extraction model and a branch for a vehicle-to-everything (V2X) feature extraction model. This solves the problem of mismatch between the models trained by existing learning-based image compression algorithms and the characteristics of video data in the V2X field, significantly improving the quality of decompressed video data. Utilizing lightweight fine-tuning techniques, based on existing publicly available V2X datasets, the parameters of the general feature extraction model are fine-tuned. This invention can quickly train a model suitable for the characteristics of V2X video data, greatly reducing the number of parameters to be learned and improving model training efficiency, especially when processing large amounts of V2X video data. The hybrid architecture designed in this invention can capture local texture features and global contextual dependencies of images, significantly reducing the average bit rate per pixel and improving the storage efficiency of V2X video data. This invention is unaffected by the complex and varied content of V2X video and maintains high stability regardless of whether it is in rural, urban, or other environments. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a method for compressing video data in a vehicle-to-everything (V2X) scenario according to an embodiment of the present invention.

[0037] Figure 2 This is a flowchart illustrating another method for compressing video data in a vehicle-to-everything (V2X) scenario according to an embodiment of the present invention. Detailed Implementation

[0038] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0039] To address the problems of existing technologies, the purpose of this invention is to provide a highly efficient compression method specifically for video data in the Internet of Vehicles (IoV) scenario, meeting the stringent requirements for video data transmission and storage in IoV environments. This invention leverages the advantages of neural networks in feature extraction and data distribution adaptation, enabling the compression algorithm to be optimized based on the specific distribution of in-vehicle camera data. Specifically, this invention introduces lightweight fine-tuning techniques such as LoRA (Low-Rank Adaptation) and Adapter Layers, allowing the neural network model to adaptively compress video data in IoV scenarios without significantly increasing computational costs. This lightweight fine-tuning method is particularly effective when vehicle computing resources are limited, optimizing model performance without significantly increasing hardware burden, thereby compressing in-vehicle camera data.

[0040] Specifically, this invention provides a method for compressing video data in a vehicle-to-everything (V2X) scenario, combined with... Figure 1 and Figure 2 This may include the following steps:

[0041] Step 1: The original image data is encoded using an image encoder to obtain the feature representation of the original image data;

[0042] The original image data is encoded using multiple basic blocks in the image encoder. Specifically, in the first basic block, the original image data is used as the input. The original image data is processed by a general feature extraction model. Specifically, the input features are processed by convolution weights W and ReLU activation function to obtain general image feature information. At the same time, the original image data is processed by a vehicle networking feature extraction model to obtain vehicle networking image feature information. The general image feature information and vehicle networking image feature information are fused together using Hout to obtain the combined feature representation of the first basic block, which is the hidden representation Hin.

[0043] The combined feature representation of the first basic block is used as the input of the next basic module. Similarly, the combined feature representation of the input is processed by the general feature extraction model and the vehicle network feature extraction model to obtain the general image feature information and the vehicle network image feature information of the combined feature representation. Feature fusion is performed on the general image feature information and the vehicle network image feature information of the combined feature representation to obtain the combined feature representation of the basic block. The combined feature representation of the basic block is used as the input of the next basic block. Similarly, the processing and feature fusion operations by the general feature extraction model and the vehicle network feature extraction model are repeated until the last basic block. The combined feature output of the last basic block is the feature representation of the original image data.

[0044] The general feature extraction model is obtained in the following way:

[0045] Image data from the general dataset ImageNet is input into the hybrid Transformer-CNN architecture model to obtain general feature information of the predicted images. Based on the general feature information of the predicted images and the general feature information of the real images, the loss value is calculated. The model parameters of the hybrid Transformer-CNN architecture model are updated based on the loss value until the loss value is less than a preset threshold, thus obtaining the general feature extraction model.

[0046] The core of the hybrid Transformer-CNN architecture model involves combining convolutional neural networks (CNNs) and Transformer modules. The CNN part employs multi-level residual convolutional blocks to extract local texture features of the image (such as edges and color distribution). The Transformer part uses a multi-head self-attention mechanism to model the global contextual dependencies of the image. Furthermore, skip connections are used to fuse the output features of the CNN and Transformer to generate the final latent representation.

[0047] The described vehicle-to-everything (V2X) feature extraction model is derived by fine-tuning the general feature extraction model based on existing publicly available V2X datasets, such as KITTI, using lightweight fine-tuning techniques. Specifically, lightweight fine-tuning techniques such as LoRA (Lower-Rank Adaptation) and Adapter Layer can be used. These techniques fine-tune the convolutional weights through additional low-rank matrices WA and WB, achieving adaptation to specific data distributions with minimal computational cost. This branch outputs V2X-specific features.

[0048] In image compression, the main purpose of quantization is to convert continuous numerical values ​​into discrete numerical values, thus facilitating entropy coding. Since continuous numerical values ​​cannot be directly used for finite symbol coding, they must first be discretized. During quantization, rounding or floor selection is generally used. However, because continuous values ​​become discrete in this floor selection step, it becomes impossible to optimize the quantization process using standard backpropagation during neural network training. In this scheme, uniform noise is added to the continuous values ​​before quantization to avoid gradient blocking caused by discontinuous values. The specific quantization process is shown in step 2.

[0049] Step 2: Add a uniformly distributed noise to the feature representation of the original image data to obtain the feature representation after adding noise. By introducing the cumulative distribution function CDF, the feature representation after adding noise is quantized to obtain the target quantized feature representation.

[0050] Then, based on the target quantized feature representation, the quantized probability distribution is calculated, specifically through the following formula:

[0051]

[0052] Where x represents the original image data, This represents the quantitative feature representation of the target. This represents the quantized probability distribution. These represent the parameters of the image encoder, where U represents a uniform distribution, and y... i This represents the i-th feature representation after adding noise. Indicates y i Quantized feature representation.

[0053] Step 3: Perform entropy encoding on each quantized feature representation in the target quantized feature representation to obtain the encoded feature representation;

[0054] Then, the probability distribution after entropy encoding is calculated, specifically through the following formula:

[0055]

[0056] in, This represents the quantitative feature representation of the target. This represents the probability distribution after entropy encoding. The parameters represent the entropy encoder. for The i-th component, Indicates that in a given Under the condition of y i The probability distribution, The i-th feature in the feature representation after adding noise is the quantized feature representation.

[0057] The purpose of entropy coding is to minimize the average code length for each symbol. The prerequisite for the optimal solution is obtaining the true probability distribution of the symbols; if an incorrect actual distribution is used, the average code length will increase. Entropy coding encodes the quantized features into a compact binary representation (e.g., "0110...1110"), further compressing the data volume. This part of the processing ensures that image data is transmitted with minimal bandwidth usage during transmission.

[0058] Step 4: Perform entropy decoding on the encoded feature representation to obtain the recovered quantized feature representation. These features are still discrete quantized values. Dequantize the recovered quantized feature representation to obtain the recovered continuous feature representation, in order to restore the distribution of the original features as much as possible. The final output after dequantization is the recovered continuous feature representation, which is used as input to the image decoder for further image reconstruction.

[0059] Step 5: Decode the recovered continuous feature representation through CNN deconvolution and attention mechanism in the image decoder to obtain the reconstructed image data;

[0060] Combination Figure 2 The image decoder consists of multiple basic blocks (symmetrical to the encoder structure), which reconstruct the feature representation layer by layer, restoring it back to a form close to the original input image. The output reconstructed image data should retain as much key information as possible from the original image to meet the image quality requirements in connected vehicle scenarios.

[0061] Furthermore, based on the reconstructed image data, the probability that the original image data can be successfully reconstructed given a target quantized feature representation is calculated, specifically through the following formula:

[0062]

[0063] Where x represents the original image data, This represents the quantitative feature representation of the target. θ represents the probability that the original image data can be successfully reconstructed given the target quantized feature representation. g The parameters represent the image decoder. Let λ represent the reconstructed image data, where λ is a hyperparameter used to measure the significance of the mean square error between the original and reconstructed image data. This represents the mean square error between the original image data and the reconstructed image data.

[0064] Step 6: Based on the quantized probability distribution, the probability of successful reconstruction of the original image data given the target quantized feature representation, and the probability distribution after entropy encoding, calculate the optimization objective function, specifically implemented through the following formula:

[0065]

[0066] Where x represents the original image data, Let q represent the target quantization feature representation, and p represent the vehicle-to-everything (V2X) feature extraction model. x Represents the distribution of the original image data. This represents the true probability distribution after quantization. To measure q and The KL divergence of the differences between them This represents the quantized probability distribution. This represents the probability distribution after entropy encoding. This represents the probability that the original image data can be successfully reconstructed given the target quantized feature representation, and const represents the pre-set residual.

[0067] The core of the objective function is to minimize a KL divergence, which measures... and true posterior The distance between them is made to approximate the true posterior distribution as closely as possible, because the true posterior distribution... It is difficult to compute, as it involves integration or summation over the entire data space. Therefore, we introduce an alternative, a computable approximate distribution.

[0068] The optimization objective consists of two parts: a weighted distortion term, which measures the loss between the restored image and the original image x, and is best when as close to 0 as possible; and a bit rate term, which measures the encoding... The required number of bits describes the length of the compressed data; the smaller the better.

[0069] During training, the posterior distribution is relevant to the actual task and describes the distribution of the latent variables. An approximate distribution is designed to simplify computation. To approximate the true posterior distribution and thus optimize the model's computational efficiency, an optimizable objective function is constructed by balancing encoding error and bit rate, and posterior distribution estimation is used to improve the model's training performance.

[0070] The optimization objective function is one of the core inventive points of this invention because it guides the joint training of various modules, including the encoder, quantization, entropy coding, and decoder, from end to end. This objective function achieves overall optimization of video data compression performance in vehicle-to-everything (V2X) scenarios by balancing coding error (distortion term) and coding cost (bit rate term). In the image coding steps below, the original image x is processed by the encoder to generate a feature representation y; in the quantization inference and entropy coding steps, y is quantized and probabilistically modeled to obtain a symbol sequence; throughout the process, the optimization objective function is used as the loss function to guide the joint optimization of parameters in each step.

[0071] Step 7: Update the parameters of the image encoder and the image decoder according to the optimization objective function;

[0072] This involves updating the parameters of the image encoder, which means updating the parameters of the general feature extraction model and the vehicle-to-everything (V2X) feature extraction model.

[0073] Step 8: Acquire multiple raw image data, repeat steps 1 to 7 until the number of repetitions of steps 1 to 7 reaches the preset number of rounds, and obtain the trained image encoder and image decoder. Use the trained image encoder to compress the image through the processes of image encoding, quantization and entropy encoding.

[0074] This invention can also be understood as follows: the present invention obtains multiple original image data and the corresponding quantized true probability distribution of the original image data. Based on this, the original image data is used as input samples, and steps 1 to 7 are executed to update the parameters of the image encoder and the image decoder. Then, a new original image data is used as input samples to execute steps 1 to 7 again to update the parameters of the image encoder and the image decoder. This process is repeated until the number of times steps 1 to 7 are executed reaches a preset number of rounds, that is, the number of times the parameters of the image encoder and the image decoder are updated reaches a preset number of rounds, thus obtaining the updated image encoder and image decoder.

[0075] Therefore, image encoding is performed on the image data to be compressed to obtain the feature representation of the image data to be compressed. A uniformly distributed noise is added to the feature representation of the image data to be compressed. The feature representation of the image data to be compressed after adding noise is quantized and entropy encoded to obtain the encoded feature representation of the image to be compressed, thus realizing the image compression process.

[0076] This invention ingeniously integrates existing learning-based image / video compression algorithms and the LoRA model fine-tuning technique. Addressing the shortcomings of existing methods in video compression within the connected vehicle (V2V) field, it proposes a comprehensive solution aimed at improving video compression efficiency and quality in the V2V domain. The following are the core highlights of this invention:

[0077] 1. Hybrid Architecture Training Paradigm: The entire training process employs a Transformer-CNN hybrid architecture, combining the local feature extraction capabilities of CNNs with the global modeling capabilities of Transformers to learn the optimal latent representation. Compared to pure CNN or pure Transformer structures, the hybrid architecture can better balance computational efficiency and compression quality, adapting to the diverse video data in connected vehicle scenarios.

[0078] 2. Enhanced Adaptability with a Dual-Branch Structure: A dual-branch structure is introduced to learn general image features and vehicle-to-everything (V2X) specific image features separately. This allows the model to not only retain general image information but also optimize for the data distribution in V2X scenarios, thereby improving compression performance.

[0079] 3. Efficient Fine-Tuning Mechanism: LoRA is used for lightweight fine-tuning to reduce computational costs and optimization difficulty, enabling the model to efficiently adapt to the data characteristics of the connected vehicle domain. LoRA can significantly improve performance on specific connected vehicle datasets by updating only a small number of parameters without affecting the model's original generalization ability.

[0080] This invention achieves highly efficient compression of video data in the field of vehicle networking by comprehensively utilizing existing learning-based image / video compression algorithms and LoRA fine-tuning technology. Compared with existing traditional video compression algorithms and neural network image compression strategies, it exhibits significant advantages and superior performance. The main effects and advantages of this invention are as follows:

[0081] 1. Improved video data compression quality: This invention designs a dual-branch structure, which solves the problem of mismatch between the model trained by existing learning-based image compression algorithms and the characteristics of video data in the field of vehicle networking, and significantly improves the quality of decompressed video data.

[0082] 2. Reduced model training time: By utilizing LoRA fine-tuning technology, this invention can quickly train a model suitable for the characteristics of vehicle network video data, greatly reducing the number of parameters to be learned and improving model training efficiency, especially when processing large amounts of vehicle network video data, the efficiency improvement is particularly significant.

[0083] 3. Reduce the average bit rate of pixels: The hybrid architecture designed in this invention can capture the local texture features and global context dependencies of images, which significantly reduces the average bit rate of pixels and improves the storage efficiency of vehicle network video data.

[0084] 4. Adaptable to complex and ever-changing video data: This invention is not affected by the complex and ever-changing video content of the Internet of Vehicles, and can maintain high stability whether in rural areas, cities or other environments.

[0085] In our experiments, we compared the compression performance of our invention with existing learning-based image / video compression algorithms on multiple connected vehicle datasets. The results show that, under the same hardware conditions, our invention improves the PSNR (Peak Signal-to-Noise Ratio) for evaluating decompressed image quality and the SSIM (Structural Similarity Index) for measuring the similarity between the decoded and original images by approximately 15%, while reducing the average bit-rate per pixel by 17%, and maintaining high reliability and stability. This invention demonstrates significant technical advantages and application value in video compression within the connected vehicle field.

[0086] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for compressing video data in a vehicle-to-everything (V2X) scenario, characterized in that, include: Step 1: The original image data is encoded using an image encoder to obtain the feature representation of the original image data; Specifically, the original image data is encoded through multiple basic blocks in the image encoder. In the first basic block, the original image data is used as input. The original image data is processed by a general feature extraction model to obtain general image feature information. At the same time, the original image data is processed by a vehicle-to-everything (V2X) feature extraction model to obtain V2X image feature information. The general image feature information and the V2X image feature information are fused to obtain the combined feature representation of the first basic block. The combined feature representation of the first basic block is used as input to the next basic module. Similarly, the processing and feature fusion operations through the general feature extraction model and the V2X feature extraction model are repeated until the last basic block. The combined feature output of the last basic block is the feature representation of the original image data. The vehicle network feature extraction model is obtained by fine-tuning the general feature extraction model based on existing public vehicle network datasets using lightweight fine-tuning techniques. Step 2: Add a uniformly distributed noise to the feature representation of the original image data to obtain the feature representation after noise addition. By introducing the cumulative distribution function CDF, the feature representation after noise addition is quantized to obtain the target quantized feature representation. Then, based on the target quantized feature representation, the probability distribution after quantization is calculated. Step 3: Perform entropy encoding on each quantized feature representation in the target quantized feature representation to obtain the encoded feature representation, and then calculate the probability distribution after entropy encoding; Step 4: Perform entropy decoding on the encoded feature representation to obtain the recovered quantized feature representation, and then perform dequantization on the recovered quantized feature representation to obtain the recovered continuous feature representation; Step 5: Decode the recovered continuous feature representation through CNN deconvolution and attention mechanism in the image decoder to obtain reconstructed image data. Then, based on the reconstructed image data, calculate the probability that the original image data is successfully reconstructed given the target quantized feature representation. Step 6: Calculate the optimization objective function based on the quantized probability distribution, the probability of successful reconstruction of the original image data given the target quantized feature representation, and the probability distribution after entropy encoding; Step 7: Update the parameters of the image encoder and the image decoder according to the optimization objective function; Step 8: Acquire multiple raw image data, repeat steps 1 to 7 until the number of repetitions of steps 1 to 7 reaches the preset number of rounds, and obtain the trained image encoder and image decoder. Use the trained image encoder to compress the image through the processes of image encoding, quantization and entropy encoding.

2. The method for compressing video data in a vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, The general feature extraction model is obtained through the following methods: Image data from the general dataset ImageNet is input into the hybrid Transformer-CNN architecture model to obtain general feature information of the predicted images. Based on the general feature information of the predicted images and the general feature information of the real images, the loss value is calculated. The model parameters of the hybrid Transformer-CNN architecture model are updated based on the loss value until the loss value is less than a preset threshold, thus obtaining the general feature extraction model.

3. The method for compressing video data in a vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, In step 2, the probability distribution after quantization is calculated based on the target quantized feature representation, specifically through the following formula: ； in, x Represents the original image data. This represents the quantitative feature representation of the target. This represents the quantized probability distribution. The parameters represent the image encoder, and the uniform distribution is indicated. This represents the i-th feature representation after adding noise. express Quantized feature representation.

4. The method for compressing video data in a vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, Step 3 calculates the probability distribution after entropy encoding, specifically using the following formula: ； in, This represents the quantitative feature representation of the target. This represents the probability distribution after entropy encoding. The parameters represent the entropy encoder. for The i One portion, Indicates under given conditions The probability distribution, The i-th feature in the feature representation after adding noise is the quantized feature representation.

5. The method for compressing video data in a vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, In step 5, based on the reconstructed image data, the probability that the original image data can be successfully reconstructed given the target quantized feature representation is calculated, specifically through the following formula: ； in, x Represents the original image data. This represents the quantitative feature representation of the target. This represents the probability that the original image data can be successfully reconstructed given the target quantized feature representation. This represents the parameters of the image decoder. This represents reconstructed image data. This is a hyperparameter used to measure the importance of the mean squared error between the original image data and the reconstructed image data. This represents the mean square error between the original image data and the reconstructed image data.

6. The method for compressing video data in a vehicle-to-everything (V2X) scenario according to claim 1, characterized in that, Step 6 is specifically implemented using the following formula: ； in, x Represents the original image data. This represents the quantitative feature representation of the target. This represents the feature extraction model for vehicle-to-everything (V2X) networks. Represents the distribution of the original image data. This represents the true probability distribution after quantization. To measure and The KL divergence of the differences between them This represents the quantized probability distribution. This represents the probability distribution after entropy encoding. This represents the probability that the original image data can be successfully reconstructed given the target quantized feature representation. This indicates the pre-set residual.

Images