Scholarlink

<p>Vehicular communication, as the core connection hub between the perception layer, decision layer, and execution layer of intelligent driving, faces an explosive demand for real-time image transmission and processing. The accuracy and real-time performance of image reconstruction directly determine the reliability of safety functions such as lane detection, pedestrian detection, and collision warning. However, conventional compressive sensing image reconstruction algorithms suffer from detail loss, structural distortion, and high computational complexity in vehicular scenarios, making them difficult to adapt to limited bandwidth and computing resources. To address this, a compressive sensing image reconstruction algorithm for vehicular communication is proposed. The algorithm first integrates Multi-Scale Residual Convolution, Coordinate Spatial Attention mechanism, and Depth wise Separable Convolution to construct an image feature extraction algorithm. It accurately captures key details in vehicular images while reducing computational cost. Then, it combines a Generative Adversarial Network optimized by a variational autoencoder with a Vision Transformer to model the correlation between low-sampling features and complete images, ensuring global coherence of the reconstruction results. Experimental results show that in dense pedestrian areas, the structural similarity index reaches 0.935, the peak signal-to-noise ratio is 33.64&#xa0;dB, the maximum memory usage is 413.6&#xa0;MB, and the response time is 162.4&#xa0;ms with 2000 data samples. The proposed algorithm outperforms comparative methods in multiple metrics, balancing reconstruction accuracy, anti-interference capability, and real-time performance, and it adapts well to dynamic vehicular communication environments, providing efficient and reliable image support for intelligent driving visual systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-scale residual convolutional block-based compressive sensing image reconstruction for vehicular communication

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月31日

查看全文
  • Jingtao Guan

摘要

Vehicular communication, as the core connection hub between the perception layer, decision layer, and execution layer of intelligent driving, faces an explosive demand for real-time image transmission and processing. The accuracy and real-time performance of image reconstruction directly determine the reliability of safety functions such as lane detection, pedestrian detection, and collision warning. However, conventional compressive sensing image reconstruction algorithms suffer from detail loss, structural distortion, and high computational complexity in vehicular scenarios, making them difficult to adapt to limited bandwidth and computing resources. To address this, a compressive sensing image reconstruction algorithm for vehicular communication is proposed. The algorithm first integrates Multi-Scale Residual Convolution, Coordinate Spatial Attention mechanism, and Depth wise Separable Convolution to construct an image feature extraction algorithm. It accurately captures key details in vehicular images while reducing computational cost. Then, it combines a Generative Adversarial Network optimized by a variational autoencoder with a Vision Transformer to model the correlation between low-sampling features and complete images, ensuring global coherence of the reconstruction results. Experimental results show that in dense pedestrian areas, the structural similarity index reaches 0.935, the peak signal-to-noise ratio is 33.64 dB, the maximum memory usage is 413.6 MB, and the response time is 162.4 ms with 2000 data samples. The proposed algorithm outperforms comparative methods in multiple metrics, balancing reconstruction accuracy, anti-interference capability, and real-time performance, and it adapts well to dynamic vehicular communication environments, providing efficient and reliable image support for intelligent driving visual systems.