Adversarial Domain Adaptation and Multi-View Semantic Fusion for CNN-Based Autonomous Driving Image Recognition

The development of autonomous driving technology presents challenges for vision systems, especially in complex environments such as changing weather, lighting, and tunnels. Traditional convolutional neural networks (CNNs) struggle with domain shift issues, and surround systems face limitations in view completion and adaptive weighting. To address these, we propose an end-to-end fusion framework. This framework integrates a Gradient Reversal Layer (GRL) into the Faster R-CNN backbone to align features across different domains. Additionally, a spatial-channel dual attention module is employed to focus on key regions. A multi-view semantic fusion network based on multi-head self-attention and high-order spectral graph convolution is designed to dynamically assign camera weights and capture high-order topological dependencies between views, effectively completing blind spots. Optimization strategies such as feature downsampling and sparse adjacency are applied to maintain real-time perception while reducing computational costs. Experimental results on a public multi-environment, multiview driving dataset show significant improvements over baseline methods in object detection and short-term trajectory prediction tasks. Ablation studies further confirm the synergistic benefits of adversarial alignment, attention focusing, and multi-view fusion. The proposed method demonstrates excellent stability and accuracy in complex driving scenarios, offering strong support for multi-sensor collaboration, self-supervised learning, and efficient hardware deployment.