Enhanced U-Net for Car Segmentation with Residual Attention and Adaptive Feature Fusion

Accurate vehicle segmentation is critical for autonomous driving and intelligent transportation systems. However, standard U-Net architectures often struggle with boundary ambiguity and background noise due to rigid feature fusion and information loss during deep encoding. To address these challenges, we propose an enhanced U-Net that establishes a logical synergy between feature purification and dynamic reconstruction. Specifically, we introduce a Residual Channel Attention Double Convolution (RCADC) module in the encoder to suppress background interference and preserve sharp edge textures through residual learning and channel-wise recalibration. Complementing this, an Adaptive Weight Feature Fusion (AWFF) unit is designed for the decoding path. Unlike static concatenation, AWFF employs learnable weights to dynamically bridge the semantic gap between purified encoder features and upsampled maps, ensuring precise boundary restoration. Experimental results on the Carvana dataset demonstrate that this closed-loop design achieves a Dice coefficient of 0.96 and a significant improvement in Edge Dice from 0.65 to 0.91, while maintaining a real-time inference speed of 48 FPS. Qualitative and quantitative evaluations further confirm the model’s robustness under challenging illumination and occlusion, offering a technically sound and computationally efficient solution for high-precision vehicle segmentation