Structural-Prior Deep Learning Network for Millimeter-Wave Radar Image Enhancement in Autonomous Driving Road Sensing

Hongyan Chen, Tonghui Huang, Yuexia Wang, Jiajia Shi, Zhihuo Xu

Millimeter-wave radar imaging plays an increasingly important role in autonomous driving road perception due to its robustness under adverse weather conditions. However, radar images are inherently contaminated by multiplicative speckle noise, which severely degrades structural continuity, weakens target boundaries, and limits the perception of road scenes and surrounding objects. To address this problem, this paper proposes a structural-prior deep learning network for millimeter-wave radar image enhancement. The proposed framework first introduces an adaptive Otsu-based masking strategy to extract salient scattering structures and generate a coarse image structural prior for subsequent restoration. Guided by this prior, the network performs progressive feature enhancement through a continuous attention mechanism that integrates residual channel attention, context-aware feature extraction, and convolutional block attention, thereby enabling effective multi-scale representation learning while suppressing signal-dependent speckle interference. In addition, a composite loss function is designed by combining logarithmic denoising gain, total variation regularization, and a β-index edge-preservation term to jointly improve noise suppression, spatial smoothness, and structural fidelity. The proposed method is evaluated on the synthetic UC Merced dataset under different noise intensities and via cross-domain inference on the real-world RADIATE millimeter-wave radar dataset for autonomous driving scenarios. Experimental results demonstrate that the proposed network consistently outperforms conventional filtering methods and representative deep learning baselines in terms of PSNR, SSIM, β-index, and ENL while providing a superior preservation of road structures, target contours, and scene geometry. Ablation studies further confirm the effectiveness of the structural-prior guidance and continuous attention design. Furthermore, the network achieves a rapid inference latency of 12.35 milliseconds. These results indicate that the proposed method provides an effective and robust solution for millimeter-wave radar image enhancement and offers practical value for downstream road-scene perception in autonomous driving environments.