Knowledge distillation for super-resolution reconstruction and segmentation in forward-facing camera images

Abstract

Semantic segmentation in vehicular vision faces a critical multi-objective challenge: maintaining high accuracy, especially for distant and small objects, while meeting the stringent low computational cost requirements of on-vehicle systems. We focused on improving perception accuracy in simulated adverse conditions where low-resolution (LR) images lead to poor recognition of small or far-distance objects. To address this, we propose knowledge distillation for super-resolution reconstruction and semantic segmentation (KD4SRSS), a novel end-to-end framework combining super-resolution reconstruction (SR) and semantic segmentation. KD4SRSS utilizes the proposed lightweight spatial boundary-aware network (SBANet), which requires only 147,312 parameters, and introduces the boundary-aware knowledge distillation (BAKD) method. BAKD efficiently transfers semantic and crucial boundary knowledge from a robust Teacher network to the Student SBANet, enabling boundary-centric SR at minimal computational expense. Experiments on the Cambridge-driving labeled video database (CamVid) and mini-database of the Cityscapes (MiniCity) datasets confirm KD4SRSS's superior performance: it achieved mean intersection over union (mIoU) scores of 64.42 % and 35.79 % respectively, representing a significant improvement of 3.02 % (CamVid) and 4.74 % (MiniCity) over the state-of-the-art (SOTA) baseline. This performance validates KD4SRSS as an optimal, robust solution for real-time applications in resource-constrained intelligent vehicle systems. © 2026 Elsevier B.V.