Deep Multi-Patch Hierarchical Network-Based Visibility Restoration Model for Autonomous Vehicles

Dilbag Singh, Ahmad Ali AlZubi, Manjit Kaur, Vijay Kumar, Heung-No Lee

Abstract: Clear visibility of the road is crucial for autonomous vehicles, but adverse weather conditions can significantly reduce road visibility, impacting their performance. To address this challenge, a visibility restoration model is essential for autonomous driving and navigation. This paper proposes a Lightweight Visibility Restoration Network (LVR-Net) designed to restore images with improved spatial and spectral information despite severe weather degradation. Initially, a Deep Multi-Patch Hierarchical Network (DMPHN) is employed, but its computational demands hinder deployment on autonomous vehicles. To overcome this, a Modified Adaptive distributed Differential Evolution (MADE) optimization technique is applied to enhance the network's size, computational speed, and overall performance. A multi-objective fitness function based on Peak Signal-to-Noise Ratio (PSNR), memory size, and computational speed is formulated. Benchmark datasets are utilized to train, validate, and test the LVR-Net. Experimental results demonstrate that the proposed LVR-Net outperforms other competitive models across various performance metrics. Additionally, its compact size (48.25 MB) and faster scene restoration make it highly suitable for deployment in autonomous vehicles.

External IDs:dblp:journals/tvt/SinghAKKL25