Improving Semantic Segmentation Under Hazy Weather for Autonomous Vehicles Using Explainable Artificial Intelligence and Adaptive Dehazing Approach

Abstract

Haze-level discriminators are crucial for autonomous vehicles to handle segmentation tasks successfully in hazy and foggy outdoor environments. Deep learning (DL) networks trained to segment clear images exhibit more false positives and fail to recognize the pixel patterns for the class categories when faced with hazy images. To address this issue, we propose a novel dehazing scheme called Adaptive Dehazing (AD) that separates the unacceptable hazy images and applies the dehazing technique only to those images before passing them to the DL for segmentation. We define various thresholds to classify hazy images into four categories: heavy, moderate, slight, and clear. Additionally, we implement a hazy image generator to create hazy synthetic images that replicate actual hazy road scene conditions for testing the proposed algorithm. Moreover, we use the Explainable Artificial Intelligence (XAI) method to understand the feature selected by different layers in the network before and after applying the AD and their contribution to obtaining a final segmented output. Extensive experiments demonstrate both quantitatively and qualitatively the performance improvement in the segmentation task by achieving an optimal balance between Intersection over Union (IoU) and pixel accuracy (PA) metrics of the segmented categories. The improvement in IoU metrics shows that the AD scheme significantly outperforms the previous state-of-the-art dehazing methods.

Keywords

• Autonomous vehicle
• dehazing
• deep learning
• semantic segmentation
• XAI