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

Abstract

Haze-level discriminators are highly desirable for autonomous vehicles to handle segmentation tasks successfully in hazy and foggy outdoor environments. Neural networks trained to detect clear images show more false positives and unrecognize the pixel patterns for the class categories when they encounter hazy images. Here, we propose a novel dehazing scheme called Adaptive Dehazing (AD) which can separate the unacceptable hazy images and apply the dehazing technique only to those images before passing it to the neural network for segmentation. Various thresholds have been defined to classify hazy images into four hazy categories: heavy, moderate, slight and clear. Also, a hazy image generator is implemented to create synthetic hazy images that replicate the actual hazy road scene conditions for testing the proposed algorithm. In addition, 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 improving the Intersection over Union (IoU) and pixel accuracy (PA) of the segmented categories. Extensive experiments demonstrate quantitatively as well as qualitatively the performance improvement in the segmentation task. Our AD scheme significantly outperforms the previous state-of-the-art dehazing methods.