H2CGAN: Manageable AI for Scene Understanding Tasks in Hazy/Rainy Environment

Abstract

This paper primarily focuses on image understanding tasks in hazy/rainy environments. We carried out research for image enhancement in hazy/rainy situations in tandem with object detection in those scenes. We first introduce an unsupervised Generative Adversarial Network (GAN) based network for hazy/rainy image translation tasks, which is trained on unpaired real-time hazy/rainy datasets. The network is trained with latent feature loss, geometric consistency, and contextual loss to deal with the instability of cycle consistency in unsupervised GANs. We compare the results obtained on Cityscapes, Reside, Dawn-Haze, Dawn-Rain, and RIS datasets with state-of-the-art (SOTA) algorithms in terms of visual quality as well as no-reference quality indices. Using our method, we achieved an overall gain of 35.5% and 36.2% in the PIQE (P) image quality score for the hazy and rainy images, respectively. We further report object detection scores on the processed images. Our proposed model outperforms other SOTA algorithms for the hazy domain, achieving an overall gain of 1.75% in mAP scores across all datasets. For the rainy domain, our model gave the third-best result with mean gain of 2.4% in mAP. This study aims to demonstrate a particular network’s ability to improve detection accuracy based on its feature activations. Our experiments demonstrate which enhancing technique generates higher feature activations from the detection network, hence resulting in higher detection scores. We provide a precise understanding of how learning works in the network for different prior knowledge situations.

Keywords

• Explainability
• H2CGAN
• image understanding
• object detection
• unconstrained environment