TheiaNet: Towards fast and inexpensive CNN design choices for image dehazing

Abstract

Abstract This work examines inexpensive design choices for dehazing as an end-to-end image-to-image mapping problem without relying on the physical scattering model. The proposed TheiaNet is free from intermediate-computation of transmission map, enabling haze removal in a highly resource constrained environments. The simplicity of the network is augmented by a spatial cleaning bottleneck block, that adds faster feature extraction without adding to trainable parameters. We also analyze the effectiveness of multi-cue color space (RGB, HSV, LAB, YCbCr) over single cue color space (RGB) for end-to-end dehazing. A comprehensive set of experiments were conducted on HazeRD, D-Hazy and the more recent Reside datasets. The proposed TheiaNet significantly outperforms the existing CNN and GAN based state-of-the-art methods in terms of PSNR and SSIM on all these datasets. It also outperforms all existing methods in term of speed, compute and memory efficiency, making it more efficient. This work highlights how judicious application-specific components can augment simple CNNs to denoise faster, and more accurately than deeper heavier networks, which is supported by an ablation analysis as well.