DSSCNet: Deep Custom Spatial and Spectral Consistency Layer-Based Dehazing Network

Abstract

Poor weather conditions, such as haze, fog, and smog, present significant challenges in capturing clear and visually appealing images. Though the existing image dehazing algorithms have achieved significant performance, they still suffer from various problems such as generalization to diverse hazy conditions, potential artifact generation, and computational complexity. Additionally, sensitivity to parameter settings, haze density variability, image content, noise, and scene-specific information remains areas of concern. To address these issues, we propose a Deep Custom Spatial and Spectral Consistency Layer-based Dehazing Network (DSSCNet) that effectively removes haze from images while preserving important spatial and spectral details. The network architecture includes a custom Haze Removal Layer (HRL), convolutional layers with ReLU activation, pooling layers, skip connections, and a custom Spatial and Spectral Consistency Layer (cSSCL). HRL estimates atmospheric light and transmission maps to generate an intermediate haze-free image. The proposed loss function combines Mean Squared Error (MSE) loss with a Consistency Loss (CL) to encourage content preservation during dehazing. Extensive experimental results demonstrate that DSSCNet outperforms competitive models in terms of various performance metrics, including contrast gain ( $c_{g}$ ), new visible edges ( $e$ ), new edge gradients ( $\bar {r}$ ), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) by average improvements of approximately 1.27%, 1.12%, 1.18%, 1.21%, and 1.24%, respectively.

Keywords

• Dehazing
• deep learning
• convolutional neural network
• atmospheric light
• transmission map
• spatial and spectral consistency