Vision Unveiling Through End-to-End Barur-Net Approach
C. Jenisha, C. Sheeba Joice · IETE Journal of Research · 2025
Dehazing images with cost-effective approaches remains a significant and ever-evolving research area in computer vision. Atmospheric elements like mist, fog, and haze can significantly degrade image quality. This paper proposes Barur-Net an end-to-end and computationally efficient approach for image dehazing. Unlike conventional methods, dependent on the Atmospheric Scattering Model (ASM), Barur-Net directly generates a haze-free image in a single step. It employs an input feature concatenation concept to enhance feature learning. A simple Residual Network (ResNet) is utilized to eliminate the problem of vanishing gradient during training. Furthermore, an encoder and decoder are adapted to maintain the feature spatial dimension, with the BOttleneck CAscade enhancer moduLe (BOCAL), employed between them to enhance the overall model performance. Barur-Net is a lightweight yet powerful dehazing framework, that effectively balances computational efficiency with performance. Further, it requires only 1.66 MB of parameters and 6.49 G FLOPs, making it significantly more efficient than existing dehazing methods. Extensive qualitative and quantitative evaluations on diverse benchmark datasets, and real-world captured hazy images in Kodaikanal, demonstrate that Barur-Net exceeds state-of-the-art dehazing methods. It achieves an average PSNR of 23.4683, SSIM of 0.948, MSE of 0.0048, and RMSE of 0.0684, establishing its superiority in image dehazing.