Hard-negative Sampling with Cascaded Fine-Tuning Network to Boost Flare Removal Performance in the Nighttime Images

Abstract

When light passes through a camera lens, it creates a residue called "flare" due to the interaction between foreign substances on the lens surface and internal glasses. At night, images can be distorted by flare due to multiple light sources, and research has been conducted using neural networks to remove the flare and solve this problem. However, to our knowledge, research on this approach has only recently begun, and the results are still limited, with only a few models available for use. Further research is needed to determine if the existing models provide optimal results. As part of the mentioned research, we propose a cascaded neural network structure as a means of fine-tuning earlier models to improve their performance. We optimize the performance of the proposed model by constructing triplets using the outputs of two identical neural networks and applying contrastive learning. To demonstrate the superiority of the proposed method, we quantitatively evaluated it by measuring PSNR and SSIM. We also visually compared the differences in image details after removing the flare. Experimental results confirmed that the images reconstructed by the proposed model were superior in terms of PSNR and SSIM in streak regions, compared to the results generated by the reference model.