Nighttime Image Flare Removal Based on Wavelet Transform and Frequency Domain Interaction

Images captured at night often suffer from flare artifacts, which are characterized by their diverse shapes and strong contamination. Removing these flares and restoring the details in the contaminated areas are urgent issues to address. The low-frequency components of an image contain information about its color and shape, which can be helpful in modeling the structural features of flares to cope with their various shapes. Meanwhile, the high-frequency components contain details and textures, which are beneficial for repairing the details in areas affected by flare contamination. In this paper, we propose a method for removing flares from nighttime images based on wavelet transform and frequency domain interaction, which consists of two stages: frequency domain feature extraction and frequency domain interaction fusion. In the frequency domain feature extraction stage, the flare features are separated into high and low-frequency features through the wavelet transform, and both high and low-frequency features are adaptively modeled by the Multi-scale Feature Extraction Block (MFEB) which includes the Adaptive Feature Fusion Block (AFFB). In the frequency domain interaction fusion stage, the high and low-frequency features undergo interaction learning through the Frequency Domain Cooperative Interaction Block (FCIB) in the frequency domain, and then the interacted high and low-frequency features are respectively sent to the MFEB for deep fusion. Finally, the de-flared image is obtained through the inverse wavelet transform. Experimental results on a benchmark dataset for nighttime image flare removal demonstrate that the proposed method achieves state-of-the-art performance in nighttime image flare removal.