A self-prompt based dual-domain network for nighttime flare removal

Existing nighttime flare removal work regards flare as a single degradation factor in spatial domain. However, flare in complex scenes consists of multiple flare types, and it is difficult to distinguish them from the background, leading to distorted results and incomplete perception. In this paper, we propose a self-prompt based dual-domain network named SPDDNet for nighttime flare removal, which encodes the data distributions of different flare types to generate prompt features and facilitates the interaction of the prompt features with the decoder to guide network for flare removal. In addition, we introduce Fast Fourier Transform and parallel attention in traditional convolutional neural network that is designed to extract global frequency features and location-dependent local information to accurately perceive the flare region. Finally, to adequately integrate spatial details, contextual information, prompt and image features, we propose a feature fusion module that generates a set of learned dynamic weights to adaptively guide the information fusion across channels. Extensive experiments on real-world and synthetic datasets strongly demonstrate the effectiveness of our proposed SPDDNet and its superior performance compared to state-of-the-art methods. Moreover, as an essential pre-processing step, the potential advantages of our method for other computer vision applications, including object detection and semantic segmentation, are demonstrated.