Research on Flare Removal Network Based on Channel Attention Mechanism and Depthwise Over-parameterized Convolution

Guanyu Zou, Hongyang Bai, Yuman Yuan, Tianyu Deng, Zhenpeng Yin, Jingyi Wei

Abstract: Lens flare significantly affects the visual quality of images, and it can prevent a computer vision program from performing well in certain tasks, such as optical flow estimation and stereo matching. Removing this issue can help ensure that the program performs optimally in future. Existing deep learning-based methods for lens flare removal suffer from the problem of partial loss of image details and structural information due to the lack of information exchange between channels and weak utilization of local contextual details in the images. To address this issue, we propose a network for nighttime lens flare removal by combining channel attention mechanism and depthwise over-parameterized convolution. Firstly, to enhance the perception and extraction of different types and directions of lens flare, we incorporate the channel attention mechanism into the attention component of the Transformer module, which strengthens the interaction between channels and enhances the feature maps of the input through channel feature enhancement. Secondly, to improve the resolution ability of lens flare details and the effectiveness of lens flare removal, we introduce a depth over-parameterized feedforward network by adding over-parameterized depth convolution blocks in the feedforward network, which enhances the utilization of local contextual information in the images. Experimental results on the Flare7k++ dataset demonstrate that our network achieves a 0.049 dB improvement in PSNR and a slight reduction of 0.0008 in LPIPS compared to the optimal method Uformer, a 0.002 improvement in SSIM compared to the optimal method Restormer.