Fine lightning segmentation algorithm based on frequency domain denoising

Lightning was challenging to capture using traditional image processing methods due to interference from city lights, clouds, and trees. However, deep learning had the potential to surpass traditional methods by enabling fine branch detection that would otherwise require the naked eye, offering significant convenience to lightning researchers. In this paper, we addressed several challenges in lightning image segmentation, including low accuracy, noise interference, limited segmentation capabilities under very dark conditions, and rough masks. We presented a refined lightning mask dataset and proposed a refined segmentation model (FD-FLSNet), which utilized frequency domain denoising. This model improved fine branch detection by fusing multi-scale features through a W-structured U2-Net + architecture. Additionally, wavelet downsampling was employed to mitigate the loss of small structural features of lightning, and the lightning shape-calibrated attention mechanism (LRCA) was integrated to better capture the spatial features of various lightning morphologies. Moreover, the segmentation of weak branches and channel lightning was improved, making it more accurate and realistic. To further enhance noise suppression, FD-LFEM was used to transform the features into the frequency domain, effectively reducing noise interference.