DFDNet: Dynamic Frequency-Guided De-Flare Network

Strong light sources in nighttime photography often cause flares that degrade image quality and compromise the performance of downstream tasks such as object detection. However, nighttime flares are usually accompanied by complex degradations, such as large-scale diffusion, directional streaks, and high-intensity saturation around light sources. Existing methods still struggle to simultaneously remove large-scale flare artifacts and recover local structures near light sources. To address these issues, we propose a dynamic frequency-guided deflare network, termed DFDNet. Specifically, we design a global dynamic frequency-domain guidance module (GDFG), which maps features into the frequency domain through the discrete Fourier transform and employs multichannel dynamic frequency weights conditioned on the input to modulate flare-related abnormal frequency responses. In this way, flare information can be separated from background content, and large-scale flare artifacts can be effectively suppressed. Moreover, we design a local detail guidance module (LDGM), which aligns local features around light sources between the restored image and the reference image through contrastive learning, thereby reducing detail damage caused by flare removal and improving fine-grained image restoration. The experimental results demonstrate that the performance of the proposed method is competitive with that of existing state-of-the-art methods. The code is available at https://github.com/qy792832/DFDNet.