Learning-Based Flare Removal for Images of Korean Wooden Joint Bracket Sets

Flare removal is an important process that aims to improve the clarity and visual quality of images by reducing artifacts such as streaks, circles, and haze caused by lens defects and light reflections. However, there are significant reflective flares in images of Korean wooden joint bracket sets that the camera directed towards the light source. We used the acquired images and applied them by adopting a learning-based method that reduces the reflective flare phenomenon. This approach involves creating pairs of flare-free original images and new images synthesized with arbitrary flares onto the original images. The pairs are then used to train a deep learning network to restore flared photos to their original state. In our experiment, we augmented the flare by adding images of Korean wooden joint bracket sets to the previous Flare dataset. Furthermore, we trained the flare removal model with pairs of original images and flare-synthesized images. As a result, we generated images with removed flare through this process.