SEGMENTATION OF LENS FLARES IN DIGITAL IMAGES

Abstract

This paper investigates the problem of lens flare detection in digital images. As a result of the analysis of modern approaches to solving this problem, it was found that existing methods, although demonstrating good results for the main cameras of modern smartphones, still have difficulties with processing images captured with telephoto lenses, since due to their structure they are more prone to reflection defects. Therefore, we focus our study on such images. A small test dataset was created for the study. It is demonstrated that the assumption of symmetric reflection to the light source relative to the image center is not always true in the case of lenses with a more complex lens system, which is the case in our dataset. We propose an algorithm that detects reflective flare based on color, area, and location characteristics. The algorithm is not as general as other state-of-the-art methods based on deep learning, as it is based on several assumptions about the nature of a flare, which are primarily typical for telephoto camera images with several bright light sources and a dark background. However, a comparison of the proposed method with several popular approaches demonstrates a significant improvement in segmentation quality for the considered test set. With further improvement, the algorithm can be useful both for visual enhancement of photos and as a data preprocessing tool in many computer vision tasks.