Comparative Study of Deep Learning Approaches for Classification of Flares in Images

Flare is common in most photographs captured against bright light sources such as glowing LED, Sun, and bulb. The lens artifacts are diverse, depending on light reflection inside the lens or clouds of dust reflecting light particles. The resulting flare may contain the color spread, bright spot, halos, haze, and streaks. However, these artifacts cause a severe problem for detecting objects in real time, generating false results. Drones and mobile robots encounter these flares frequently, failing to furnish features of the scene in numerous cases. In addition, it is common to use simulated flare in the movie generation field to improve the viewer’s inclination in terms of aesthetic view. We are not carrying on added flashes; instead, we focus on natural flare exposure in the robotic sector. This work focuses on distinguishing whether the image is a flare or not, applying the current state-of-the-art deep learning approaches. We experiment on the synthetic and natural dataset with diverse lens flare for flare classification. We obtained the accuracy of ResNet18 (acc: 96.8), AlexNet (acc: 91.6), MobileNetv2 (acc: 97.4), SqueezeNet (acc: 93.3), VGGNet (acc: 96.2). Combining these approaches acts as a preprocessing step that promotes outdoor robots or drones to eliminate the flare frames if present, leading to better accuracy in the next steps such as object detection or tracking of an object.