Learning to Detect Multiple Photographic Defects

Although problems relating to specific image correction have been explored intensively, the problem of simultaneous diagnosis for multiple photographic defects remains relatively untouched. Solutions to this problem attempt to predict the existence, severity, and locations of common defects. This paper proposes a first attempt at a solution to the general defect diagnosis problem based on our novel dataset. We formulate the defect diagnosis problem as a multi-task prediction problem and utilize multi-column deep neural networks (DNN) to approach the problem. We propose DNN models with holistic and multi-patch inputs and combine their predicted scores to integrate multi-scale information. During experiments, we validate the complementarity of both kinds of inputs. We also validate that our combined predictions have a more consistent ranking correlation with our ground truth than the average of individual users’ judgments. Furthermore, we apply the fully convolutional version of our trained model to visualize defect severity heat maps, which can effectively identify defective regions of input images. We propose that our work will provide casual photographers with better experiences when using image editing software to improve image quality. Another promising avenue for future application involves the equipping of photo summarization systems with defect cues to focus more on defect-free photos.