Learning CNN with Explicit Reflection Model to Remove Reflections in the Wild

When photographing through glass surfaces, the quality of the photo may deteriorate due to reflections on the glass. There are many ways to deal with this, using multiple images obtained under different conditions, but much more interesting is the problem of removing reflections on a single image, which is currently one of the most difficult problems in computational photography. In recent years, deep learning methods have surpassed traditional approaches and defined the current state in this field, but these methods use synthetic data for training and have insufficient generalization ability to work with real data. To combat this, we propose a combination of an implicit reflection model and a deep neural network. This method generalizes well to real data, using only synthetic data for training. Experimental results show that the proposed approach is superior to modern methods based on open test datasets SIR2, CEILNetandDPRR. The paper considers the latest methods for the reflection removal problem using neural networks, methods for generating data, and proposes a new model and method for stabilizing learning. A detailed description of the training procedure is provided. The comparison of work on several sets of test data was made, during which it was shown that most methods can produce undesirable artifacts, and the results were analyzed.