Progressive integration network for single‐image rain removal

Abstract

Abstract Due to the inherent overlap between the background and rain streaks, most deraining methods have suffered greatly from the over‐smoothing effect, which restrains the deraining models’ performance to a great extent. To address this issue, a progressive integration network (PIN) is presented to reduce this effect. Specifically, in our network, the high‐level representations will be used to enhance the low‐level feature maps through our proposed global integration and local integration blocks. Moreover, based on the difficulty of prediction, a novel curriculum‐learning mechanism is explored to ease the procedure of training. Extensive experimental results demonstrate that the proposed method achieves considerable performance in contrast to other deraining methods.