Unsupervised double-branch rain-carrying and rain-free image processing method

Abstract

The invention discloses an unsupervised double-branch rain-free image processing method, which comprises the following steps of: circularly processing an originally input image through a deep network based on a long-short term memory module, and effectively acquiring rain stripe feature representation of the original image through a double-branch and circular consistency structure formed by the double branches; meanwhile, on the basis of processing of a double-branch generator and a discriminator, through an unsupervised deep learning method, the obtained rain stripe features with attention are subjected to merging operation with an original image, and thus a rain-removed image and a rain-carried image are obtained; secondly, re-processing based on circulation consistency, taking the rain-removed image and the rain-free image obtained in the first step as input, sending the input to another pair of generator and discriminator for processing, and after circulation consistency processing, enabling the network to be more stable and improving the accuracy of the rain-free image and the rain-carrying image to a certain extent.

Description

technical field [0001] The invention relates to the field of image processing methods, in particular to an unsupervised two-branch image processing method with rain and without rain. Background technique [0002] Images and videos acquired from outdoor vision systems (such as autonomous driving, surveillance, and person / vehicle tracking) in rainy weather are often degraded by the weather phenomenon of rain. This degenerate situation will directly reduce the effect of subsequent high-level tasks, such as object detection, image recognition and saliency detection. Therefore, the single-image deraining task will have a wide range of applications. However, single-image deraining is still an important but challenging problem due to the complexity of rain streaks in practical processing and the uncertainty of its solution. [0003] The rainy image mainly contains rain streaks and raindrops, which form a rain mask before the real image. For torrential rains, they may create a ha...

AI Technical Summary

Problems solved by technology

Even for synthetic datasets, it is still difficult to compute an exact mapping between rain and rain-free images without some strict pairwise constraints; (2) for existing synthetic datasets (such as Rain1400 and Rain800), Supervised models still cannot achieve ideal recovery results, i.e. there is still a lot of room for improvement

Images