REMOVAL OF RAIN COMPONENTS FROM SINGLE IMAGES USING A RECURRENT NEURAL NETWORK

Context. Removing the undesirable consequences of rain effects from single images is an actual problem in many computer vision tasks, because streaks can significantly degrade visual quality and seriously interfere with operation various intelligent systems, which are used for their processing further analysis.
 Objective. The goal work to develop a method detecting removing images, based on using convolutional neural network recurrent structure.
 Method. main component proposed network, has multi-stage structure. A feature this architecture use repeated blocks (layers), at output you get intermediate result «cleaning» original image. Moreover each next layer we image less influence components than previous one. Each contains two independent sub-networks (branches) parallel processing. branch designed detect remove effect attention improve speed up process (for map formation).
 Results. An approach been developed automatically images. “cleaning” structure, was trained Rain100H Rain100L datasets. results experiments, testifies effectiveness expediency solving practical tasks pre-processing “contaminated” presented.
 Conclusions. advantage that architecture, it allows be potentially applied under conditions limited computing resources. successfully development systems area monitoring surveillance cameras, autonomous vehicles control, aerial photography results, etc. In future, should considered possibility forming separate sub-network eliminate blurring train datasets contain samples different rain, will make more «resistant» forms increase “cleaning”.