Adaptive convolution residual error correction single image rain removal method

Abstract

The invention provides an adaptive convolution residual error correction single image rain removal method, which is characterized in that a rain line correction factor (refine factor) is added, an existing rain map model is improved, and the influence of the rain line on each pixel in the rain map is described more accurately. An adaptive selection convolution network (SKNet) is constructed, information of corresponding dimensions of different convolution kernels is adaptively selected, the information of different convolution kernels is further learned and fused, and the expressive force of the network is improved. And finally, an adaptive convolution residual error correction network (SKRF) network is built for directly learning the rain line graph and a residual error correction coefficient (RF), thus reducing a mapping interval, and reducing background misjudgment. Compared with an existing method, the method can obtain higher accuracy. The picture result is improved in objective index and rain line removal effect of the generated picture. According to the invention, feature information of channels corresponding to convolution kernels of different sizes can be adaptively selected, and the rain influence on each pixel point can be achieved more accurately.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method for removing rain from a single image with residual correction based on deep learning adaptive convolution. Background technique [0002] Computer vision relies on image quality, and images collected from outdoors are often affected by bad weather including rain, snow, and fog. Rainy day is one of the most common weather in nature. Raindrops will form chaotic rain lines in the air. At this time, white lines with high pixel values ​​will appear in some areas of the collected images. At the same time, rain will form water mist in the air. These factors The sight of urban people is affected. The single image rain removal algorithm has certain application value in technologies such as automatic driving and video surveillance. [0003] The image deraining algorithm at the current stage mainly refers to the following two directions: single image deraining algorithm,...

AI Technical Summary

Problems solved by technology

[0004] Existing research methods can initially achieve the effect of rain removal. Among them, most of the traditional rain removal algorithms are based on mathematical modeling for pixel-level optimization solutions, which are difficult to guarantee in terms of running speed and are not very practical.

At the same time, the visual effect of the picture after removing the rain is not good, and there are often residual rain lines and background loss.

[0005] The application of deep learning-based methods has greatly improved the performance of the algorithm, but the expressive ability is still limited. Some existing algorithms use relatively complex structures, including increasing the number of network layers and branches, to increase the expressive power of the network. The network is too complex

[0006] The process of network deraining includes the detection and removal of rain lines and the restoration of the background. Some algorithms use image decomposition technology to expect rain lines to be easier to detect, but this will introduce additional steps, and this operation will result in background details in the generated image lost

Images