Single image rain removal system based on convolutional neural network

• Main Authors: Yang,Shih-Yi, 楊適翊
• Other Authors: Chu,Yuan-Sun
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/regg9k

碩士 === 國立中正大學 === 電機工程研究所 === 107 === In recent years, due to the rapid development of deep learning, the accuracy of object recognition and object detection has been greatly improved. Nowadays, smart phones are also equipped with smart AI chips, which enable smart phones to have intelligent power saving functions and camera smart scene recognition. Such functions, the accuracy of these object recognition and object detection can be high in the case of general sunlight or indoors without other environmental influences, but in the case of outdoor rain or fog, Its accuracy will affect the identified or detected objects due to the effects of rain and fog, resulting in a significant drop in accuracy. In this paper, we propose a classifier that classifies the dense relationship of rain in the image through the classifier, and then uses different treatment methods for dense rain to effectively remove the rain, so that the rain can be eliminated as much as possible. This paper also proposes two different methods of rain removal, one for the traditional image to remove rain and the other for the learning mode to remove the rain, to prove that the reduction effect will be better after adding the classifier. Compared with the rain-removal system without classifier, the two types of rain-removal systems have an average increase of about 0.02 in SSIM, an average increase of about 2.3 dB in PSNR, and the classification-type learning-type rain-removal system proposed by the rear wheel is better. The degree of reduction has an average SSIM of 0.8796 and an average PSNR of 28.59 dB. It is better than the current general learning type rain removal system.