基于卷积神经网络的单幅图像去雨-信号与信息处理专业论文.docxVIP

优秀毕业论文 精品参考文献资料 摘要通过 摘要 通过SCreen blend model对干净的清晰图像人工加雨，形成相应的有雨图像与无 雨图像对，建Z-Y{Jll练神经网络的图像库。相关实验表明，这种仿真获得的训练 库训练所得的卷积神经网络对于实际有雨图像雨线去除依然有效。 关键词：单幅图像去雨；变分法；卷积神经网络；深度学习 万方数据 ABSTRACTAffected ABSTRACT Affected by the bad weather,the images obtained by outdoor visual systems always degrade．Rain is one of the common bad weather．At home and abroad，there are some solutions about removal of rain，but most of it aims to videos．It Can’t apply to single image since no temporal information Can be obtained．Recently,the study of rain removal from a single image gradually receive more attention．However,the existing methods Call’t meet the demand of practical application．Because it makes a trade off between rain removal and image sharpness，and it has low computing efficiency． Therefore，this paper proposes three new methods to remove the rain streaks in a single image based on variational method and convolutional neural network，the main study content and results are as follows： 1．Removing rain from a single image method via gradient孽o regularization was proposed．A guided Zo smoothing filter is designed，which can simply remove rain．The filter introduces gradient粤o regularization，which make the filter smooth the input image according to the gradient magnitude of guided image．The final refined result is obtained by minimization operation between the input image and the coarse rain·free image． 2．Removing rain from a single image method via gradient modification is proposed，It puts forward a new energy function aiming the problem of loss information．The method assumes that rain is contained in horizontal gradient．So the sum of the vertical gradient and the smoothing horizontal gradient by guided filter Can be a modified gradient．A new energy function is proposed using the modified gradient．The function Can be solved with variational method，which call keep more details． 万方数据 ABSTRACT3 ABSTRACT 3．Removing rain from a single image method via CNN is proposed．The existing methods of single ima