譚臺哲 柳博



摘要:為更好地將圖像去雨算法應(yīng)用在戶外監(jiān)控、手機移動終端上,提出一種基于局部空間注意力機制的輕量級卷積神經(jīng)網(wǎng)絡(luò)。將圖像去雨看作殘差學(xué)習(xí),既有利于從有雨圖中去掉雨滴,又便于模型的訓(xùn)練與優(yōu)化。深度可分離卷積作為模型提取特征的卷積操作,在不降低模型的性能情況下,顯著降低模型的參數(shù)量與計算量。局部空問注意力模塊利用空洞卷積提供較大的感受野來提取豐富的語義信息,有利于雨滴的檢測與去除。在多個公開的數(shù)據(jù)集上進行對比與測試,證明模型去雨效果較好且速度較快。
關(guān)鍵詞:單幅圖像去雨;分組卷積;空洞卷積;空問注意力;殘差學(xué)習(xí)
中圖分類號:TP18 文獻標(biāo)識碼:A
文章編號:1009-3044(2020)20-0028-04
Single Image De-Rain Method Based on Croup Convolution and Spatial Attention Mechanism
TAN Tai-zhe, BO Liu
(School of Computers, Guangdong University of Technology, Guangzhou 510000, China)
Abstract : In order to hetter apply the image de-raining algorithm to outdoor monitoring and mobile terminals , a lightweight convolu-tional neural network based on the local spatial attenticm mechanism is proposed. Taking the image to rain as residual learning isnot only beneficial for removing raindrops from the raining image, but also for training and optimization of the model. depthi~'iseconvolution. as a convolution operation for extracting features, significantly reduces the amount of parameters and calculations ofthe model without reducing the performance of the model. The local spatial attention module uses dilate convolution to provide alarger receptive fielcl to extract rich semantic information. which is conducive to the detection and removal of raindrops. Compari-son and testing on multiple public data sets prove that the model has better rain removal effect and faster speed.
Key words : single image derain;group convolution; dilate convolution; spatial attention: residual leaming
戶外視頻監(jiān)控、無人駕駛、白然場景下的文本識別等基于計算機視覺算法的應(yīng)用容易受到天氣的影響,由于現(xiàn)有的算法設(shè)計以及訓(xùn)練模型所使用的數(shù)據(jù)集都基于天氣情況較好這一假設(shè),在如下雨、下雪、霧天等情況下算法的性能會有所降低。另一方面,隨著手機等終端設(shè)備的普及,雨天拍攝的照片由于雨痕、雨滴、雨霧的存在影響拍照的主體,影響圖像的美感,因此對基于視覺算法的應(yīng)用而言去雨算法可以作為算法應(yīng)用的預(yù)處理,提升算法在不良天下的性能,提升算法的魯棒性。對人眼視覺感知而言,圖像去雨可以將影響拍照主題的雨滴去除掉,恢復(fù)主體的信息。圖像去雨可以分為單幅圖像去雨和視頻圖像去雨,視頻圖像去雨因為有連續(xù)圖像信息可以使用,相較于單幅圖像難度較低,所以視頻的去雨的研究較為成熟?!?br>