A Research Of Single Image De-Raining Based On Generative Adversarial Nets

Rain is one of the most common weathers.Images and videos obtained from a rainy day by camera will have some problems of visual quality such as blurry,losing detail and color distortion,etc.It will cause performance reduction of computer vision algorithms such as image classification,object detection and semantic segmentation.Therefore,research on image de-raining has important significance for improving the stability of outdoor computer vision systems.Compared with video de-raining task,single image de-raining is a more challenging task due to the lack of temporal information.This article focuses on the single image de-raining,the contributions are as follows:1)A GAN(Generative adversarial network)-based single image de-raining network is proposed(GAN-SID)In this article,the single image de-raining is considered as an image-to-image translation task.In order to solve this problem,a novel GAN-based network is proposed.The network can pay more attention on rain steaks due to utilizing Squeeze-andExcitation(SE)module in transform part of generator.Unlike Batch Normalization(BN)used by most deep learning works,Instance Normalization(IN)is used to eliminate the interactions between different samples within the same batch.In addition,the proposed method uses a trainable perceptual loss function that enables the output image to close to the real image at various semantic levels.The experiments show that the proposed method outperforms many advanced image de-raining algorithms.The Structural Similarity Index(SSIM)is improved by 0.22,and the Peak Signal-to-Noise Ratio(PSNR)is improved by 3.02 d B.Compared with other algorithms in visual quality,the proposed method also performs better.2)An enhance image de-raining network based on super-resolution technology is proposed(GAN-SIDR)There is still some ambiguity in the results of proposed GAN-SID.To solve this problem,an enhanced rain removal network called GAN-SIDR is proposed.Before the input of generator,a rain pattern estimation network is proposed to estimate the rain steaks.The estimated rain steaks can indicate the reconstruction of clear image.After the generator,a refine network is proposed to eliminate blur.The estimation network uses a three-way dense residual network with different kernel size to estimate the distribution of rain patterns on different receptive fields.In refine network,a series of improved Res Net blocks are used as backbone to refine the result without increasing much computation.The experiments show that the proposed network can effectively remove the blur and eliminate the rain residue.