基于多尺度和循环生成对抗的连接式去雨网络

Connected Deraining Network Based on Multi-Scale and Cyclic Generative Adversarial

【目的】图像去雨能够作为其他计算机视觉任务的预处理步骤,使自动驾驶、目标识别等其他计算机视觉任务的结果进一步得到提升。【方法】本文将多尺度信息交换与循环生成对抗网络进行了连接,提出的方法根据训练步骤分为两个部分,首先通过多尺度信息交换得到雨水条纹信息进行初去雨,然后通过循环生成对抗网络对初去雨图像进行进一步增强,以得到效果最佳的去雨图像。【结果】该方法能够有效地去除图像中的雨水信息,恢复出清晰的图像。本文去雨结果在PSNR(Peak Signal to Noise Ratio)和SSIM(Structural Similarity)评价指标上取得了较高的结果,能够更好地保留图像的细节。【结论】通过在合成数据集以及真实图像上与其他图像去雨方法的结果进行对比,本文的方法取得了较好的效果,能够更好地为其他计算机视觉任务提供支持。

[Objective]Image deraining can be used as a preprocessing step for computer vision tasks so that the results of computer vision tasks such as automatic drive and target recognition can be improved.[Methods]In this paper,multi-scale information exchange is connected with a cyclic generative adversarial network.The proposed method is divided into two parts according to the training steps.First,the rain streak information is obtained through multi-scale information exchange for initial rain removal.Then the image with initial rain removal is further enhanced by a cyclic generative adversarial network so as to obtain the best rain removal effect.[Results]This method can effectively remove the rain information in the image and restore a clear image.This method presented in this paper has achieved good rain removal results in PSNR(Peak Signal to Noise Ratio)and SSIM(Structural Similarity)evaluation indexes and can better preserve the details of the image.[Conclusions]By comparing the results on synthetic datasets and real images with other image rain removal methods,this method has achieved better results and can provide better support for other computer vision tasks.