多尺度密集时序卷积网络的单幅图像去雨方法

Single Image De-raining Method for Multi-scale Dense Temporal Convolutional Networks

雨滴会降低户外拍摄图像质量,影响图像视觉效果及后续图像分析工作。针对目前去雨算法存在颜色失真、去雨过度化等问题,为了提高计算机视觉算法在中、大雨天气下的准确性,提出多尺度DenseTimeNet(密集时间序列卷积神经网络)的单幅图像去雨方法。该网络由多个尺度DenseTimeNetBlock(密集时序卷积网络密集块)组成,通过卷积下采样技术得到不同尺度下雨线特征信息与降低图像维度后利用时域卷积寻找的时间维度特征信息。在不同维度下学习雨景图和无雨图之间的映射关系,网络主体由密集卷积块和残差网络组成,可加速算法收敛速度,更深度学习图像纹理特征,使特征信息在网络结构进行深度传播,可以更好地复原残损图像。在不同方向,不同大小的雨滴图像上对所提方法进行验证,实验结果表明,该方法相较于现有算法,图像去雨效果良好。

Raindrops will reduce the quality of outdoor captured images and affect the visual effect and subsequent image analysis. In order to improve the accuracy of computer vision algorithm in medium and heavy rain,a single-image rainfall method of DenseTimeNet(dense time series convolutional neural network) is proposed to solve the problems of color distortion and excessive rainfall. The network consists of multiple sizes of DenseTimeNetBlocks(dense time series convolutional network dense blocks). The feature information of rain line at different scales and the feature information of time dimension after image dimension reduction are obtained by convolution down-sampling technique. The mapping relationship between rain scene map and non-rain map is studied in different dimensions. The main body of the network is composed of dense convolution block and residual network,which can accelerate the convergence speed of the algorithm,further study the texture features of the image,make the feature information spread deeply in the network structure,and better recover the damaged image. The proposed method is validated on raindrop images of different sizes and directions. Experiment shows that this method has better image rainfall effect than the existing algorithms.