基于多尺度编解码网络的道路交通模糊图像盲复原

Blind Restoration of Road Traffic Blurred Image Based on Multi-scale Encoder-decoder Network

提出一种多尺度编解码深度卷积神经网络结构,使用生成对抗的思想对模糊图像直接进行盲复原。首先,设计一种优化多尺度残差块应用在编解码器内部,在减少参数量的同时提高了网络非线性表达能力;其次,分别计算多尺度网络每层对应的L2损失,确保逐级去模糊后的图像更加接近真实图像;最后,在GoPro数据集和真实道路交通模糊图像上进行仿真,结果表明,所提方法能够得到清晰度更高的复原结果。

A multi-scale encoder-decoder depth convolution neural network structure was proposed,which used the idea of generative adversarial to blindly restore blurred images directly.Firstly,an optimization multi-scale residual block was designed and applied inside the encoder-decoder,which reduced the number of parameters and improved the network′s nonlinear expression ability.Secondly,the L2 loss corresponding to each layer of the multi-scale network was calculated separately to ensure that the image after stepwise deblurring was closer to real image.Finally,the simulation was carried out on the GoPro data set and the blurred image of real road traffic.And the experimental results showed that the proposed method could obtain higher resolution results.