目前, 基于卷积神经网络的超分辨率重建方法具有参数数量大, 时效性偏低, 边缘细节信息丢失的缺陷。针对该问题, 提出基于边缘修正的多尺度卷积神经网络超分辨率重建算法。首先在训练阶段, 利用低频信息的冗余性设置参数共享层, 将同一组滤波器应用到不同放大倍数的训练网络中, 构建多任务学习框架; 然后在重建阶段, 从样本训练库中学习可以高分辨率图像边缘修正系数, 采用邻域像素差值线性运算将边缘系数与重建的高分辨率图像进行融合, 矫正边缘信息的偏差, 弥补丢失细节; 最后根据随机梯度下降法和反向传播法, 利用梯度不断更新权重参数使网络达到最优化。实验结果表明, 该算法的重建效果较为显著, 边缘锐度较高, 消除了模糊和锯齿现象, 并且通过参数共享大幅减少参数量, 满足实时性的要求。

At present, the super-resolution reconstruction methods based on convolutional neural network have the defects of large amount of parameters, low timeliness and loss of edge detail information. In order to solve these problems, we propose a super-resolution reconstruction algorithm of multiscale convolution neural network based on edge correction. Firstly, in the training phase, we set the parameter sharing layer by using the redundancy of low frequency information, In other words, the same set of filters applied to different magnification training networks to build the multi-task learning framework. In the reconstruction phase, the edge correction coefficient of high-resolution image is learned from the sample training library. The neighborhood pixel difference is used to fuse the edge coefficient and the reconstructed high resolution image, and to correct the deviation of the edge information and make up for the missing details. Finally, according to the stochastic gradient descent and back-propagation, we use the gradient to continuously update the weight parameters to make the network reach the maximum optimization. Experimental results show that the proposed algorithm has the significant reconstruction effect, high edge sharpness, elimination of blurring and aliasing, and greatly reduces the amount of parameters through parameter sharing to meet real-time requirements.