关键词: 卷积神经网络, 图像超分辨率重建, 残差网络

Abstract:

The traditional convolutional neural network uses end-to-end mapping between high/low resolution images based on sparsely expressed super-resolution images. The input image is high resolution and the output image is low resolution. SRCNN with three convolution layers has a certain reconstruction effect, but the field of receptivity is low. Therefore, a deeper network structure is proposed. This improvement makes the latter network layer have a larger field of perception, so that the result pixels can be inferred from more pixels. However, considering the influence of network structure deepening on transmission rate, the learning rate is improved by introducing a combination of local residual learning and global residual learning. The convergence rate is effectively accelerated by this method. The experimental results are verified. Compared with existing Bicubic, SRCNN and VDSR, the reconstruction effect has been improved in peak signal-to-noise ratio, structural similarity and visual effect.

Key words: convolutional neural network, image super-resolution reconstruction, residual network