关键词: 超分辨率, 深度学习, 空间元学习, 残差密集模块, 权值预测

Abstract: For the problem that the existing deep-learning based super-resolution reconstruction methods mainly study on the reconstruction problem of amplifying integer times, not on the cases of amplifying arbitrary times (e.g. non-integer times), a super-resolution reconstruction method with arbitrary magnification based on spatial meta-learning was proposed. Firstly, the coordinate projection was used to find the correspondence between the coordinates of high-resolution image and low-resolution image. Secondly, based on the meta-learning network, considering the spatial information of feature map, the extracted spatial features and coordinate positions were combined as the input of weighted prediction network. Finally, the convolution kernels predicted by the weighted prediction network were combined with the feature map in order to amplify the size of feature map effectively and obtain the high-resolution image with arbitrary magnification. The proposed spatial meta-learning module was able to be combined with other deep networks to obtain super-resolution reconstruction methods with arbitrary magnification. The provided super-resolution reconstruction method with arbitrary magnification (non-integer magnification) was able to solve the reconstruction problem with a fixed size but non-integer scale in the real life. Experimental results show that, when the space complexity (network parameters) is equivalent, the time complexity (computational cost) of the proposed method is 25%-50% of that of the other reconstruction methods, the Peak Signal-to-Noise Ratio (PSNR) of the proposed method is 0.01-5 dB higher than that of the others, and the Structural Similarity (SSIM) of the proposed method is 0.03-0.11 higher than that of the others.

Key words: super-resolution, deep learning, spatial meta-learning, residual dense module, weight prediction