关键词: 超分辨率, 轻量化, 异构卷积, 特征交叉

Abstract:

In recent years, the deep learning technology has significantly improved the performance of single image super-resolution reconstruction (SISR). However, SISR algorithm based on thedeep learning technology has some problems, such as alarge quantity of model parameters, complex network structure and high resource consumption. In order to solve these problems, this paper proposes a lightweight image super-resolution reconstruction algorithm based on spatial feature cross-fusion. The algorithm uses the multiple local feature fusion modules and the feature cross-enhancement modules to form a nonlinear mapping unit, and learns thestepwise polymerization image features through residuals to extract more accurate residual information. At the same time, the symmetric structure is used to map thefeatures to two branches, and the high-frequency components are extracted by performing thefeature intersection and themultiplication of corresponding elements, which refines thefeatures and increases the network nonlinearity. In each feature cross-enhancement module, the heterogeneous convolution is used instead of standard convolution to split and fuse two branches, which effectively reduces the parameters of the network and makes the network achieve a relative balance between parameters and performance. Finally, a multi-level integration module is used to enhance the correlation of features in different stages. Experiments on benchmark data sets show that the proposed method not only reduces the model parameters, but also achieves good results in peak signal-to-noise ratio and structural similarity, and the edge structure of the reconstructed image is complete, the overall outline is clear and the details are more abundant.

Key words: super-resolution, lightweight, heterogeneous convolution, feature crossing