Lightweight image super-resolution with feature cheap convolution and attention mechanism

Since deep learning is introduced into the field of super-resolution (SR), many deep learning-based SR methods have been proposed and achieved good results. At present, most neural networks use ordinary convolution and deeper neural layer in image super-resolution reconstruction in order to achieve better results. Therefore, most existing models have a massive amount of parameters and calculation, which limits the practical application. To solve this problem, in this paper, we propose an efficient feature cheap convolution SR (FCCSR) model, which consists of several multi-level information fusion Blocks (MIFB). In the MIFB, the features are extracted by combining normal convolution and cheap convolution, and the features obtained from each layer of convolution are fused. Under limited parameters and reconstruction speed, it enables the whole model to obtain the features from shallow to deep layers in the low-resolution (LR) images. Secondly, a novel lightweight channel attention module is designed to obtain fewer parameters and better performance. Finally, we add a gradient loss function to the original L1 loss function, making the model pay more attention to the high-frequency part of the LR image and making the reconstructed image’s texture details more straightforward. The experimental results show that the proposed model has reached the most advanced level regarding image quality, memory consumption, parameter number, and calculation amount. The operation speed of FCCSR is roughly the same as that of the current STOA models, and the number of parameters is only 0.63 M. On some data sets, the best performance of FCCSR can be improved by 7%.