CFDN: cross-scale feature distillation network for lightweight single image super-resolution

Feature distillation plays an important role in lightweight single image super-resolution (SISR) by retaining and extracting hierarchical features step-by-step. This allows the network to possess a deep and compact structure, enabling it to achieve lightweight while maintaining good performance. However, the features prematurely retained by the distilled branches cannot be fully utilized, making it difficult for the model to effectively learn the multi-scale semantics and details. To solve above problems, we propose a Cross-scale Feature Distillation Network (CFDN), which consists of the basic module called Cross-scale Feature Distillation Block (CFDB) and sub-module called Multiple Receptive field Aggregation Block (MRAB). Different from the previous distillation structures that lost insight of extracting multi-scale information due to the use of simple distillation units, the proposed CFDB module designed with a novel distillation strategy can capture multi-scale information, thereby providing the diversity of features to the retained and purified branches. In each retained/distilled branch of CFDB, the sub-module MRAB equipped with pixel attention and dilated convolutions is designed to capture multi-scale semantics and details by enlarging receptive field processively. Experiments on five datasets show that the proposed network achieves superior performance over the other state-of-the-art lightweight SR methods.