Pixel Distillation: A New Knowledge Distillation Scheme for Low-Resolution Image Recognition

The great success of deep learning is mainly due to the large-scale network architecture and the high-quality training data. However, it is still challenging to deploy recent deep models on portable devices with limited memory and imaging ability. Some existing works have engaged to compress the model via knowledge distillation. Unfortunately, these methods cannot deal with images with reduced image quality, such as the low-resolution (LR) images. To this end, we make a pioneering effort to distill helpful knowledge from a heavy network model learned from high-resolution (HR) images to a compact network model that will handle LR images, thus advancing the current knowledge distillation technique with the novel pixel distillation. To achieve this goal, we propose a Teacher-Assistant-Student (TAS) framework, which disentangles knowledge distillation into the model compression stage and the high resolution representation transfer stage. By equipping a novel Feature Super Resolution (FSR) module, our approach can learn lightweight network model that can achieve similar accuracy as the heavy teacher model but with much fewer parameters, faster inference speed, and lower-resolution inputs. Comprehensive experiments on three widely-used benchmarks, \ie, CUB-200-2011, PASCAL VOC 2007, and ImageNetSub, demonstrate the effectiveness of our approach.