Residual Super-Resolution Single Shot Network for Low-Resolution Object Detection

For object detection in computer vision, detection models trained by high-resolution images often fail to recognize or localize objects on low-resolution images. To tackle this problem, we propose a fully convolutional network named residual super-resolution single shot network (RSRSSN). RSRSSN consists of two sub-networks, super-resolution sub-network and detection sub-network. The super-resolution sub-network in RSRSSN is achieved by stacking of identity residual blocks while the detection sub-network adopts the single shot multibox detector (SSD). Based on multi-task learning, we design a novel objective function called feature maps multibox loss to enforce low-resolution images to produce similar feature maps with their corresponding high-resolution ones. This information sharing mechanism is proved to be critical for solving the resolution mismatch problem in the experiments. A two-step training scheme is also proposed to train the RSRSSN in an end-to-end manner. Without any data augmentation, RSRSSN outperforms the SSD on both down-sampled PASCAL VOC and MS COCO in real-time object detection.