Consistency- and dependence-guided knowledge distillation for object detection in remote sensing images

Abstract

As one of the challenging tasks in the remote sensing (RS), object detection has been successfully applied in many fields. Convolution neural network (CNN) has recently attracted extensive attention and is widely used in the natural image processing. Nevertheless, RS images have cluttered scenes compared with natural images. As a result, the existing detectors perform poorly in RS images, especially with the complicated backgrounds. Moreover, the detection inference time and model volume of detectors in RS images often go unrecognized. To address the above issues, this study proposes a novel method for object detection in RS images, which is called the consistency- and dependence-guided knowledge distillation (CDKD). To this end, the spatial- and channel-oriented structure discriminative modules (SCSDM) are put forward to extract the discriminative spatial locations and channels to which the teacher model pays attention. SCSDM improves the feature representation of the student model by effectively eliminating the influence of noises and the complicated backgrounds. Then, the consistency and dependence of the features between the teacher model and the student model are constructed under the guidance of SCSDM. Experimental results over public datasets for RS images demonstrate that our CDKD method surpasses the state-of-the-art methods effectively. Most of all, on the RSOD dataset, our CDKD method achieves 92% mean average precision with 3.3M model volume and 588.2 frames per second.