CRKD-YOLO: Cross-Resolution Knowledge Distillation for Low-Resolution Remote Sensing Image Object Detection

Abstract

The majority of advanced remote sensing object detection technologies excel in accurately detecting objects from high-resolution images. However, in practical scenarios, it is often necessary to detect objects in images of varying resolutions due to differences in imaging equipment. When dealing with lower-resolution images, the limited detailed information and blurry boundaries lead to a noticeable decrease in detection accuracy. To address this problem, we propose an efficient object detection method for low-resolution remote sensing images based on the YOLO detector, named CRKD-YOLO. The method constructs a cross-resolution knowledge distillation (CRKD) framework to resolve the issue of feature mismatch, enabling the model with low-resolution inputs to learn more refined feature representations from high-resolution images. Furthermore, to effectively leverage the limited detailed information in low-resolution images, we propose the backbone augment feature pyramid network (BAFPN). It enhances detection accuracy for low-resolution remote sensing images while making the model more lightweight. Massive experiments on DOTA, DIOR, NWPU VHR-10, DroneVehicle, and VEDAI demonstrate that our CRKD-YOLO achieves significant improvements, even achieving higher accuracy compare to training and testing high-resolution images with baseline. Our code is published at https://github.com/Jianfantasy/CRKD-YOLO