Detect Deeply Only Once: Arbitrary-Oriented Object Detection in Remote Sensing Images Based on Improved YOLOv5

To solve the problem that the existing arbitrary-oriented object detection cannot satisfy both high accuracy and high efficiency, a model called Detect Deeply Only Once is proposed. The model is based on the YOLOv5s_CSL network and is improved in three ways: (1) Three attention mechanisms are separately added to the appropriate positions of the model. All three attention mechanisms pay more attention to the features of interest and improve the detection accuracy. (2) The repeated weighted bi-directional feature pyramid network and path aggregation network are combined, which strengthens feature fusion ability among multiple layers. (3) Loss rank mining is employed after the final feature map, which filters some elements representing easy examples in the final feature map and forces the detectors to concentrate on hard examples during training. Meanwhile, the role of each module is verified on the HRSC2016 dataset, and the proposed method shows good superiority in terms of higher detection accuracy than the benchmark model and the state-of-the-art models on both HRSC2016 dataset and UCAS_AOD dataset.

I. Introduction

Due to the rapid development of remote sensing technology, the data sources of remote sensing images (RSIs) have exploded in recent years, and high-resolution and high-quality RSIs have been generated in response. Object detection in the RSIs is a significant task of computer vision, which has great application value, especially for monitoring of ecological environment, construction of smart cities, and striking of military targets. However, there are the problems of complex background, arbitrary angle, small objects, etc. in the RSIs, which lead to the task of object detection in the RSIs is full of challenges. There is an imminent need to develop specific models to maximize the detection accuracy [1].