A Rotating Object Detection Method Based on Anchor-Free in Remote Sensing Scene Images

Abstract

The core of the object detection task is to be able to accurately classify and regress the object position. The current two-stage algorithm based on convolutional neural networks has achieved good results in horizontal box detection, but the effect in rotating bounding box detection is limited, and the anchor-based detection algorithm has problems, such as complicated hyper-parameter design and imbalance between positive and negative samples. In order to better solve the above problems, the model in this paper adopts a single-stage anchor-free method, uses the high resolution network(HRNet) to obtain the high-resolution feature map of cross fusion, and uses a rotating Gaussian kernel for predicting center point, which can increase the number of positive samples at the center point, simultaneously realize the improvement of detection accuracy and speed. Experimental results on HRSC2016 and DOTA datasets, which demostrates that our method exceed the state-of-art models, and indicates the effectiveness of the proposed methods.