关键词:
Object recognition
摘要:
Remote sensing-oriented object detection is a challenging task in computer vision because traditional horizontal bounding box representations cannot accurately locate remote sensing targets that have various scales, arbitrary orientations, and dense arrangements. The widely used five-parameter oriented bounding box representation increases the complexity of model training because of the periodicity of the orientation angle and the interchangeability of edges. To address these issues, this study proposes an elliptical equation-based remote sensing-oriented object detection network called EllipticNet. EllipticNet decouples the problem of predicting the orientation angle into two subproblems quantitative angle regression and rotation direction classification. The proposed method combines the major and minor axes of the ellipse and its center to describe the remote sensing-oriented target accurately, thereby overcoming the boundary discontinuity problem of five-parameter oriented bounding box representation. Additionally, a novel ellipse-constrained loss function is designed to enhance the intrinsic geometric relationship between ellipse parameters, thus improving the robustness of EllipticNet training. A layer-wise dilated spatial pyramid pooling module is also proposed to substantially enhance EllipticNet’s ability to represent multiscale features. The proposed method is validated on three commonly used public remote sensing datasets, namely, DOTA, HRSC2016, and UCAS_AOD. Results demonstrate that the proposed method is competitive in terms of performance and efficiency and has practical value in remote sensing-oriented object detection. © 2025 Science Press. All rights reserved.