AFOD: Two-stage object detection based on anchor-free remote sensing photos

Abstract

Aerial photo target detection in remote sensing utilizes high-resolution aerial images, along with computer vision techniques, to identify and pinpoint specific objects. To tackle imprecise detection caused by the random arrangement of objects, a two-stage model named anchor-free orientation detection, based on an anchor-free rotating frame, has been introduced. This model aims to deliver encouraging outcomes in the analysis of high-resolution aerial photos. Initially, the model adopts faster Region with CNN feature (faster R-CNN) as a foundational framework. Eliminating the anchor configuration and introducing supplementary angle parameters accommodates the identification of rotating frame objects. Subsequently, it integrates the spatial attention module to seize global semantic information and establish an approximate detection frame with certainty. Additionally, the channel attention module extracts critical features from the semantic data within the predicted frame. Ultimately, the faster R-CNN detection head is employed to refine, leading to enhanced model outcomes and further bolstered regression and classification precision. After validation, the accuracy of the model detection reaches 88.15 and 77.18% on the publicly accessible aerial remote sensing datasets HRSC2016 and DOTA, respectively, which is better than other advanced rotating frame object detection methods.

Keywords

• aerial remote sensing photos
• spatial attention
• channel attention
• anchor-free
• rotating object detection