MFADet-A multi-level fusion of rotation detection frame model algorithm
He Xiao, Shenghua Lin, Ningyi Xiao, Yaosheng Chen, Jiahui Yang · 2024
Objects in remote sensing images are often arranged in arbitrary orientations and are relatively small. Common object detection algorithms typically employ horizontal detection methods, which are insufficient for remote sensing scenarios. Therefore, this paper proposes a multi-level fusion rotation box detection network, MFADet. This network leverages multi-level feature fusion to extract useful features pertinent to target information and enhances these features using an attention mechanism, thereby improving detection accuracy in complex backgrounds. Furthermore, to better accommodate the distinctiveness of rotation boxes, the method employs feature channels and feature interpolation. In the fine-tuning stage, an area interpolation method is used to average the regions of interest. During the rotated box feature extraction stage, a combination of fully connected layers and convolutional layers is employed, and CIoULoss is utilized as the regression loss function. Evaluations on the large-scale remote sensing dataset DOTA show an average precision of 78.03%, while on the ship remote sensing dataset HRSC2016, an average precision of 91.62% was achieved. Comparative experiments demonstrate the effectiveness of the MFADet network.