PFRNet: A Small Object Detection Method Based on Parallel Feature Extraction and Attention Mechanism

Abstract

To address the challenge of detecting small objects in aerial and satellite remote sensing images with low-resolution, we propose a high-precision object detection method based on PFRNet. PFRNet incorporates parallel feature extraction branches and a progressive feature refinement mechanism, significantly enhancing the model’s ability to perceive detailed features. In addition, PFRNet introduces the spatial pyramid pooling fusion with spatial attention (SPPFSPA) module, which integrates multi-scale features with an attention mechanism, enabling the model to better focus on areas of interest, thereby improving detection performance. Results demonstrate that PFRNet achieves outstanding detection accuracy, markedly outperforming other algorithms, particularly in small object detection. Visualization analysis reveals that the PFR module effectively captures richer and more comprehensive visual features in images, providing robust input for subsequent detection tasks, which is crucial for PFRNet’s superior performance. Overall, the proposed PFRNet model makes significant strides in small object detection in UAV aerial and satellite remote sensing images, offering strong support for applications such as intelligent transportation and precision agriculture.