MSPA-YOLO: Remote Sensing Target Recognition Method Based on Improved YOLOv10

Object detection and recognition in optical remote sensing imagery play a crucial and indispensable role in civilian applications and military operations. The existing detection algorithms are plagued by a multitude of limitations. Specifically, they struggle to strike a balance between accuracy and efficiency while effectively mitigating the interference posed by intricate and cluttered backgrounds. In this paper, MSPA-YOLO model based on YOLOv10 algorithm is proposed for remote sensing image recognition. This model improves the attention mechanism of feature extraction part, extracts and optimizes features from different scales of receptive field ranges by Patch-Aware module (PA) combined with multi-branch structure and self-attention mechanism, and suppresses complex background interference by improved spatial and channel parallel attention mechanism. In addition, a Lightweight Attention Feature Fusion module (LAF) is designed to replace the multi-layer traditional convolution operation with spatial and channel attention mechanisms to improve feature utilization and detection efficiency while ensuring detection accuracy. Finally, experiments on DOTA dataset show that our proposed approach outperforms existing methodologies, offering a more effective solution to strike a balance between accuracy and efficiency in remote sensing image target detection.