UAV remote sensing detection and target recognition based on SCP-YOLO

Lihui Wang,Zhan Miao,Endong Liu

DOI: https://doi.org/10.1007/s00521-024-09938-x

2024-06-21

Neural Computing and Applications

Abstract:Identifying small, overlapping wheat ears in UAV images continues to be a difficult task. This paper proposes SCP-YOLO, a novel detection model that addresses this limitation. Initially, the dataset comprises remote sensing images of wheat kernels captured at two periods and three altitudes. Using the YOLOv8n network as a baseline, SCP-YOLO processes the network's low-resolution feature layer using the space-to-depth (SPD) approach. At the stage of feature fusion, the Context-Aggregation structure is executed to facilitate the aggregation and interaction of data on the feature map on a global scale. The PConv method ingeniously implements the lightweight detection head structure. On top of that, a new detection scale that integrates more superficial information with location data is positioned. The experimental outcomes demonstrate that the proposed method outperformed several established state-of-the-art detection models by achieving a detection speed of 90 frames per second and an AP@50 value of 96.3%. Compared with the baseline network, the AP@0.5, and AP@0.5:95 exhibited respective increases of 2.5% and 6.3%, respectively. Experimental results indicate that the methodology demonstrates exceptional robustness for six scenario datasets. About counting, it is exact and capable of quantifying wheat ears in images acquired through remote sensing.

computer science, artificial intelligence