To attend to the challenges of low efficiency in vehicle image detection from UAV aerial imagery, problems in small target function extraction, and the large parameter size of existing models, we propose the OSD-YOLOv10 algorithm, an improved variation based on YOLOv10n.The proposed algorithm includes a number of crucial developments: First, we utilize online convolutional reparameterization to build the OCRConv module and style a lightweight feature extraction structure, SPCC, to replace the traditional C2f module, thereby lowering computational load and parameter count.
Second, we integrate an efficient dual-layer feed-forward hybrid attention module to enhance the models feature extraction capabilities.We also construct a double small-target detection layer that integrates shallow and ultra-shallow functions to improve small-target detection.
We introduce the DySample vibrant upsampling module to enhance feature fusion in the neck network from a point sampling perspective.Extensive experiments on the VisDrone-DET2019 and UAVDT datasets show that OSD-YOLOv10 attains a 40.7% decrease in specification count and a 3.6% decline in floating-point operations, while improving precision and indicate typical accuracy by 1.3% and 1.6%, respectively.Compared to other YOLO series and lightweight models, OSD-YOLOv10 displays exceptional detection precision and lower computational intricacy, accomplishing an optimal balance between high accuracy and low resource consumption.These developments make it particularly appropriate for release in UAV onboard hardware for vehicle target detection jobs.
Code will be available online (https://github.com/Z76y/OSD-YOLO).
The full paper is offered here.Source: Nature scientific reports