Enhanced YOLOv8 model for accurate and real-time remote sensing target detection
Year: 2025
Type: article
Abstract: Current remote sensing image object detection algorithms struggle with false positives, missed targets, and subpar accuracy. We proposed an improved YOLOv8 network (PIYN) solution to tackle these issues through targeted modifications to the YOLOv8 architecture. The backbone network of YOLOv8 uses a Cross-Stage Partial (CSP) with 2 convolutions, called a faster C2f module. Firstly, the C2f module is infused with the Efficient Multi-Scale Attention (EMA) mechanism, enhancing the module's ability to process information at various scales. Secondly, a compact Path Aggregation Network (Compact-PAN) structure within the neck network reduces the computational complexity of the model. Finally, replacing the Complete Intersection over Union (CIoU) loss function with Weighted Intersection over Union (WIoU) refines the model's detection accuracy. We also applied K-fold cross-validation on the dataset to mitigate the overfitting issue. Extensive Dataset for Object deTection in Aerial images (DOTA) and Dataset for Object Recognition in Optical Remote sensing imagery (DIOR) datasets experiments reveal PIYN's effectiveness: a 2.43% and 2.56% mAP increase over YOLOv8, respectively, coupled with a 4.49% GFLOPs reduction. This demonstrates PIYN's ability to boost accuracy while maintaining efficiency and solidifies its progressive and practical impact, with implications for smart city scenarios.
Source: Computer Standards & Interfaces
Authors: Israr Ahmad, Fengjun Shang, Muhammad Faheem
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,670
OpenAlex ID: W4416451744