PBM‐YOLO: A Performance Balanced Floating Garbage Detection Model for Water Surface Environments

ABSTRACT The detection of floating garbage on the water surface is essential for ecological protection and resource recycling. However, it faces challenges such as the difficulty of extracting features from small objects, complex background interference, and hardware resource limitations. This paper proposes PBM‐YOLO, a performance‐balanced model based on the YOLOv8n framework, incorporating three key modules: the frequency‐spatial fusion feature extraction module (FSEM), the context feature fusion module (CFFM), and the lightweight detection head (LWDH). FSEM adopts a dual‐branch architecture, wherein the fast Fourier transform (FFT) captures global frequency features, while the Scharr operator enhances spatial edge information. The fused features are further refined via a coordinated attention mechanism to amplify target characteristics, significantly improving detection in complex backgrounds. CFFM performs an adaptive fusion of multi‐resolution features through lightweight channel mapping and a dynamic residual mechanism, thereby enhancing the representation of small objects. LWDH compresses model parameters to 1.89 M by leveraging parameter sharing, detail‐enhanced convolution (DEConv) and a channel‐wise learnable scaling and bias mechanism (BiScale), facilitating deployment on edge devices. Experiments on the FloW‐Img dataset and Self‐dataset demonstrate that PBM‐YOLO achieves mean average precision at 50% intersection over union (mAP50) scores of 90.7% and 85.1% and mean average p averaged over intersection over union (IoU) thresholds from 0.5 to 0.95 (mAP50:95) scores of 48.1% and 50.4%, respectively. Compared to YOLOv8n, it reduces the number of parameters by 40% while maintaining a well‐balanced performance profile.