Faster-PGYOLO: an efficient framework for floating debris detection in inland waters

Unmanned cleaning vessels are increasingly deployed to detect and clean floating debris in inland waters. However, existing deep learning methods, though effective, often require excessive parameters and computational resources, limiting their practicality on edge devices like unmanned cleaning vessels with limited computational capacity. To address this, we propose a novel hardware-friendly approach utilizing YOLOv8s, specifically designed for detecting floating debris with unmanned cleaning vessels. Firstly, by integrating a high-resolution detection head and pruning redundant network layers, we construct a reshaped network structure to improve overall detection performance. Then, we optimize the C2f module by incorporating FasterNet's backbone, which curtails non-essential computation tasks. Furthermore, we embed the Fusion Global Coordinate Attention (FGCA) module in the path aggregation network (PAN) pathway and the parameter-free average attention module (PfAAM) in the feature pyramid network (FPN) pathway, resulting in a pioneering neck fusion network, GPAN + PFPN, capable of handling complex features in diverse adverse scenes with minimal computational burden. Numerous experiments demonstrate our method achieves a detection accuracy of 93.9% with only 2.25 million parameters and a 4.9 MB model size, making it hardware-friendly, and suitable for widespread use in edge devices. Our code and datasets are provided.