IR-YOLOv8 for lightweight object detection in inland waterway environments

Abstract

Rapid and accurate object perception is essential for inland-waterway navigation, yet remains challenging because of complex visual environments, the lack of task-specific datasets, and limited computational resources. To address these challenges, we construct ODDIIRS, an inland-waterway object detection dataset containing 10,198 images across 12 categories, including five vessel types, four categories of floating debris, and three categories of navigation facilities or landmarks. Based on ODDIIRS, we propose IR-YOLOv8, a lightweight detector designed to improve small-object perception, feature preservation, prediction efficiency, and localization quality. The model incorporates a high-resolution P2 prediction branch to better represent small and distant targets. ADown reduces the loss of fine-grained information during multi-level downsampling. To offset the additional cost introduced by four-scale prediction, LSCD_Head employs shared convolution and Group Normalization to reduce the parameter count and computational cost across the P2-P5 feature levels. During training, PIoU v2 is used to improve bounding-box regression and accelerate early optimization without introducing additional operators during inference. Experiments on ODDIIRS show that, compared with YOLOv8n, IR-YOLOv8 reduces the parameter count and model size by 35.88% and 31.89%, respectively. Precision, recall, F1-score, mAP@0.5, and mAP@0.5:0.95 improve by 0.58, 3.37, 2.00, 1.92, and 1.25 percentage points, reaching 93.23%, 92.60%, 92.91%, 94.89%, and 65.68%, respectively. With an input resolution of \(640\times 640\) , the model achieves an inference-only throughput of 356.24 FPS on an NVIDIA RTX 3060 GPU. These results indicate that IR-YOLOv8 provides a good balance between detection accuracy and model compactness.