A Method for Inland Vessel Object Detection Based on PEW-YOLOv8

In inland vessel object detection, many targets fall into the category of small objects, occupying limited pixels in images. Additionally, interference from complex environments often leads to insufficient detection accuracy, frequent false positives, and missed detections. To address these challenges, this study proposes an object detection algorithm based on PEW-YOLOv8, which integrates YOLOv8 with a P2 detection layer, EfficientNetV2, and the WIoUinner loss function. A new P2 shallow detection layer with a resolution of 160×160 is introduced to enhance small target detection. A 32-dimensional feature space reconstruction is employed to achieve dynamic weight allocation across multi-scale features. Furthermore, a bidirectional interaction mechanism between high- and low-level features is designed to improve feature extraction for small vessel objects. To address the increased parameter burden caused by multi-level detection heads, an improved EfficientNetV2 architecture is adopted, which incor-porates a GELU-activated Stem module to mitigate gradient explosion and unstable training. During training, the channel count is expanded fourfold while simplifying the convolutional structure, significantly accelerating the training process without sacrificing model quality. Besides, the WIoUinner loss function with a dynamic non-monotonic focusing mechanism is designed, which introduces auxiliary prediction boxes with varying scales to accelerate the convergence of bounding boxes. When the predicted and ground truth boxes are closed aligned, the model places greater emphasis on the distance between center points, reducing the penalty from geometric metrics and improving generalization capability. The algorithm is validated using a dataset that combines the publicly available Seaships dataset with a self-constructed inland vessel dataset. Experimental results demonstrate that compared to YOLOv10, PEW-YOLOv8 achieves an average detection accuracy of 94.8%, a 3% improvement. Computational complexity is significantly reduced, with FLOPs optimized to 3.7 G, representing a 43.1% reduction, which demonstrates the model's advantages in both accuracy and efficiency for inland vessel detection tasks. Heatmap analysis further confirms the model's ability to effectively focus on inland vessel features, demonstrating robust detection performance in complex inland waterway scenarios.