关键词: YOLOv8, 深度学习, 目标检测, 损失函数

Abstract:  Currently， object detection has been found extensive applications across various domains and is progressively reaching a state of maturity. However， in the task of detecting riverine discarded bottles， the ongoing challenges with existing image processing technologies include low accuracy， high costs， and deployment difficulties， which continue to be significant obstacles in this research field. Therefore， we present an enhanced deep learning detection model based on YOLOv8. Firstly， to tackle the common issue of false positives and false negatives for small objects in unmanned vessel images， we introduce the Bi-PAN-FPN concept to improve the Neck component of YOLOv8-n. By carefully considering and reusing multi-scale features， this approach aims to achieve a more sophisticated and comprehensive feature fusion process while minimizing parameter costs. Secondly， the CIoU loss function is optimized by introducing the EIoU and addressing the imbalance between hard and easy samples. This is achieved by separating the influence of aspect ratio factors between predicted and ground-truth boxes. These enhancements aim to improve the optimization capability of the model. We conduct experiments using the publicly available FloW dataset and includes multiple evaluations， such as ablation experiments， comparative experiments， performance analysis of loss functions， and contrast experiments in special scenarios. These evaluations provide comprehensive evidence of the feasibility and effectiveness of the proposed method from multiple perspectives. The results of the ablation experiments indicate that the enhanced YOLOv8-n model exhibits a significant improvement in reducing false negatives compared to the baseline network， achieving an average precision of 85.2%. This marks a 2.7 percentage points increase over the baseline network model， demonstrating a notable enhancement in detection performance. The results of the comparative experiments indicate that the average precision of the improved model surpasses six other models， namely Mobilenet-SSDv2， YOLOv4s， YOLOv5s， YOLOv7-tiny， YOLOv3-SPP， and YOLOv5-MobileNetV3s， by 60.15%， 18.99%， 3.90%， 7.30%， 28.7%， and 55.47%， respectively. These findings highlight the superior overall performance of the improved model in terms of FPS， parameter quantity， and model size. Additionally， the improved model demonstrates exceptional performance in special scenarios. Therefore， the comprehensive performance of the improved model in the paper outperforms the currently popular models for river garbage detection. As a result， it is more suitable for real-time detection of discarded bottles in rivers.

Key words: YOLOv8, deep learning, target detection, loss function