基于YOLOv9-B模型的港口船舶红外检测方法

Infrared ship detection in harbor based on YOLOv9-B model

为了进一步提高港口作业安全与智能化水平,本文针对海上复杂环境下的在港船舶智能检测方法进行研究。主要考虑客观环境引起的成像模糊及拍摄角度不同导致的船舶目标较小等因素造成的检测不准确问题,提出了一种基于YOLOv9-B的高精度红外船舶检测模型。首先,设计一种多尺度空间注意力机制,采用多个空洞卷积取代原本空间注意力中的普通卷积,扩大感受野捕获更多全局信息。然后,设计一种分支融合注意力机制,通过引入便捷通道注意力和多尺度空间注意力机制来增强小目标和模糊目标关注度,减少特征融合过程中的目标信息损失。最后,将YOLOv9中RepNCSPELAN4模块替换为C2f模块,加强特征提取能力,提高模型检测准确度。在红外船舶数据集和本文自建数据集进行消融实验,结果显示,相较于YOLOv9模型,本文模型在mAP上分别提升了1.6%和1.9%,检测速度分别提升了3.2和1.2 fps。同时,对比实验表明,相较于其他主流模型,本文模型更具优越性。

To further enhance the safety and itelligence of port operations,itelligent detection methods for ships in the harbor under complex conditions are studied in this paper.It primarily addresses sses of detection inaccuracies caused by imaging blur due to environmental factors and the small size of vessel targets resulting from shooting angles.A high-precision infrared vessel detection model based on Y0L0v9-B is proposed.Finstly,a multi-scale spatial attention module is designed,wherein traditional convolutions in spatial attention are replaced with multiple dilated convolutions to expand the receptive field and capture more local information.Secondly,a branch fusion attention mechanism is devised to enhance the focus on small and blurry targets by introducing efficient channel attention and multi-scale spatial attention,thereby reducing the loss of target information during feature fusion.Finally,the RepNCSPELAN4 module in Y0L0v9 is replaced with the C2f module to strengthen feature extraction capabilities and improve detection accuracy.The ablation experiments are conducted on the infrared ship dataset and the self-constructed dataset in this paper,and the results show that compared with the YOLOv9 model,the proposed model improves the mAP by 1.6%and 1.9%,and improves the detection speed by 3.2 and 1.2 fps,respectively.At the same time,the comparative experiments show that the proposed model is superior to other mainstream models.