摘要
针对红外图像船舶目标检测方法的准确性和实时性还不能满足海防场景需求的问题,提出一种基于改进YOLOv7算法的轻量级船舶检测算法。该算法首先在Backbone网络中引入MobileNetv3主干网络,实现模型轻量化处理。然后在Neck网络引入注意力机制抑制噪声与干扰,以提高网络的特征提取能力,并采用双向加权特征金字塔,以提升特征融合能力。最后引入Wise IoU优化损失函数,提高模型的收敛速度与精度。在艾睿数据集上的实验结果表明,相较于YOLOv7,所提改进算法的精确率、召回率、平均精度均值(mAP)@0.5与mAP@0.5∶0.95分别提升0.9个百分点、2.5个百分点、1.2个百分点和1.2个百分点,模型参数量降低了约38.4%,浮点运算数(FLOPs)降低了约65.5%。所提改进算法在满足检测速度要求的同时得到了更优的检测精度,有效地实现了高速、高精度的船舶检测。
To address the issue of inadequate accuracy and real-time performance of infrared ship target detection methods on coastal defense scenarios,a novel lightweight ship detection algorithm based on improved YOLOv7 framework is proposed.This framework incorporates several enhancements to augment its capabilities.First,to achieve model lightweight processing,the algorithm integrates the MobileNetv3 network into the architecture of the Backbone network.This addition contributes to efficient computation and model size reduction.Second,an attention mechanism is introduced within the Neck network to mitigate noise and interference,thereby improving the network’s feature extraction capability.In addition,we employ a bidirectional weighted feature pyramid to enhance feature fusion within the network,promoting more effective information integration.Finally,the algorithm incorporates Wise IoU to optimize the loss function,improving convergence speed and model accuracy.Experimental evaluations on the Arrow dataset demonstrate noteworthy improvements over the standard YOLOv7 approach.Specifically,the proposed enhanced algorithm exhibits a 0.9 percentage points increase in accuracy,2.5 percentage points increase in recall,and 1.2 percentage points increase in mean average precision(mAP)at IoU thresholds of 0.5 and 0.5∶0.95.In addition,it achieves approximately 38.4%reduction in model parameters and a 65.5%reduction in floating point operations per second(FLOPs).This enhanced algorithm delivers superior inspection accuracy while meeting the speed requirements for efficient ship inspection.Consequently,it effectively enables high-speed and high-precision ship detection.
作者
张燊
胡林
孙祥娥
刘美华
Zhang Shen;Hu Lin;Sun Xiang’e;Liu Meihua(School of Electronic Information,Yangtze University,Jingzhou 434023,Hubei,China;Intelligence Research Institute,Yangtze University,Jingzhou 434023,Hubei,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第22期248-254,共7页
Laser & Optoelectronics Progress
基金
国家自然科学基金(62273060)。
