基于改进YOLOx-nano的海上红外目标检测算法  

Marine Infrared Target Detection Algorithm Based on Improved YOLOx-nano

作　　者：张俊 位门 吕璐 ZHANG Jun;WEI Men;LV Lu(The 58th Research Institute of CETC,WuXi 430070,China)

机构地区：[1]中国电子科技集团公司第五十八研究所,江苏无锡430070

出　　处：《红外》2025年第2期49-56,共8页Infrared

摘　　要：提出了一种基于改进YOLOx-nano的海上红外目标检测算法。通过对检测头的分类与定位任务进行解耦,并引入改进的特征金字塔网络(Feature Pyramid Network, FPN)结构,不仅提升了模型的精度和收敛速度,而且提高了红外大目标检测能力。将改进的压缩和激励网络(Squeeze-and-Excitation Network, SENet)通道注意力机制模块加入到模型中,增强了模型的非线性表达能力,同时提高了有效特征学习能力。为了加快嵌入式平台模型的前向推理速度,引入剪枝技术来实现模型剪枝,在保证召回率不降低的情况下减少模型参数。通过测试集对本文算法进行了验证。结果表明,该算法的平均精度(Average Precision, AP)比原始YOLOx-nano算法提高了1.35%,达到了93.92%。本文算法平衡了模型精度与耗时的矛盾关系,在提升性能的同时,保证了模型检测的速度。A marine infrared target detection algorithm based on improved YOLOx-nano is proposed.By decoupling the classification and positioning tasks of the detection head and introducing an improved feature pyramid network(FPN)structure,not only the accuracy and convergence speed of the model are improved,but also the infrared large target detection capability is improved.The improved squeeze-and-excitation network(SENet)channel attention mechanism module is added to the model to enhance the nonlinear expression ability of the model and improve the effective feature learning ability.In order to speed up the forward reasoning speed of the embedded platform model,the pruning technology is introduced to implement model pruning,and the model parameters are reduced without reducing the recall rate.The algorithm in this paper is verified by the test set.The results show that the average precision(AP)of the algorithm is 1.35%higher than that of the original YOLOx-nano algorithm,reaching 93.92%.The algorithm in this paper balances the contradictory relationship between model accuracy and time consumption,and ensures the speed of model detection while improving performance.

关 键 词：红外目标检测 YOLOx-nano 模型剪枝 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]