基于通道剪枝的改进YOLOv7-tiny舰船识别算法  

An Improved YOLOv7-tiny Ship Recognition Algorithm Based on Channel Pruning

作　　者：张上 熊中越 王恒涛 ZHANG Shang;XIONG Zhongyue;WANG Hengtao(College of Computer and Information Technology,China Three Gorges University,Yichang 443000,China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443000

出　　处：《电光与控制》2025年第4期31-36,共6页Electronics Optics & Control

基　　金：国家级大学生创新创业训练计划(202011075013,202111075012)。

摘　　要：海上舰船目标识别是海洋监测的重要一环,也是世界各海岸地带国家国土安全的重要解决方案之一。针对SAR图像舰船目标检测存在识别精度低、训练模型大等问题,提出了一种基于通道剪枝的改进YOLOv7-tiny海上舰船识别算法。首先,采用MobileNetV3替代原有主干网络,以降低模型的计算量和体积,实现模型轻量化;其次,引入MPDIoU简化计算过程,优化模型的收敛性;最后,通过通道剪枝提高模型精度,同时平衡模型体积和计算量的降低幅度,进一步优化算法模型。实验结果表明,改进算法相对于YOLOv7-tiny,召回率提升了5.85个百分点,mAP提升了3.69个百分点,参数量减少了63.35%,计算量减少了70%。Ship target identification at sea is a crucial part of maritime monitoring and a significant solution for national security in coastal regions worldwide.Aiming at the problems of low recognition accuracy and large training model in ship target detection in SAR images,an improved YOLOv7-tiny maritime ship recognition algorithm based on channel pruning is proposed.Firstly,the original backbone network is replaced by MobileNetV3 to reduce the calculation and volume of the model and realize the lightweight of the model.Secondly,MPDIoU is introduced to simplify the calculation process and optimize the convergence of the model.Finally,through channel pruning,the model accuracy is improved,while the reduction of model volume and calculation amount is balanced,and the network model is further optimized.The experimental results show that compared with YOLOv7-tiny,the improved algorithm improves the recall and mAP by 5.85 and 3.69 percentage points respectively,the parameter is reduced by 63.35%,and the FLOPs is reduced by 70%.

关 键 词：目标检测 YOLOv7-tiny SAR图像 轻量化模型 通道剪枝 损失函数 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]