基于YOLOv3-Tiny改进的船舶目标检测研究  

IMPROVED SHIP TARGET DETECTION BASED ON YOLOV3-TINY

作　　者：朱伟 段先华 程婧怡 Zhu Wei;Duan Xianhua;Cheng Jingyi(School of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212100,Jiangsu,China)

机构地区：[1]江苏科技大学计算机学院,江苏镇江212100

出　　处：《计算机应用与软件》2024年第6期169-174,229,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61806087)。

摘　　要：针对船舶实时性检测中出现的检测精度低、漏检问题,改进一种基于YOLOv3-Tiny的船舶目标检测算法。通过引入深度可分离卷积作为主干网络,提高通道数量,减少模型的参数量和运算量;采用H-Swish和Leaky ReLU激活函数改进卷积结构,提取更多特征信息;利用GIOU(Generalized Intersection Over Union)损失优化边界框,突显目标区域重合度,提高精度。在混合船舶数据集上检测结果表明,改进后YOLOv3-Tiny的检测精度为83.40%,较原算法提高5.33百分点,召回率和检测速度也均优于原算法,适用于船舶实时性检测。Aiming at the problems of low detection accuracy and missed detection in real-time detection of ships,this paper improves a ship target detection algorithm based on YOLOV3-Tiny.By introducing depth wise separable convolution as the backbone network,the number of channels was increased and the number of parameters and computation of the model were reduced.The H-Swish and Leaky ReLU activation functions were used to improve the convolution structure in order to extract more feature information.GIOU loss was used to optimize the bounding box to highlight the coincidence degree of the target area and improve the accuracy.The detection results on the mixed ship data set show that the detection accuracy of the improved YOLOv3-Tiny is 83.40%,which is 5.33 percentage points higher than the original algorithm.Its recall rate and detection speed are also better than the original algorithm,which is suitable for real-time ship detection.

关 键 词：船舶检测 YOLOv3-Tiny 深度可分离卷积 H-Swish GIOU 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]