摘要
针对海上雾天获取到的图像中小目标船舶识别效果低下、漏检率高等问题,提出了一种融合MSRCP算法的改进YOLOv5模型。在输入端加入MSRCP算法对图像进行预处理,提高远处船舶的特征;采用改进k-means聚类方法设计先验框,加快模型收敛速度,使锚框和边界框更匹配;在网络部分采用了SoftPool池化替换原来的MaxPool池化,保留更多的图像特征,提高图像的检测精度。经实验,改进后的算法MAP值提高了12%,平均召回率提升了16%,检测速度达到40帧/秒,能够在满足实时性检测的前提下,更好地完成对大雾天气下的船舶识别。
Aiming at the problems of low recognition effect and high missed detection rate of small target ships in the images obtained in fog on the sea,this paper proposes an improved YOLOv5 model integrating the MSRCP algorithm.The MSRCP algorithm is added to the input to preprocess the image to improve the characteristics of distant ships.The improved k-means clustering method is used to design a priori box,which accelerates the convergence speed of the model and makes the anchor box and the boundary box match better.In the network part,Softpool pooling is used to replace the original Maxpool pooling,retain more image features and improve the detection accuracy of images.The map value of the improved algorithm is increased by 12%,the average recall rate is increased by 16%,and the detection speed is up to 40 frames/second.On the premise of meeting the real-time detection effect,it can better complete the ship recognition in foggy weather.
作者
李伟
张雪
单雄飞
宁君
LI Wei;ZHANG Xue;SHAN Xiong-fei;NING Jun(College of Navigation,Dalian Maritime University,Dalian Liaoning 116026,China)
出处
《计算机仿真》
2024年第5期204-208,482,共6页
Computer Simulation
关键词
目标检测
船舶识别
雾天船舶检测
Target detection
Ship recognition
Ship detection in fog
