基于改进Mask R-CNN的无人船视觉检测  

Visual Detection of Unmanned Boat Based on Improved Mask R-CNN

作　　者：佟剑峰 于雨 TONG Jian-feng;YU Yu(Institute of Oceanographic Instrumentation,Qilu University of Technology(Shandong Academy of Sciences),Qingdao Shandong 266061,China;Shandong Key Laboratory of Ocean Environmental Monitoring Technology,Qingdao Shandong 266061,China;National Engineering and Technological Research Center of Marine Monitoring Equipment,Qingdao Shandong 266061,China)

机构地区：[1]齐鲁工业大学(山东省科学院)山东省科学院海洋仪器仪表研究所,山东青岛266061 [2]山东省海洋监测仪器装备技术重点实验室,山东青岛266061 [3]国家海洋监测设备工程技术研究中心,山东青岛266061

出　　处：《船海工程》2025年第1期6-12,共7页Ship & Ocean Engineering

基　　金：国家自然科学基金(41706101)。

摘　　要：针对无人船视觉检测算法准确率低及水界线检测困难的问题,构造自定义数据集并提出改进Mask R-CNN网络模型的无人船视觉检测算法。改进后的算法以Mask R-CNN网络模型为检测框架,将骨干网络由ResNet50替换成ResNeXt50并加入SENet注意力机制模块提高模型的特征提取能力;在特征金字塔网络(feature pyramid network,FPN)中加入多尺度的特征提取模块(inception模块),提高特征图的融合效果;加入多尺度锚框(anchors),提高模型对于多尺度目标的检测效果;通过直方图均衡化、调整对比度的方式对输入图像进行预处理,优化输入图像。结果表明,改进后的Mask R-CNN网络模型相比于原始的网络模型在目标检测任务中平均精度均值(mean average precision,mAP)提高了8.86%,交并比为0.5条件下的平均精度(IOU=0.5 average precision,AP 50)提高了9.39%;在实例分割任务中mAP提高了4.55%,AP 50提高了4.63%。相关改进,提高了无人船视觉检测的效率。An improved Mask R-CNN network model was proposed to address the problems of low accuracy in the detection algorithm for unmanned boat vision and the difficulty of detecting the water boundary.The improved algorithm took the Mask R-CNN network model as the detection framework,replaced the backbone from ResNet50 to ResNeXt50,and added the SENet attention mechanism to improve the feature extraction ability of the model.It also added the inception module to the FPN structure.Additionally,it added multiscale anchors to improve the detection effect of the model for multiscale targets.The input image was preprocessed by histogram equalization and increasing contrast.The experimental results showed that the improved model can improve mAP by 8.86%and AP50 by 9.39%in the object detection task,and improves mAP by 4.55%and AP50 by 4.63%in the instance segmentation task,so as to increase efficiency of unmanned boat visual detection.

关 键 词：无人船视觉 Mask R-CNN网络模型 骨干网络 注意力机制 特征融合 

分 类 号：U675.7[交通运输工程—船舶及航道工程]