摘要
目的:为了提高水下复杂环境下检测海参和海星等水下目标物的鲁棒性。方法:提出了基于改进Mask R-CNN的海参和海星实例分割算法,该算法以Mask R-CNN结构为主框架,将Swin Transformer主干网络代替Mask R-CNN原本的ResNet卷积神经网络;同时采用Water-Net网络对海参和海星实例数据集进行图像增强;最后采用Soft-NMS的方法替换经典的NMS算法。结果:在本文自己标定的数据集上进行实验,与改进前Mask R-CNN相比,本文算法检测框检测mAP可达到70.6%,提升6.4%;实例分割mAP达到了69.2%,提高了4.7%,并且正确率收敛于97%。结论:与其他主流目标检测算法相比较,本文提出的方法具有更高的检测精度,在水下目标检测任务上更加具有优势。
Aims:This paper aims to improve the robustness of detecting sea cucumber and starfish in the complex underwater environment.Methods:An improved instance segmentation algorithm based on the Mask R-CNN was proposed.In the algorithm,the ResNet convolutional neural network of Mask R-CNN was substituted with the Swin-Transformer backbone network.At the same time,the Water-Net network was applied for image enhancement of the sea cucumber and starfish data.Finally,the NMS algorithm was replaced with the Soft-NMS method.Results:Compared with the classic Mask R-CNN network,the bounding box mAP of the presented algorithm was 70.6%with 6.4%of increase.The instance segmentation mAP was 69.2%with 4.7%of increase;and the accuracy rate was up to 97%.Conclusions:The proposed method has higher detection accuracy and obvious advantage in underwater target detection.
作者
胡栩榛
严天宏
HU Xuzhen;YAN Tianhong(College of Mechanical and Electrical Engineering,China Jiliang University,Hangzhou 310018,China)
出处
《中国计量大学学报》
2023年第1期34-43,50,共11页
Journal of China University of Metrology
基金
国家重点研发计划项目(No.2019YFC1408304)
浙江省自然科学基金项目(No.LTGG23E090002)。
