| 基于深度学习的循环水养殖鳗鲡(Anguilla)计数研究 |
| |
| 引用本文： | 李凯,江兴龙,陈尔康,陈彭,许志扬,林茜.基于深度学习的循环水养殖鳗鲡(Anguilla)计数研究[J].海洋与湖沼,2022,53(3):664-674. |
| |
| 作者姓名： | 李凯 江兴龙 陈尔康 陈彭 许志扬 林茜 |
| |
| 作者单位： | 集美大学水产学院 福建厦门 361021;鳗鲡现代产业技术教育部工程研究中心 福建厦门 361021;鳗鲡现代产业技术教育部工程研究中心 福建厦门 361021;集美大学海洋与信息工程学院 福建厦门 361021 |
| |
| 基金项目： | 国家重点研发计划“特色鱼类精准高效养殖关键技术集成与示范”,2020YFD0900102号。 |
| |
| 摘 要： | 鳗鲡(Anguilla)作为我国优质水产养殖种类,精准掌握其数量对高效养殖有重要意义。为实现对循环水养殖鳗鲡的准确计数,提出了一种基于深度学习的改进Faster RCNN模型。针对检测目标即鳗鲡头部尺寸小的问题,选择在特征提取网络ResNet50中加入FPN结构来作为模型的骨干网络,以提取并融合多尺度的特征;针对原模型锚框都是基于人工经验设置的,并不适用于鳗鲡数据集的问题,使用k-means聚类算法对训练集中标注的鳗鲡头部检测框进行聚类分析,获得了适合鳗鲡数据集的15种不同尺度的锚框;针对图像中存在鳗鲡头部重叠的问题,选择使用Soft-NMS算法替代原NMS算法对RPN部分生成的候选框进行筛选,以减少模型对鳗鲡重叠部分的漏检情况。试验结果表明:改进后的Faster RCNN模型对鳗鲡头部的检测精度(mAP0.5)高达96.5%,较原Faster RCNN模型(Backbone为ResNet50)显著提升了14%,与SSD300和YOLOV3模型相比分别显著提升了24.9%和15%;在鳗鲡计数上,利用改进后的Faster RCNN模型检测结果进行计数,计数准确率达到90%以上,提升了模型对鳗鲡的检测识别能力。
|
| 关 键 词： | 鳗鲡计数 深度学习 Faster RCNN模型 FPN结构 k-means聚类算法 Soft-NMS算法 |
| 收稿时间： | 2021/12/19 0:00:00 |
| 修稿时间： | 2022/1/26 0:00:00 |
AUTO-COUNTING THE EEL ANGUILLA IN RECIRCULATING AQUACULTURE SYSTEM VIA DEEP LEARNING |
| |
| LI Kai,JIANG Xing-Long,CHEN Er-Kang,CHEN Peng,XU Zhi-Yang,LIN Qian.AUTO-COUNTING THE EEL ANGUILLA IN RECIRCULATING AQUACULTURE SYSTEM VIA DEEP LEARNING[J].Oceanologia Et Limnologia Sinica,2022,53(3):664-674. |
| |
| Authors: | LI Kai JIANG Xing-Long CHEN Er-Kang CHEN Peng XU Zhi-Yang LIN Qian |
| |
| Institution: | Fisheries College, Jimei University, Xiamen 361021, China;Engineering Research Center of the Modern Technology for Eel Industry, Ministry of Education, Xiamen 361021, China;Engineering Research Center of the Modern Technology for Eel Industry, Ministry of Education, Xiamen 361021, China;School of Ocean Information Engineering, Jimei University, Xiamen 361021, China |
| |
| Abstract: | Eel is an economically valuable species in China. To realize accurate and efficient auto-counting of eel cultured in recirculating aquaculture system, an improved Faster RCNN model by deep learning was proposed. Because the head size of eel is small, how to recognize eel individual correctly became an issue of this study. To solve the issue, the FPN structure was added to the feature extraction network ResNet50 as the backbone network of model to extract and fuse multi-scale features. However, the original model anchors are based on artificial experience, which is not suitable for eel dataset process. Therefore, the eel-head detection boxes marked in the training dataset were clustered and analyzed in k-means clustering algorithm, and 15 different scales of anchors suitable for the eel dataset were obtained. In addition, to avoid eel head overlap in the images, Soft-NMS algorithm instead of the original NMS algorithm was applied to select candidate boxes generated by RPN to reduce the missed detection of eel due to the image overlap. Results show that the detection accuracy of the improved Faster RCNN model for eel-head counting reached 96.5%, which is 14%, 24.9%, and 15% higher than that of the original Faster RCNN (the backbone of ResNet50), SSD300, and YOLOV3 models, respectively. With the improved Faster RCNN model, the correct counting accuracy could reach >90%, by which the eel detection and identification in recirculating aquaculture system could be enhanced considerably. |
| |
| Keywords: | eel counting deep learning Faster RCNN FPN k-means Soft-NMS |
|
| 点击此处可从《海洋与湖沼》浏览原始摘要信息 |
| 点击此处可从《海洋与湖沼》下载免费的PDF全文 |
|
