| 摘要: |
| 为实现对养殖鳗鲡(Anguilla)摄食强度的准确评估, 提出了一种基于双流残差卷积神经网络的鳗鲡摄食强度评估方法, 该方法针对传统双流网络(Two-stream)中存在的问题做出了相应的改进。首先针对传统双流网络存在网络结构较浅, 无法提取到充分的鳗鲡摄食行为特征的问题, 选择使用ResNet50网络进行替换, 以提取到更具代表性的特征。其次针对传统双流网络最后的分类结果是把空间流和时间流的得分取平均值融合而获得, 这种方式较为简单, 且其空间流和时间流网络为独立进行训练, 容易导致网络出现学习不到鳗鲡摄食行为的时空关联特征的问题, 选择使用特征层融合方式对空间流和时间流网络提取获得的特征进行融合, 让网络能够并行进行训练, 以提取到时空信息间的关联特征。试验结果表明: 文内提出的基于双流残差卷积神经网络的鳗鲡摄食强度评估方法准确率达到98.6%, 与单通道的空间流和时间流网络相比, 准确率分别提升了5.8%和8.5%, 与传统的双流网络相比准确率也提升了3.2%。 |
| 关键词: 鳗鲡 摄食强度 双流残差卷积神经网络 ResNet50 并行训练 特征层融合 |
| DOI：10.11693/hyhz20221100291 |
| 分类号:Q959.9; S965 |
| 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; 福建省科技厅高校产学合作项目, 2020N5009号。 |
附件 |
|
| EVALUATION ON FEEDING INTENSITY OF AQUACULTURE EEL (ANGUILLA) BY DOUBLE-FLOW RESIDUAL CONVOLUTION NEURAL NETWORK |
|
LI Kai1,2, JIANG Xing-Long1,2, XU Zhi-Yang1,2, LIN Qian1,2
|
|
1.Fisheries College, Jimei University, Xiamen 361021, China;2.Engineering Research Center of the Modern Technology for Eel Industry, Ministry of Education, Xiamen, 361021, China
|
| Abstract: |
| To accurately evaluate the feeding intensity in eel (Anguilla) culture, the eel intensity evaluation method based on double-flow residual convolution neural network was proposed, by which the problems existing in traditional double-flow network (Two-stream) was solved. The traditional two-flow network is shallow in network structure and not able to extract sufficient eel feeding behavior information. Therefore, ResNet50 network was chosen to extract more representative features. The final classification score of the traditional double-flow network could be obtained by combining the average scores of spatial flow and time flow, and the fusion method was relatively simple, and the spatial flow and time flow network were trained independently, which could lead to an issue that the network cannot learn the spatio-temporal correlation characteristics of eel feeding behavior. We chose to use the feature layer fusion method to fuse the features extracted from the spatial flow and time flow network, by which the network was trained in parallel to extract the correlation features of the spatio-temporal network. Results show that the classification accuracy of the eel feeding intensity evaluation in double-flow residual convolution neural network reached 98.6%, which was 5.8% and 8.5% higher than that of single-channel spatial flow and time flow network, respectively. Compared with the traditional double-flow network, the classification accuracy was improved by 3.2%. |
| Key words: eel feeding intensity double-flow residual convolution neural network ResNet50 parallel training feature layer fusion |
