基于迁移学习的非约束环境下热带鱼类识别

Tropical Fish Identification Under Unconstrained Environment Based on Transfer Learning

针对海底非约束环境下视频背景多模、环境变化复杂导致图像识别困难的问题,提出基于迁移学习的非约束环境下热带海域鱼类识别方法。首先构建ResNet深度卷积神经网络;其次引入迁移学习进行网络训练,对比迁移学习前后的识别效果。结果表明,在引入迁移学习下,accuracy和loss指标均优于非迁移学习的情况,并且在训练到150个epoch时,各项指标开始收敛,能够较好地完成非约束环境下热带海域的鱼类识别任务。

Aiming at the problems that seabed unconstrained environment video often with complex multi-mode background, complex weather variation, which leads to difficulty in image recognition, we proposed a tropical fish identification method under unconstrained environment based on transfer learning. Firstly, we constructed the ResNet deep convolution neural network. And then, we introduced transfer learning for network training, and compared the recognition effect before and after transfer learning. The results show that the indexes of accuracy and loss with transfer learning are better than those without transfer learning. When the training reaches 150 epochs, each index begins to converge, which can better complete the task of tropical fish identification under unconstrained environment.