基于迁移学习的水产动物图像识别方法

针对传统水产动物图像识别方法步骤复杂、准确率差、泛化性差，而深度卷积神经网络（DCNN）模型开发难度大等问题，提出一种基于参数迁移策略采用微调方式再训练源模型的方法。首先，对图像进行数据增强等预处理；然后，在修改源模型全连接分类层的基础上，进一步将高层卷积模块的权重设置为可训练以进行自适应调整；最后，以验证集识别率与训练时间作为评估指标，针对不同源模型采用不同的可训练参数占比进行性能实验。实验结果表明，通过再训练得到的图像识别模型准确率可达到97.4%，相比源模型最多可提高20个百分点；在可训练参数占比为75%左右时可得到较理想的性能。通过实验证实了采用微调方法可以在低成本开发条件下得到性能良好的深度神经网络图像识别模型。

Aquatic animal image classification method based on transfer learning

Aiming at the problems that traditional aquatic animal image recognition methods have complex steps, low accuracy and poor generalization, and it is difficult to develop Deep Convolutional Neural Network (DCNN) model, a method based on parameter transfer strategy using fine-tune to retrain pre-trained model was proposed. Firstly, the image was preprocessed by data enhancement and so on. Secondly, on the basis of modifying the source model's fully connected classification layer, the weights of high-level convolution modules were set to be trained for adaptive adjustment. Finally, using training time and recognition accuracy on validation set as the evaluation indexes, the performance experiments were conducted on various network structures and different proportion of trainable parameters. The experimental results show that the highest retrained model classification accuracy can reach 97.4%, 20 percentage points higher than the source model, the ideal performance can be obtained when the proportion of trainable parameters is around 75%. It is proved that the fine-tune method can obtain a deep neural network image classification model with good performance under low-cost development condition.