基于FTVGG16卷积神经网络的鱼类识别方法

Fish Identification Method Based on FTVGG16 Convolutional Neural Network

针对大多数应用场景中,大多数鱼类呈不规则条状,鱼类目标小,受他物遮挡和光线干扰,且一些基于颜色、形状、纹理特征的传统鱼类识别方法在提取图像特征方面存在计算复杂、特征提取具有盲目和不确定性,最终导致识别准确率低、分类效果差等问题,本文在分析已有的VGG16卷积神经网络良好的图像特征提取器的基础上,使用Image Net大规模数据集上预训练的VGG16权重作为新模型的初始化权重,通过增加批规范层(Batch normalization,BN)、池化层、Dropout层、全连接层(Fully connected,FC)、softmax层,采用带有约束的正则权重项作为模型的损失函数,并使用Adam优化算法对模型的参数进行更新,汲取深度学习中迁移学习理论,构建了FTVGG16卷积神经网络(Fine-tuning VGG16 convolutional neural network,FTVGG16)。测试结果表明:FTVGG16模型在很大程度上能够克服训练的过拟合,收敛速度明显加快,训练时间明显减少,针对鱼类目标很小、背景干扰很强的图像,FTVGG16模型平均准确率为97. 66%,对部分鱼的平均识别准确率达到了99. 43%。

Computer vision technology is widely applied in fish individual identification. Nevertheless, there are some problems such as small fish targets, occlusion of objects and light interference in videos and images. Some fish identification methods based on color, shape and texture also exit complicated calculations in feature extraction, such as non-migration of features will result in low recognition accuracy and poor classification. With the help of analysis of image feature extraction of the existing VGG16 convolutional neural network model, the FTVGG16 convolutional neural network (Fine-tuning VGG16 convolutional neural network) was designed. As it was known, the basic deep learning tool used in this work was convolutional neural networks. The FTVGG16 convolutional neural network was composed of convolutional layers, batch normalization layers, pooling layers, Dropout layers, fully connected layers and softmax layers. The experimental results showed that the average recognition accuracy of the FTVGG16 model for fish was about 97.66%, and the average recognition rate of some fishes could reach 99.43%. It had high recognition accuracy and robustness in pictures with small fish targets and strong background interference. It could be operated through an appropriate, easy-to-use, and user-friendly web application for the specific case of fish identification.