考虑类间差异损失的中华白海豚个体识别
Identification of Sousa Chinensis Considering the Inter-Class Loss
-
摘要: 为了提升中华白海豚个体识别的准确率, 提出了基于类间差异损失的细粒度识别模型. 首先, 基于不同类别图像的特征之间具有差异性的事实, 设计了类间差异损失函数; 其次, 在VGG16的基础上根据类别数调整了部分卷积层的通道数, 并设计了两种全连接方式用于识别不同规模的个体; 最后, 将所提损失函数与已有损失进行组合, 激励网络学习到更具区分度的特征. 选取在厦门湾拍摄的2177幅中华白海豚图像, 人工标注为30头个体. 在该数据集上进行实验, 所提损失函数可以将准确率提升1.05%, 达到98.65%, 且比主流的细粒度识别算法至少高出0.9%.Abstract: To improve the accuracy of individual identification of Sousa Chinensis, a fine-grained identification model based on the loss of inter-class difference is proposed. Firstly, based on the fact that the features of different types of images are different, a loss function of inter-class difference is designed; Secondly, the channel number of several convolution layers about VGG16 is adjusted according to the number of categories. And two full con-nection modes are designed for identification of large-scale and small-scale individuals respectively; Finally, the proposed loss and the existing loss are combined to stimulate the network to learn more distinguishing features. 2177 images of Sousa Chinensis taken in Xiamen Bay are selected and manually labeled as 30 individuals. Ex-periments on this dataset show that the proposed loss function can improve the accuracy rate by 1.05% to 98.65%, which is at least 0.9% higher than the mainstream fine-grained identification algorithms.