Chen Hu, Sun Che, Liao Peixi, Lai Yancun, Fan Fei, Lin Yi, Deng Zhenhua, Zhang Yi
College of Computer Science, Sichuan University, Chengdu, Sichuan, China.
Department of Scientific Research and Education, The Sixth People's Hospital of Chengdu, Chengdu, Sichuan, China.
Patterns (N Y). 2022 Apr 1;3(5):100485. doi: 10.1016/j.patter.2022.100485. eCollection 2022 May 13.
When accidents occur, panoramic dental images play a significant role in identifying unknown bodies. In recent years, deep neural networks have been applied to address this task. However, while tooth contours are significant in classical methods, few studies using deep learning methods devise an architecture specifically to introduce tooth contours into their models. Since fine-grained image identification aims to distinguish subordinate categories by specific parts, we devise a fine-grained human identification model that leverages the distribution of tooth masks to distinguish different individuals with local and subtle differences in their teeth. First, a bilateral branched architecture is designed, of which one branch was designed as the image feature extractor, while the other was the mask feature extractor. In this step, the mask feature interacts with the extracted image feature to perform elementwise reweighting. Additionally, an improved attention mechanism was used to make our model concentrate more on informative positions. Furthermore, we improved the ArcFace loss by adding a learnable parameter to increase the loss of those hard samples, thereby exploiting the potential of our loss function. Our model was tested on a large dataset consisting of 23,715 panoramic X-ray dental images with tooth masks from 10,113 patients, achieving an average rank-1 accuracy of 88.62% and rank-10 accuracy of 96.16%.
事故发生时，全景牙科图像在识别无名尸体方面发挥着重要作用。近年来，深度神经网络已被应用于解决这一任务。然而，虽然牙齿轮廓在传统方法中很重要，但很少有使用深度学习方法的研究设计一种专门将牙齿轮廓引入其模型的架构。由于细粒度图像识别旨在通过特定部分区分从属类别，我们设计了一种细粒度人体识别模型，该模型利用牙齿掩码的分布来区分牙齿存在局部细微差异的不同个体。首先，设计了一种双边分支架构，其中一个分支被设计为图像特征提取器，另一个是掩码特征提取器。在这一步中，掩码特征与提取的图像特征相互作用以进行逐元素重新加权。此外，使用了一种改进的注意力机制，使我们的模型更多地关注信息丰富的位置。此外，我们通过添加一个可学习参数来改进ArcFace损失，以增加那些难样本的损失，从而挖掘我们损失函数的潜力。我们的模型在一个由来自10113名患者的23715张带有牙齿掩码的全景X射线牙科图像组成的大型数据集上进行了测试，平均排名第一的准确率为88.62%，排名前十的准确率为96.16%。
