基于深度学习与泰勒展开的吸毒成瘾者识别

Identification of Drug Addicts by Deep Learning and Taylor Expansion

研究旨在利用深度学习方法通过面部图像以及抽象特征中的局部信息对吸毒成瘾者的成瘾程度和社区矫正时间进行识别。提出了一个基于泰勒展开式的神经网络模型,以深度残差网络作为主干网络,并嵌入泰勒特征图,使模型的训练时间减少、特征提取更加准确从而达到实时性的目标。实验过程中先对ResNet18进行预训练,再对嵌入的泰勒特征模式进行微调,网络末端通过全连接层与Softmax函数的组合进行分类,随机梯度下降的优化目标采用了交叉熵损失。此方法对于吸毒成瘾程度的识别准确度达到80.35%,对于社区矫正时间的识别准确率达到62.11%,该模型性能得到有效提升。

The research aims to use deep learning methods to identify drug addicts through facial images as well as local information in abstract features. The results included their degree of addiction and time to community correction. We propose a neural network model based on Taylor expansion, which uses a deep residual network as the backbone network and embeds the Taylor feature map to reduce the training time of the model and make feature extraction more accurate and fast to achieve the goal of real-time performance. During the experiment, ResNet18 was pre-trained first, and then the embedded Taylor feature map was fine-tuned. The end of the network was classified by the combination of the fully connected layer and the Softmax function. The optimization objective of stochastic gradient descent used cross-entropy loss. The recognition accuracy of this method for drug addiction level reaches 80.35%,and the recognition accuracy for community correction time reaches 62.11%,the model performance is improved effectively.