Face Attributes and Detection of Drug Addicts

Abstract

This paper addresses the problem of detecting the prolonged drug abuse marks using “soft” face biometrics. We propose a two-stage approach that integrates both the machine learning approach and the probabilistic reasoning. We identified “drug affect facial attributes” for attribute classification, and then performed probabilistic inference for detecting the drug addicts using the five face attributes. The experiments were performed using pre-trained convolutional neural networks, GoogleNet, ResNet50 and VGG16, for face attribute classification. PCA was applied on the extracted features for dimensionality reduction along with Fisher's linear discriminant feature selection method, and then a classification of attributes was performed using a linear SVM. The average accuracy of the five “drug affect facial attribute” detection reached, in particular, 90% using the ResNet50 model. We then applied this statistics to create a Bayesian network which represented the causal model for the final decision-making to classify the subjects as drug addicts. This approach reached the accuracy of 84%.