Image Recognition in New and Emerging Drugs (NEDs) Package with Convolutional Neural Network

 

Chih-Ping Yen

Department of Information Management, Central Police University,

Taoyuan 33304, Taiwan, ROC

ORCID: 0000-0002-1189-4922, peter@mail.cpu.edu.tw

 

Abstract

Today, many new and emerging drugs (NEDs) are packaged in a "foodized" way by mixing ingredients such as coffee, candy powder, jelly, etc. In an attempt to attract young people to take and avoid police investigations, which leads to the spread and abuse of drugs. Because people can't recognize the contents of the package as drugs, they are almost exposed to drugs without defense, and naïve teenagers are the most direct victims. In order to quickly identify whether it is an NED package that has been seized in the past by first-line staff. Based on deep learning, this study will propose a Multi-channel convolutional neural network (MCCNN) architecture to check the suspected packaging by taking photos with smart phones. We will use very few NEDs packaging images and applied 7 kinds of data augmentation methods, including brightness, scale, translation, shearing transformation, blur, rotation, and random crop to expand the training image data. Next, the proposed MCCNN method compares Uniform LBP (i.e. non-CNN) and other well-known CNN classification methods include AlexNet, VGG-16, VGG-19, GoogleNet, and ResNet. Finally, experiments have proven that the proposed MCCNN has the best accuracy compared with the non-CNN and state-of-the-art CNN methods, and reaches 98.56%.

 

Keywords: New and emerging drugs (NEDs), Deep learning, Multi-channel convolutional neural network (MCCNN), Data augmentation