Enhancing Medicine Authenticity in Nigeria Using Deep Features Concatenation of VGG16 and ResNet-50

Abstract

This study tackles the public health risk of fake medicines in Nigeria's pharmaceutical sector, as several studies revealed a significant percentage of medicines used by Nigerians are fake. To authenticate and further classify the brands of the medicine, a self-captured dataset of medicines packages were used with a focus on features like packaging design (logo), text and National Agency for Food and Drug Administration Control (NAFDAC) number and a concatenated model was developed. Due to product availability, only 25 of Nigeria's registered pharmaceutical brands were considered in this study, with few fake drug package samples obtained from NAFDAC. Other samples were created using adapted techniques from existing studies. The proposed model utilized the high-level features layer of two (2) different CNNs, ResNet-50 and VGG16. The extracted features from the CNNs were concatenated to form a robust model (V16R50-Concat) and extract required features from the medicine package images, capable of detecting the medicine's genuineness and classifying the brands. The V16R50-Concat model demonstrated improved performance in brand classification and fake medicine detection, with an accuracy of 96.0% and 95.7%, respectively. The outstanding accuracy of the test data achieved in this study outperformed benchmark studies in brand recognition and detection, demonstrating potential for improved public health in Nigeria.

Keywords: CNN, V16R50-Concat, Fake, Medicine, Brand, Logo, Packages.