Deep Learning Based Pharmaceutical Identification Through the Optimization of Feature Manifolds in the MobileNet-V2 Architecture

Abstract

Medication misidentification is a major health risk topatient safety, especially in the elderly who have to deal with anumber of medications. The proposed paper describes a newdeep learning method of drug identification based on animproved MobileNet-V2 architecture with feature manifoldoptimization. The proposed approach is a solution to the issue ofinter-class similarity between pharmaceutical products, whichinvolves a two-step refinement approach to fine-tuning, whichsubsequently enhances the ability to extract features. The initialphase defines feature correspondence globally by training on 20epochs, whereas the subsequent phase unfreezes the last 30layers and specializes by training on a reduced learning rate of1e-5. The results of the experiments on a filtered subset of 10high-volume prescriptions on the basis of the NIH Pill ImageArchive database prove the usefulness of the suggested method,with a peak validation accuracy of 72.46 and the ultimatevalidation loss of 0.8155. The PCA visualization makes sure thatthere is clear territory building of the features, which verifies theoptimization of feature manifolds for the recognition of visually-similar classes of pharmaceuticals. Lightweight architectureprovides the ability to deploy the lightweight architecture onmobile devices to provide real-time consumer-gradepharmaceutical identification without external computation.