Real-Time Deep Learning-Based Automatic Pill Classification

This paper proposes a method of combining pill image processing with a convolution neural network modified VGG16 based on Adam optimization that ensures the high accuracy of defect detection and classification. The novelty lies in: (i) the development of real-time deep learning-based automatic pill defect detection; (ii) the proposed training structure: modified VGG16 structure and Adam optimization algorithms to find the optimal solution and (ii) the design and implementation of a model construction and training based on VGG16. For the training stage, qualified pill images in the real manufacturing system are collected as samples. In the prediction stage, the proposed method is implemented to validate pills and categorize them into three different categories (contamination, good, crack). The results validate the effectiveness of the real-time. It provided a high accuracy on real-time systems. High accuracy and fast processing time make the proposed method highly potential to apply to industrial systems.