Drug and Vitamin Package Image Classification Using HOG and HSV Color Histogram Features: A Baseline Study on a Balanced Dataset

Abstract

Image-based identification of drug and vitamin packages plays an important role in supporting inventory management, pharmaceutical distribution, and visual product verification. Although deep learning approaches have demonstrated superior performance in image classification tasks, establishing a conventional baseline remains essential for objectively assessing the improvement achieved by more advanced models. This study evaluates the effectiveness of combining Histogram of Oriented Gradients (HOG), HSV color histogram features, and the Nearest Centroid classifier as a baseline model for drug and vitamin package image classification. These methods were selected because they are computationally efficient, interpretable, and widely adopted as conventional benchmark approaches before implementing deep learning models. The dataset consists of 10,000 balanced images representing 10 pharmaceutical product classes. The data were stratified into 8,000 training images and 2,000 testing images. HOG descriptors and HSV color histograms were extracted from each image and subsequently classified using the Nearest Centroid algorithm. Experimental results show that the proposed baseline achieved an accuracy of 9.35% and a macro F1-score of 8.34%. These findings indicate that handcrafted visual features are insufficient to capture the complex visual characteristics of pharmaceutical packages, including product names, logos, colors, illustrations, and layout structures. Therefore, this study establishes an objective baseline while providing empirical evidence that more discriminative feature learning methods based on deep learning are required for reliable pharmaceutical package image classification.