Análise de Fotografias de Pílulas por Redes Neurais Convolucionais

Abstract

The automatic recognition of pills through photography may reduce administration medication error and also enable forensic intelligence to establish links among drugs, laboratories, local and international trafficking routes. Considering the remarkable progress and performance of the convolutional neural networks (CNNs) to extract image characteristics, we propose the use of CNNs for automatic recognition of pill images. The results of classification and content-based image retrieval (CBIR) are evaluated quantitatively by the Accuracy and mean average precision (MAP) measures, and qualitatively by cluster visualization of the U-Matrix. The results demonstrated that the LeNet outperformed the pre-trained Inception Resnet on classification and CBIR tests for legal pills, while Inception Resnet outperformed LeNet on illegal pills.