Visual Explanations of ResNet 101 for Blister Package Classification

Abstract

This paper presents an efficient approach to generate visual explanations from a ResNet 101 model for identification of medication. The blister package dataset is used to train a deep learning model built on the PyTorch framework's ResNet 101 pre-trained model. Visual inspections and a quantitative localization benchmark demonstrate that the model approach correctly identifies the critical components of blister packs for medicine identification. A Gradient-weighted Class Activation Mapping (Grad-CAM) method is used to extract the feature map, and then the attention mechanism is utilized to extract the high-level attention maps, that emphasizes the part of the image that is significant to the target class, which can be considered as a visual representation. The experimental results indicated that the Grad-CAM achieved the better visualization and interpretation of ResNet 101 in blister package classification.