Triplet attention-based deep learning model for hierarchical image classification of household items for robotic applications

Hierarchical classification is important in automation and robotics in terms of providing extra pieces of information about the classified object. In this paper, a hierarchical classification model using a triplet attention-embedded deep learning model is proposed for household items for robotic applications. Here deep learning models are applied to modified office-home image datasets of the household to give out two outputs from the model. The first prediction gives the name of the class, and the second output gives the hierarchical output indicating the location of the object identified. Triplet attention layers are added intermittently for improvement of the overall performance of the classification. The attention-based deep learning model gives 86% accuracy for the coarse class and 80% accuracy for the fine class at 25 epochs. The analysis is also done based on precision, recall, and F1-score. Precision values are 100% for the coarse class and 83.1% for the fine class. Recall is 82.5% for coarse and 77.7% for the fine class. F1-score is 83% for the coarse class and 78.9% for the fine class. This experiment lays a foundation for hierarchical image classification. The performance was compared with the hierarchical classification models that were previously applied for fashion datasets.