CNN Based Reliable Classification of Household Chores Objects for Service Robotics Applications

Abstract

Household chores service is one of the desirable functions for a service robot. Object classification is the most important function when searching for objects. We consider that the major concern is that the robot should not misclassify the object. If the robot misclassifies household object images, it will then perform the household tasks with the wrong object. This may cause serious damages to a service robot and to the user. This concept gives us the insight that the precision of the classification must be very high by setting a confidence threshold so that it then can claim a reliable service robot. By exploring this concept, we develop a more convincing indicator, Classification Reliability, to reveal the reliability of deep learning model. Moreover, we develop a fine-tune rule base to continuously regenerate more proper training dataset for the CNN model to increase reliability. Experimental results demonstrate that the CNN model fine-tuned by our closed-loop system achieves the reliability which is higher than the other similar effects such as DenseNet on the CIFAR-10 dataset.