Automatic classification of refrigerator using doubly convolutional neural network with jointly optimized classification loss and similarity loss

Abstract

Modern production lines for refrigerator take advantage of automated inspection equipment that relies on cameras. As an emerging problem, refrigerator classification based on images from its front view is potentially invaluable for industrial automation of refrigerator. However, it remains an incredibly challenging task because refrigerator is commonly viewed against dense clutter in a background. In this paper, we propose an automatic refrigerator image classification method which is based on a new architecture of convolutional neural network (CNN). It resolves the hardships in refrigerator image classification by leveraging a data-driven mechanism and jointly optimizing both classification and similarity constraints. To our best knowledge, this is probably the first time that the deep-learning architecture is applied to the field of household appliance of the refrigerator. Due to the experiments carried out using 31,247 images of 30 categories of refrigerators, our CNN architecture produces an extremely impressive accuracy of 99.96%.