Novelty Detection of Beverage Bottle Images Based on Transfer Learning

Abstract

The recent advances in image recognition are commonly based on the convolution neural network (CNN). Many CNN architectures have been investigated with a great success to build models that can recognize images correctly corresponding to the known classes. However, many applications require a model that can reject a novelty item that is not part of the known classes. Our goal is to solve the novelty detection problem by utilizing a pre-trained model approach. The pre-trained CNN models from the four well-known CNN architectures are used to extract the training features. The OC-SVM and Isolation Forest are implemented to train novelty detection models and to be investigated for performance evaluations. The F1 and F2 score are adopted as evaluation metrics and show that OC-SVM model trained on the features from NASNetLarge armed with a feature reduction achieves the best results in terms of detecting the novelty item comparing to other experimented CNN architectures. © 2020 IEEE.