Classification System with Capability to Reject Unknowns

In this paper, we propose a novel method for object classification with capability to reject unknown inputs. In the real world application such as an image-recognition-based checkout system, it is crucial to reject unknown inputs while correctly classifying registered objects. Conventional deep-learning-based classification systems with softmax output suffer from overconfident score on unknown objects. We tackled the problem by the following two approaches. First, we incorporated a metric-learning-based method proposed for face verification into object classification. Second, we utilize available unregistered objects (known unknowns) in the training phase by proposing a novel “Margined Unknown Loss”. In the experiment, we showed the effectiveness of the proposed method by confirming that it outperformed conventional softmax-based approaches which also use the known unknowns, on two datasets, MNIST dataset and a retail product dataset, in terms of Recall at a low false positive rate.