Open Set Deep Networks Based on Extreme Value Theory (EVT) for Open Set Recognition in Skin Disease Classification

A computerized skin disease classification system generally works on closed-set data, meaning images from unknown classes will still be classified as one of the known classes. In the Teledermatology system, skin disease classes are usually defined before the training process. However, in the real world application, it may receive images that belong to a new class or disease. To avoid misclassification, we have implemented the Extreme Value Theory of Weibull distribution function for out of distribution detection and incorporated the OpenMax layer to the deep networks for open-set recognition in skin disease classification. The system can classify seven classes of common skin disease in Indonesia with an accuracy of 71.64% using Inception v3 for closed-set data, while it achieved an accuracy of 83.33% for open-set recognition. The result indicates that the proposed method in this study has reached the purpose of recognizing open-set data in skin disease classification.