Incremental Image Classification Method Based on Proxy Features

A typical downside of deep learning methods is the problem of catastrophic forgetting, that is, once a new dataset is used to train an existing model, the model will lose the ability to identify the original dataset. In this paper, an incremental image classification method based on proxy features is proposed, which can add new categories by reasoning. Firstly, the image features are extracted by the feature extraction model, and a triplet loss function based on proxy features is constructed to improve the feature mapping ability of the feature extraction model as well as the convergence speed of the model in the training process; Secondly, a method of computing proxy features is proposed, based on which the proxy feature library is constructed, and the increment of the proxy feature library is used to achieve the increment of the image category; Finally, the performance of the proposed algorithm is compared with other similar algorithms, and the experimental results show that the proposed triplet loss function based on proxy features performs better than that of the traditional triplet loss function. In the well-established dish recognition dataset, the classification accuracy of the base class dataset is 93.9%, and the average classification accuracy of the incremental classification dataset is 89.4%, which verifies the feasibility and effectiveness of the proposed method.