Contrast Metric Enhancement Based on Memory Extraction for Online Class-incremental Learning in Image Classification

In view of the catastrophic forgetting of previous knowledge in class incremental learning for image classification, existing replay-based methods focus on memory updating and sampling, while overlooking the feature relationships between old and new samples. To this end, we propose a method called contrast metric enhancement based on memory extraction(cME<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) for Online Class-incremental Learning in Image Classification, which designs two new types of positive and negative sample-pairs, enhances the reuse of old sample feature information, and strengthens the ability of model to express redundant features and common features. It improves the distribution of samples in embedding space based on the nearest class mean classifier. Finally, the effectiveness and efficiency of the proposed method were verified by comparison experiment and ablation experiment.