Class Incremental Learning Based on Playback Images Generated by Classification Network

Image classification has surpassed human performance in numerous domains. However, in real-world scenarios, challenges such as storage constraints, privacy concerns, and commercial data protection necessitate the enhancement of existing models to accommodate new class classifications — a critical area of research known as class incremental learning. Data replay stands out as a pivotal technique within this domain. The DeepInversion algorithm ingeniously employs classification networks to generate training set images of notable quality. In this study, we improve the DeepInversion algorithm, leveraging a pre-trained model to yield superior quality playback images for class incremental learning. Within the context of class incremental learning, we introduce a cosine orthogonal classification loss function, formulated on the basis of linear layer analysis, to guide image generation and augment class incremental learning. This loss function is designed to ensure the mutual orthogonality of all class centers while minimizing intra-class distances. In the realm of knowledge distillation, we harness a combination of limited real images, a profusion of synthesized images, and Mix virtual images to facilitate feature cosine distance distillation. Sufficient comparative experiments and analyses with similar latest methods underscore the efficacy of our proposed approach, and obtain the SOTA results. The code for the paper can be found at http://github.com/YunXiaooooo/Class-incremental-learning-based-on-playback-images-generated-by-classification-network