Class-incremental learning (CIL) based on out-of-distribution (OOD) detection is a technique aiming at alleviating the problem of catastrophic forgetting. It enables the model continuously learn the features of new classes to accomplish new classification tasks without losing the ability to classify old classes as new classification tasks continue to arrive. Most existing autoencoder (AE) for OOD detection focus on reconstructing more image details, which doesn’t overcome the interference of background noise and suffers from CIL performance decline. Therefore, we propose a novel CIL framework based on OOD detection. Firstly, we utilize the classifier to supervise the training of the AE, which enables the model to capture more underlying features of in-distribution (ID) data and reduce the interference of background noise. Secondly, we generate masking vector by Taylor expansion to impose constraint on the latent space of the AE, which extract important information from latent vector while removing redundant information. Finally, to ensure stable training of our model, we adjust the proportion of important information in the latent vector during training our improved AE. The OOD analysis experiment results demonstrate that AE performs better than existing methods in OOD detection on CIFAR-100 dataset. Meanwhile, the CIL experiment results on MNIST, CIFAR-10, CIFAR-100, and TinyImageNet datasets demonstrate that our CIL framework performs better than the latest method.