A New One-Class Classification Method with Multiple Encoder-Decoder Pairs for Images
Dongxiang Chen, Chungang Yan, Mimi Wang · 2019
One of the main destinations of image classification methods is to screen out the images belonging to the target class (positive) and identify the images of other classes (negative). Although most classifiers are trained on both positive samples and negative samples, in reality, negative samples are often unavailable. One-class classifiers trained only on positive samples, are proposed to solve this problem. However, how to train effectively a classifier remains a daunting challenge. Considering the success of deep learning in the field of computer vision in recent years, we propose a one-class classification model for images based on convolutional neural network (CNN). The model consists of two parts of networks with different responsibilities. The network of the first part works as the discriminator, used to identify whether the images are positive. The networks of the second part, each of which consists of an encoder-decoder pair work as the guiders of the discriminator. They guide the discriminator to learn what images it should identify to be negative. Different encoder-decoder pairs can restore images to different degrees. Images restored to different degrees can be used to train the discriminator. The discriminator learns that images under a specific restored degree don't belong to the target class. The experiment results on CIFAR-10 show that our method can achieve good performance, with less difficulty in training than the GAN-based counterparts'.