Abstract
Generative Adversarial Networks (GAN) receive great attentions recently dueto its excellent performance in image generation, transformation, andsuper-resolution. However, GAN has rarely been studied and trained forclassification, leading that the generated images may not be appropriate forclassification. In this paper, we propose a novel Generative AdversarialClassifier (GAC) particularly for low-resolution Handwriting CharacterRecognition. Specifically, involving additionally a classifier in the trainingprocess of normal GANs, GAC is calibrated for learning suitable structures andrestored characters images that benefits the classification. Experimentalresults show that our proposed method can achieve remarkable performance inhandwriting characters 8x super-resolution, approximately 10% and 20% higherthan the present state-of-the-art methods respectively on benchmark dataCASIA-HWDB1.1 and MNIST.