Domain Transferred Image Recognition via Generative Adversarial Network

Recent studies have demonstrated that neural networks exhibit excellent performance in information hiding and image domain transfer. Considering the tremendous progress that deep learning has made in image recognition, we explore whether neural networks can recognize the imperceptible image in the transferred domain. Our target is to transfer natural images into images that belong to a different domain, while at the same time, the attribute of natural images can be recognized on domain transferred images directly. To address this issue, we proposed domain transferred image recognition to achieve image recognition directly on the transferred images without the original images. In our proposed system, a generator is designed for the domain transfer and a recognizer is responsible for image recognition. To be flexible for the natural image restoration in some cases, we also incorporate an additional generator in our method. In addition, a discriminator will play an indispensable role in the image domain transfer. Finally, we demonstrate that our method can successfully identify the natural images on transferred images without access to original images.