Convolutional Neural Networks (CNN) based single image dehazing methods have recently gained much attention. However, as they heavily rely on synthetic haze images, existing CNN-based dehazing methods have limitations in achieving visually pleasant results, especially for real haze images. Inspired by the recent advances in generative adversarial networks (GAN), this paper proposes a novel end-to-end image dehazing network for image dehazing. Different from the existing CNN-based dehazing methods that were trained with paired hazy and hazy-free images, the proposed network was trained with paired and unpaired hazy datasets. To this end, the perception loss expressing high-level semantic information has been proposed. Experimental results show that the proposed method achieve substantial improvements over current state-of-the-art dehazing methods.