In this paper, a real-time watermarking scheme based on Regularized Extreme Learning Machine (RELM) is proposed. Using the information provided by the reference positions, RELM can be trained at the embedding procedure and watermark is adaptively embedded into the blue channel of the original image by considering the human visual system. Due to the extreme training speed (always hundreds of times faster than BP neural network and Support Vector Machine (SVM)) and good generalized performance, the trained RELM can exactly extract the watermark from the watermarked image against image processing attacks within very short time, and this makes this method applicable in real-time environment. Extensive experimental results illustrate that our technique outperforms Kutter's and Yu's method against simple and multiple attacks.