The present invention relates to a method for blindly embedding and extracting a watermark by using wavelet transform and a human visual system (HVS) model, which obtains both robustness and invisibility by applying the HVS model of NVF or JND imitating a human visual system to a middle frequency band for wavelet transformation and using a quantization step determined adaptively according to the importance of wavelet coefficient. A method according to the present invention includes the steps of: decomposing an original image by a wavelet and selecting a middle frequency band as a watermark embedment area; obtaining an HVS model having a human visual recognition information at each embedding position; determining the importance of the coefficient for each embedding position, and adaptively determining a quantization step for each embedment location by using the importance and the HVS model; quantizing each pair of embedment area by the quantization step and variably embedding a watermark sequence into a middle frequency band according to a value of the watermark; and performing inverse wavelet transform on an overall area into which the watermark sequence is embedded, and generating a watermarked image.