The invention discloses a multi-level manifold learning medical image color perceiving method in order to solve the problems in the prior art that a great amount of computation is required and that colors are not matched well. The method comprises the following steps: according to a black and white medical image, obtaining the gradient feature images in multiple directions; based on the similarities among all the pixels in the bottom layer, selecting the representative points from each region, establishing a hierarchical structure from bottom to top and the transmitting similarities among the representative points at multiple levels; using the manifold learning method to reduce the dimensions of the representative points at the top layer and maintaining the geometrical structures among the representative points as much as possible to obtain the three-dimensional coordinates of high-dimensional representative pixels; interpolating the three-dimensional coordinates of the representative points, and returning to obtain the three-dimensional coordinates of all the pixels; and according to the pixel color information of a given part, transmitting the color information and obtaining the color perception of all the pixels. The technical scheme of the invention realizes the color perception of a medical image quickly, highlights the image area structure and avoids the problem of color mismatch.