Subspace Learning-based Face Modality Transfer

Human face plays an important role in human social communication. Each face carries unique individual information, including race, gender, age, etc. Images are the most common information carriers for faces. And a face can be expressed by different image modals. Although these images depict the same individual, the information carried is greatly different. Transfer between different modal images has many practical applications. In this paper, subspace learning is imported to analyze the intrinsic connection between different modal facial images, so as to synthesize facial photos from simple line drawings and colorize gray scale facial images automatically.Simple line drawing is the abstract form of image which only contains contour information. Synthesizing facial photos from simple line drawings need to predict lost texture information. In this paper, we propose a method based on Markov random fields and locality-constrained linear coding to synthesize facial photos from simple line drawings. Taking account of the special characteristic of the simple line drawing, an edge descriptor is extracted from each patch of the input image. Experimental results have validated the effectiveness of the proposed method.Aiming at the drawback of existing colorization algorithms that require human interactions, we present a new framework based on subspace learning to automatically transform the given gray scale facial image to the corresponding color one. Coupled dictionary learning is employed to capture the nonlinear relation between gray scale space and color space. In addition, Markov random fields is imported to ensure the global constraint. The experimental results demonstrate that the proposed framework is effective to colorize the gray scale facial images to the corresponding color ones automatically.