From Classification to Regression: Model Transfer for Visual Aesthetic Quality Assessment

Visual aesthetic quality assessment has played an important role in increasing number of computer vision applications. Particularly, estimating the quality score precisely is a main task of aesthetic quality assessment, but the training samples labeled with score are usually expensive to obtain. In this paper, we propose a transfer learning method which can improve the performance of aesthetic score prediction by using the coarse labeled samples, which are much easier to obtain. The proposed method incorporates the coarse information from source domain into the target domain by a novel multi-task framework, which can revise the model in target task. The effectiveness of our method is proven by experimental results that the error is reduced obviously with the help of source domain.