A Transfer Learning Model based on Residual learning and Maxout For Sketch Works Ranking

Abstract

Art education is a special subject in the education system. However, art education still shows a considerable degree of immaturity and imperfection in modern education. For example, a huge gap in the number of art teachers, uneven teaching quality in rural and remote areas, inconsistent standards and extremely cumbersome manual scoring. Intelligently ranking the artworks can effectively alleviate the above problems. In this paper, we expand our dataset SCNU-SW and then give an enhanced transfer learning model to automatically rank sketch works. First, ResNeSt50 is selected as the backbone of transfer learning. Second, data Augmentation, Maxout, Dropout and Relu (AMDR) modules are added in a certain order into ResNeSt50 such that the classification performance and generalization ability of model can be enhanced. Third, we verify the generality of the AMDR module on most convolutional networks in sketch works ranking field. The experimental results on SCNU-SW610 show that our model achieves classification accuracy of 86.6% for ranking sketch works and outperforms the most mainstream models.