Sketch works ranking based on improved transfer learning model

Classification and evaluation of sketch works is an important step in sketch teaching. The speed for manual evaluation of massive works is very slow and its cost is also high. Regarding the sketch works after scoring is hard to be collected, the limitations from small dataset create many challenges for training a great scoring model. In this paper, we firstly collects 400 teaching sketches of college students from South China Normal University and constructs a dataset. Then an improved transfer learning model based on ResNet50 was proposed for learning the high-level abstract characteristics of sketch works because transfer learning can reuse knowledge from a large dataset and employ the feature extraction capability in sketch subject scenario. In our improved model, three strategies (i.e., data augmentation, dropout and feature fusion) are used to prevent the model from early overfitting and improve the accuracy and stability of the model. Comparing with traditional feature extraction algorithms, our model provides an end-to-end mechanism. Moreover, compared with the accuracy of AlexNet and ResNet50, it is improved by 24.2% and 7.82% respectively. Our results indicate that the three strategies have outstanding effects for the transfer learning model for sketch works ranking.