부산대학교 도서관

The rating of quicksketch works is an important job in the daily life of art teachers, and it is a challenge to apply artificial intelligence to the automatic rating of quicksketch works. At present, there is little research on the application of computer technology to the rating of quicksketch works. Most of the existing methods still focus on the recognition of art works by hand features or depth features extracted from deep convolutional neural networks. We believe that the rating of quicksketch works can benefit from both the drawing order of works and the structural correlation of drawing objects. In this paper, we propose a novel network structure for the rating of quicksketch works. Firstly, the network learns the deep abstract features through the pre-training model ConvNext-Base. Secondly, the feature map of the abstract features is divided into multiple patches, and learns the stroke order of the quicksketch works and the structural correlation of the painting objects through the graph structure feature module and stroke feature module respectively. Finally, We used a multi-layer perceptron to get the final quicksketch rating score. Experiments show that our model outperforms multiple baseline models, both the CNN series and the Transformer series, on the dataset we built.