DualRW: a dual fusion network for rating quicksketch works

Quicksketch rating is a crucial component of the China Fine Arts College Entrance Examination and an important measure for evaluating the quality of students’ quicksketch works. With the advancements in artificial intelligence, intelligent rating of quicksketches has emerged as a promising and challenging task. However, very few studies have applied computer technology to the rating of quicksketch works. Most existing methods focus on the recognition of fine art works using hand-crafted or deep features extracted. In this paper, we propose a novel architecture named Dual Fusion Network for the rating of quicksketch works, which utilizes both position information and channel information. We collect 2442 quicksketches from the Guangdong Art Joint Examination and construct a dataset named SCNU-QuickSketch. The proposed architecture comprises a multi-scale feature extraction module, position attention subnetwork, and channel interaction subnetwork. The multi-scale feature extraction module extracts various feature maps from the convolutional neural network. This process effectively aggregates global information for the position attention and channel interaction subnetworks. The position attention subnetwork strengthens the interaction of different position information of the feature map, and the channel interaction subnetwork enhances the representation ability of specific semantics to achieve more detailed details of quicksketch works. The experimental results show that the Dual Fusion Network achieves 79.9% accuracy on the SCNU-QuickSketch dataset, which is 5.8% and 3.5% higher than the ResNet50 and Swin-B, respectively, and visualization shows the effectiveness of our proposed subnetworks. This study provides insights into the application of computer technology in the evaluation process.