Dual-Stream Interaction Network for Assessing Image Technique and Aesthetic Quality

With the rapid proliferation of electronic devices, an increasing volume of images is becoming prevalent in people’s daily visual experiences. It is imperative to design an accurate method that holistically assesses images from both aesthetic quality and technique quality dimensions. However, most current studies treat image technique quality and aesthetic quality assessment as separate tasks, overlooking their intrinsic interrelation. In fact, there exists a mutually reinforcing relationship between image aesthetic quality and technique quality. In this paper, a dual-stream interaction network (DSINet) is proposed to predict the aesthetic quality and technique quality of images in parallel. Specifically, the feature distilling module (FDM) is designed for the first time to extract and enhance the representational abilities of aesthetic and distortion features. Then, an aesthetic-oriented feature interaction module (AFI) and a distortion-oriented feature interaction module (DFI) are introduced to establish the relationship between aesthetic features and distortion features. Finally, with the help of visual perception rules generated by a semantic adaptive hyper network (SAHNet), aesthetic and distortion features are further mapped into image aesthetic quality and technique quality scores, respectively. The experimental results demonstrate that our proposed approach outperforms state-of-the-art methods in the existing technique quality assessment and aesthetic quality assessment datasets.