DPAR-Net: Group-Conditional Art Exam Assessment via Dual-Path Reference Encoding

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Image Aesthetic Assessment (IAA) has developed rapidly in recent years, but most existing studies focus on predicting the absolute quality of individual images. In contrast, intra-group relative assessment under a unified evaluation context in art exam scenarios remains insufficiently explored. To fill this gap, we construct the Art Exam Dataset (AED), which contains 3360 art exam works from 93 exam themes, covering three major categories: sketch, quick sketch, and color painting. Based on AED, we propose the Dual-path Reference Encoder with Adaptive Routing Network (DPAR-Net), which provides a preliminary group-conditioned assessment framework for art exam works. DPAR-Net consists of a Dual-path Reference Encoder and an Adaptive Routing Fusion module. The Dual-path Reference Encoder models both a context-aware representation with respect to the group reference and a residual-relative representation that emphasizes sample-wise deviation from the group context. The Adaptive Routing Fusion module dynamically integrates the single-image baseline prediction with two correction candidates according to the sample state. Experiments on AED show that DPAR-Net outperforms representative general aesthetic assessment methods, validating the effectiveness of group-context correction and adaptive routing fusion. This work provides a data foundation and methodological reference for fine-grained ranking assessment of art exam works, and also offers a new experimental scenario for group-conditional aesthetic assessment.

This publication has 14 references indexed in Scilit: