Article: Analysis of children's painting quality using reinforcement learning and U-Net algorithm Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.81 pp.71 - 89 Abstract: In this study, an innovative model fusing reinforcement learning and U-Net is proposed to address the demand for automated analysis of children's painting quality. The model extracts multi-level visual features of paintings using an enhanced U-Net and introduces a reinforcement learning agent to simulate the sequential decision-making process of expert evaluation, thereby enabling comprehensive analysis of multidimensional quality indexes such as composition, line, colour, and integrity. The experiment is based on a self-built children's painting data set containing 15,832 labelled samples. The model achieves an evaluation accuracy of 92.6% on the test set, and its Spearman's rank correlation coefficient is 0.891, which is significantly better than those of the traditional CNN method and the basic U-Net model. The ablation experiment further shows that introducing a reinforcement learning reward mechanism reduces the model's average absolute error in the composition rationality analysis task by 0.15, effectively improving the stability and interpretability of the evaluation results. This study verifies the model's effectiveness for analysing children's painting quality and provides a new technical path for intelligent evaluation of aesthetic education. Inderscience Publishers - linking academia, business and industry through research

Title: Analysis of children's painting quality using reinforcement learning and U-Net algorithm

Authors: Xin Qi

Addresses: Fushun Vocational Technology Institute, Fushun, Liaoning 113122, China

Abstract: In this study, an innovative model fusing reinforcement learning and U-Net is proposed to address the demand for automated analysis of children's painting quality. The model extracts multi-level visual features of paintings using an enhanced U-Net and introduces a reinforcement learning agent to simulate the sequential decision-making process of expert evaluation, thereby enabling comprehensive analysis of multidimensional quality indexes such as composition, line, colour, and integrity. The experiment is based on a self-built children's painting data set containing 15,832 labelled samples. The model achieves an evaluation accuracy of 92.6% on the test set, and its Spearman's rank correlation coefficient is 0.891, which is significantly better than those of the traditional CNN method and the basic U-Net model. The ablation experiment further shows that introducing a reinforcement learning reward mechanism reduces the model's average absolute error in the composition rationality analysis task by 0.15, effectively improving the stability and interpretability of the evaluation results. This study verifies the model's effectiveness for analysing children's painting quality and provides a new technical path for intelligent evaluation of aesthetic education.

Keywords: reinforcement learning; U-Net algorithm; children's drawing analysis; image quality evaluation; intelligent education.

DOI: 10.1504/IJICT.2026.155112

International Journal of Information and Communication Technology, 2026 Vol.27 No.81, pp.71 - 89

Received: 26 Feb 2026
Accepted: 29 Mar 2026
Published online: 27 Jul 2026 *

Title: Analysis of children's painting quality using reinforcement learning and U-Net algorithm

Keep up-to-date