Quantitative evaluation of coal mine dust based on image quality assessment and meta-learning

Abstract High-concentration dust poses significant hazards to coal mining and underground equipment operations. As a critical technology for ensuring safe coal production, current quantitative analysis methods face two major limitations: the model prediction confidence levels lack a direct correlation with dust concentration, and moreover, they exhibit poor generalization performance in real-world engineering applications.Based on this, this study proposes the integration of image quality assessment (IQA) technology into underground coal mine dust concentration quantification. We conduct a detailed analysis of how dust concentration impacts image feature parameters and identify seven features exhibiting high correlations with dust concentration-related quality scores. Furthermore, a quality score regression model (RBF-SVR) is proposed, which leverages these features to establish a feature-to-quality score mapping model, effectively achieving accurate prediction of dust concentration quality scores. To address the challenges of few-shot and cross-scene applications, we further propose a Meta-Learning Multi Layer Perceptron (MetaLeakyMLP) meta-regression model based on a Model-Agnostic Meta-Learning (MAML) framework with inner-loop multi-stage query error dynamic weighted optimization. Through comparative experiments combining multiple metrics and mainstream models, the results demonstrate that our proposed MetaLeakyMLP significantly outperforms state-of-the-art image quality assessment algorithms in evaluating dust-containing images.