Composition-Aware Feature Engineering for Lightweight Minimalist Image Quality Assessment Toward Mobile Deployment

Abstract

Aesthetic quality assessment (AQA) research has produced accurate but computationally demanding deep learning methods that are difficult to deploy on mobile devices. This study instead targets minimalist style detection as an interpretable, lightweight composition-analysis problem, extracting 45 handcrafted features per image from global statistics, Rule-of-Thirds spatial analysis, and composition ratios. Binary training labels are derived from a weighted composition score over four of these features, thresholded at the 75th percentile; K-Means, GMM, and DBSCAN clustering are evaluated on the same feature space for exploratory structure discovery, not as the label source. Classifiers trained on the resulting 49-dimensional feature set (45 composition features plus four metadata fields) reach a best of 99.24% accuracy, 0.9998 AUC-ROC, and 0.9848 minority-class F1 (XGBoost); because the composition score is itself computed from features available to the classifier, these metrics indicate rule recovery, not independent evidence of learned aesthetic discrimination. A Random Forest variant converts successfully to Apple Core ML, showing the approach is compatible with on-device deployment in principle; the higher-performing XGBoost model could not be converted with the available tooling, and no on-device latency measurements are reported.