Adaptive Hybrid Seam Carving Through Stable Alpha-Regime Prediction: A Cross-Dataset Evaluation

Abstract

Hybrid seam carving combines gradient-based structural energy and saliency-based semantic energy using a weighting parameter, alpha. Using a fixed alpha for every image may not adequately represent variations in image content. This study investigates whether four lightweight image features—edge density, mean gradient magnitude, entropy, and saliency coverage—can predict broad alpha-weighting regimes. For each image, nine candidate alpha values were evaluated, and provisional labels were assigned according to similarity with a conventional seam-carving reference. These labels therefore represent reference-derived weighting preferences rather than perceptually optimal alpha values. Exact-value regression produced negative cross-validated R² values, while a preliminary three-class formulation failed to identify the Transition category reliably. The task was consequently reformulated as binary classification of stable Semantic (α≤0.4) and Structural (α≥0.8) regimes. Label-stability analysis showed that 88.5% of RetargetMe images and 91.4% of NRID images had near-maximum-SSIM alpha values confined to one reference-derived regime. On 61 stable binary RetargetMe samples, class-balanced Logistic Regression achieved a balanced accuracy of 0.728±0.157, a macro F1-score of 0.658±0.140, and a ROC-AUC of 0.728±0.194 under repeated stratified cross-validation. For the selected configuration, a permutation test produced p=0.002. On 23 independent stable binary NRID samples, the model obtained a balanced accuracy of 0.623, a macro F1-score of 0.518, and an MCC of 0.226. The findings demonstrate statistically significant reference-regime predictability within RetargetMe but only modest cross-dataset transfer. Therefore, the proposed framework should be interpreted as a feasibility study of lightweight reference-derived alpha-regime prediction rather than a complete solution for perceptually optimal alpha selection.