A Hybrid Semantic-Aesthetic Neural Network Driven by Multi-Layer and Perceptual Learning for Intelligent Packaging Picture Resizing

This study addresses the challenge of preserving both semantic meaning and aesthetic appeal in content-aware resizing of intelligent packaging images. Existing methods often struggle to maintain the complex visual and structural configurations characteristic of packaging designs during resizing, leading to loss of important semantic and aesthetic information. To overcome this limitation, a hybrid semantic-aesthetic neural network, HSANet, is proposed, which integrates multi-layer perceptual features to robustly represent packaging designs. A core component of the approach is the hierarchical aesthetic-semantic learning (HASL) model, which simulates human visual perception by leveraging semantic saliency and aesthetic quality metrics. Through locality-preserved aesthetic learning, HASL adaptively identifies semantically meaningful and visually pleasing regions across diverse packaging layouts, enabling the generation of content-preserving paths (CPP) that maintain spatial coherence while emphasizing regions aligned with consumer visual attention. These CPP-based deep features are incorporated into a dual-stream optimization framework, facilitating intelligent image resizing that balances semantic integrity with aesthetic preservation. Extensive empirical validation on the Packaging Aesthetics Dataset (PAD) and six benchmark datasets demonstrates significant improvements over existing methods in both semantic and aesthetic dimensions. The findings also reveal enhanced performance in automated packaging adaptation for various display formats.