Integrated multi-task feature learning and interactive active optimization for scene retargeting in preschool educational applications.

In artificial intelligence (AI), effective adaptation of educational imagery across diverse screen formats is essential, particularly in preschool education, where visual content must simultaneously engage and instruct young learners. This study introduces a novel scene retargeting model tailored to preserve pedagogically significant visual elements during image resizing. The proposed framework leverages the binarized normed gradients (BING) objectness metric to efficiently identify and prioritize key regions within educational images, such as objects and facial features. A core component of our approach is integrating a locality-preserved and interactive active optimization (LIAO) mechanism, which simulates human visual attention by generating gaze shift paths (GSPs) that guide feature prioritization. These GSPs are further transformed into hierarchical deep features using a multi-layer representation, followed by refinement through a Gaussian mixture model (GMM) to enhance scene understanding and retargeting fidelity. Experimental evaluations demonstrate that the proposed model not only surpasses five state-of-the-art methods in performance but also achieves a 3% improvement in accuracy compared to the next-best approach, all while reducing inference time by over 50%. The results confirm the model's effectiveness and efficiency, offering a robust solution for educational content adaptation that aligns with cognitive and pedagogical requirements in early childhood learning environments.