Abstract
Multiclass classification plays a vital role in image-based applications, such as food recognition systems, where the automatic identification of diverse food items is essential. Although modern Convolutional Neural Networks achieve high overall accuracy, they typically treat all classes equally. This uniform approach presents challenges in domains that require increased accuracy and recall for specific categories, such as hospital meal services, where dietary compliance is critical. To address this, the Priority-Aware Network (PA-Net) is proposed as a novel framework that introduces class prioritization directly into the learning process. At its core, PA-Net features a Priority-Aware Module integrated with two tailored loss functions: Priority-Weighted Loss, which applies static importance based on domain knowledge, and Priority-Weighted Recall Loss, which dynamically adjusts focus according to class-specific performance during training. The framework is evaluated on both a clinical healthcare meal dataset from The Canberra Hospital (TCH) and the public Food-101 benchmark. The results show that PA-Net improves the recall of priority classes without a!ecting overall accuracy. In the TCH dataset, it significantly reduces misclassification of clinically sensitive items, confirming its applicability in real-world healthcare settings. Similarly, improvements on Food-101 demonstrate the method’s broader generalizability. This work contributes a scalable and flexible approach to class-sensitive learning, advancing neural network optimization by aligning training with practical class importance.Beyond food recognition, PA-Net’s principles can extend to other safety-critical applications such as medical diagnostics, industrial inspection, and autonomous driving. Future work includes enhancing PA-Net’s adaptive mechanisms, improving scalability to large class vocabularies, and advancing interpretability and fairness to support deployment in real-world, risk-sensitive environments.
| Date of Award | 2025 |
|---|---|
| Original language | English |
| Supervisor | Wanli MA (Supervisor), Shuangzhe LIU (Supervisor) & Dat TRAN (Supervisor) |