XJTLU team¡¯s AI framework offers solution to data scarcity

A research team from Xi¡¯an Jiaotong-Liverpool University has developed an artificial intelligence (AI) framework designed to help deep-learning models learn more effectively with minimal data.

Led by PhD candidate Siyue Yao from XJTLU¡¯s School of Advanced Technology, the project ¨C Crucial-Diff ¨C marks a shift from the conventional focus on collecting ¡°more data¡± to identifying and generating ¡°the right data¡±.

Yao explains that traditional data-augmentation techniques, such as rotating or cropping images, only multiply similar samples and offer limited support for deeper learning. Instead, Crucial-Diff targets the AI model¡¯s blind spots.

¡°Our aim is not to expose AI to more data but to show it the most valuable data. Improving AI¡¯s recognition accuracy requires crucial samples, those that reveal its weaknesses and compel it to improve,¡± says Yao, whose work was published in the leading international journal IEEE Transactions on Image Processing.

Crucial-Diff works by generating challenging training samples that the model is likely to misclassify, much like a coach designing exercises to address a student¡¯s shortcomings. It comprises two key modules: a scene-agnostic feature extractor, which interprets underlying patterns in images; and a weakness-aware sample miner that identifies where the model struggles and then generates targeted synthetic data.

Defect images generated from real samples. Crucial-Diff¡¯s output (top right) simulates hard-to-detect anomalies encountered in real applications, while existing methods (bottom right) repeat typical patterns.

In addition to creating realistic images, the framework produces pixel-level annotations that typically require extensive manual labelling. This reduces labour demands and accelerates model training, a key advantage in cases when expert-labelled data are costly or difficult to obtain.

The team has tested the framework in several real-world scenarios.

In industrial quality inspection, where defects such as scratches, cracks, and air bubbles are rare and variable, Crucial-Diff produced challenging synthetic images, including defects under uneven lighting and with blurred edges. When these were added to the training set, the AI model¡¯s defect-detection accuracy increased to 83.63%, outperforming traditional augmentation methods.

Defect-detection performance after training with samples generated by Crucial-Diff.

In medical imaging, Crucial-Diff generated endoscopic images with variations in lighting, viewing angles, and blockages to enhance AI performance in identifying polyps, providing a more reliable tool for assisted diagnosis and highlighting the framework¡¯s potential for supporting medical applications that encounter privacy concerns and limited access to patient data.

Researchers say the innovation marks a conceptual shift in AI training.

¡°Industries increasingly rely on AI for critical tasks, yet continue to face data scarcity. Crucial-Diff demonstrates that smarter learning, rather than larger datasets, might be key to future progress,¡± says Professor Eng Gee Lim at XJTLU¡¯s School of Advanced Technology, who jointly supervised the project with Dr Mingjie Sun, an Associate Professor at Soochow University.

From left: Professor Eng Gee Lim, Siyue Yao, and Dr Mingjie Sun.

Yao plans to extend the approach to more complex areas and explore autonomous learning in dynamic, multidimensional settings, with a view to supporting the development of AI systems that remain robust even without massive datasets.