A multi-task effectiveness metric and an adaptive co-training method for enhancing learning performance with few samples

Abstract

The integration of deep learning (DL) into vision inspection methods is increasingly recognized as a valuable approach to substantially enhance the adaptability and robustness. However, it is well known that high-performance neural networks typically require large training datasets with high-quality manual annotations, which are difficult to obtain in many manufacturing processes. To enhance the performance of DL methods for vision task with few samples, this paper proposes a novel metric called Effectiveness of Auxiliary Task (EAT) and presents a multi-task learning approach utilizing this metric for selecting effective auxiliary task branch and adaptive co-training them with main tasks. Experiments conducted on two vision tasks with few samples show that the proposed approach effectively eliminates ineffective task branches and enhances the contribution of the selected tasks to the main task: reducing the average normalized pixel error from 0.0613 to 0.0143 in pose key-points detection and elevating the Intersection over Union (IoU) from 0.6383 to 0.6921 in surface defect segmentation. Remarkably, these enhancements are achieved without necessitating additional manual labeling efforts.