人工知能学会

Counting industrial objects is challenging due to their similar appearances and complex shapes. This paper adapts Few-Shot Counting (FSC) to minimize labeled data requirements while improving accuracy. We use FamNet with rule-based feature detection to enhance robustness in industrial settings. Additionally, we introduce the INDT dataset, focusing on diverse industrial objects. Our approach integrates density map estimation with feature detection to improve interpretability and reduce over-counting errors. Experimental results show improved accuracy on industrial objects and strong generalization to other datasets, highlighting FSC’s potential for industrial automation, with future work aimed at optimizing model structure and feature extraction for further performance improvements.