JAPAN TAPPI JOURNAL
Online ISSN : 1881-1000
Print ISSN : 0022-815X
ISSN-L : 0022-815X
Research Report (Original Paper)
Analysis of Few-Shot Learning Characteristics in a Mixed-Paper Image Recognition AI
─ Effects of Training Image Composition and the Weighted Feature Distribution Index (WFDI)─
Takao Ando
Author information
JOURNAL FREE ACCESS

2026 Volume 80 Issue 10 Pages 982-991

Details
Abstract

In recent years, the supply of recovered paper in Japan has declined, increasing the importance of efficient mixed-paper sorting and collection. The authors previously developed “ECOPON,” a mixed-paper sorting support application using the Vision Transformer (ViT)-based image recognition AI “ADFI.” Because mixed paper exhibits substantial variation in shape, print, and appearance, and large training datasets are difficult to obtain under practical recycling conditions, understanding the characteristics of few-shot learning is essential.

This study investigated the few-shot learning characteristics of ADFI using three mixed-paper categories: boxboard packages, envelopes, and flyers. The number of training images was gradually increased, and changes in classification accuracy, classification confidence (p1), classification margin (Δ), and the Weighted Feature Distribution Index (WFDI) were analyzed. WFDI was newly defined using human-interpretable features (size, print, and shape) to quantify variation among training images. In addition, classification behavior was examined by fixing the number of training images at three while changing image combinations.

Classification accuracy reached its maximum with four to five training images, whereas p1 and Δ decreased around five images and partially recovered at six images. These results indicate that classification performance in few-shot learning does not necessarily improve monotonically as the number of training images increases. When three training images were used, higher WFDI values were associated with slight increases in p1 and Δ but a slight decrease in classification accuracy. These findings suggest that different selections of training images alter the distribution of samples within the feature space, affecting decision-boundary estimation and classification performance.

Although WFDI does not directly measure the internal feature space of ViT, it may serve as a proxy indicator of feature-space spread and representativeness through human-interpretable intra-class variation. The proposed WFDI provides a practical framework for evaluating training-image diversity and may contribute to the design of few-shot learning systems for mixed-paper recycling support.

Content from these authors
© 2026 Japan Technical Association of the Pulp and Paper lndustry
Previous article Next article
feedback
Top