An improved automatic defect identification system on natural leather via generative adversarial network

ABSTRACT

The past decade has witnessed an impetus in the rise of artificial intelligence techniques in a diverse range of computer vision-relevant applications. Specifically, automated optical inspection technology has been garnered substantial research attention due to its advantages of enabling amelioration in the manufacturing process and improving the efficiency of industrial production. To date, the studies and evaluations into the defects on the raw leather pieces involving observational analysis tend to rely heavily on domain experts who have undergone intensive training. Herein, an automatic inspection system is introduced to distinguish whether a leather patch contains any defective areas. In particular, the dataset involved in this experiment comprises a total of 375 images that are made up of three classes, viz, black line defect, wrinkle defect and non-defective. Notably, owing to the data scarcity issue, the Generative Adversarial Network (GAN) is adopted to discover the feature regularities to produce plausible additional training samples. The proposed algorithm is systematically analysed, and compelling performance is yielded when compared to previous works. Succinctly, with a relatively small amount of readily captured training data, the classification performance is capable of achieving 100% accuracy. As such, the proposed system may be served an appealing strategy, especially in dealing with leather defect recognition tasks.

KEYWORDS:

• Leather
• defect
• classification
• GAN
• small sample

Acknowledgments

The authors would like to thank colleagues and anonymous reviewers for their valuable comments and suggestions to improve the quality of the paper. This work was supported in part by Ministry of Science and Technology Taiwan under Grant MOST 109-2221-E-035-065-MY2, MOST 108-2218-E-009-054-MY2 and MOST 108-2218-E-035-018-.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the Ministry of Science and Technology, Taiwan [MOST 110-2221-E-035-052-, MOST 108-2218-E-009-054-MY2, MOST 109-2221-E-035-065-MY2].