Multiregion Difference-Based Defect Detection for Complex-Texture Caps

Abstract

Complex textures of bottle caps pose challenges for defect detection with smooth upper, transitional middle, and densely striped lower regions. This study proposes a multiregion detection method integrating background subtraction, Difference of Gaussians, and translation-registration difference. The sidewall is divided into three regions by texture. The upper region employs background subtraction and Difference of Gaussians to extract high-contrast dark defects and enhance shallow indentations. The middle region utilizes background subtraction to detect grayscale anomalies. The lower region combines dual-scale background subtraction with translation-registration difference to detect point-like and block-like dark defects and enhance tiny damage embedded within the stripes. This method avoids the annotation difficulties that deep learning faces with low-contrast tiny defects. Experiments on 300 caps demonstrate an accuracy of 94.7%. Furthermore, improvements of 12.4% and 14.7% over global background subtraction and Difference of Gaussians are achieved respectively, offering a viable solution for defect detection with complex textures.