An image-based approach for defect detection on decorative sheets

Abstract

In this paper, we propose a novel image-based approach for defect detection on decorative sheets. First, an image-based data augmentation approach is applied to deal with imbalanced image sets and severely rare defeat images. Two deep convolutional neural networks (CNNs) are then trained on augmented image sets using feature-extraction-based transfer learning techniques. Finally two CNNs are combined to classify defects through a multi-model ensemble framework, aiming to reduce the false negative rate (FNR) as much as possible. Extensive experiments on augmented artificial images and realistic defeat images both achieve surprisingly FNR accuracy results, which substantiate the proposed approach is promising for defect detection on decorative sheets.