A learning-based approach for aluminum tube defect detection using imbalanced dataset

Abstract

Modern industrial products are developing in intelligence, information, and integration. Hence higher quality requirements are put forward for the reliability of its parts. Computer vision-based automatic defect detection method shows a promising application prospect due to its advantages such as contactless and fastness. However, it is challenging to detect surface defects since the shape and location of the defects vary randomly. Moreover, this task becomes more difficult due to the defects' rarity, limiting the dataset's size and rendering it imbalanced. This study proposed a novel defect detection algorithm based on a metaheuristic self-organizing neural network (MSOM) to deal with the imbalanced dataset. The Maximum mean discrepancy was introduced to build the fitness function and the optimal structure of MSOM was determined by solving the fitness function iteratively. Experimental results showed that our method outperformed seven state-of-the-art algorithms on the imbalanced dataset of aluminum tube defects.