The authors of the Magnetic Tile Surface Defect Dataset explore the significance of magnetic tiles in engines, where maintaining a productive automatic assembly line for magnetic tile manufacturing remains a fundamental challenge. It consists of 1344 images with the following defect types: blowhole, crack, fray, break, uneven (caused by the grinding process), and images with no defects.

Surface defect detection is crucial for filtering out subpar products, yet this process often involves significant manual effort. In Zhejiang Province, China, the largest magnetic tile production base globally, approximately three-quarters of the workforce is engaged in manual product quality inspections. In response to this, various image processing techniques, such as those based on grayscale value, gradient edge, wavelet, curvelet, and shearlet transformations, have been proposed to automate inspection tasks. Recently, deep neural network models have gained prominence for surface defect detection, demonstrating state-of-the-art performance in classification tasks. These advancements underscore the feasibility of employing computer vision for surface examinations and the potential to reduce reliance on human labor.

To delve deeper into the possibilities, authors suggest a real-time model called MCuePush U-Net, specifically designed for saliency detection of surface defect. Magnetic tiles typically feature 4–6 curved surfaces, causing image distortion. Challenges in automatic damage detection include the complexity of texture, the variety of defect shapes, and the randomness of illumination conditions, contributing to grayscale image noise. Target defects, such as blowholes, cracks, breaks, and frays, often manifest on curved surfaces without fixed patterns, introducing randomness to image detection tasks. Overcoming these challenges is essential for building robust computer vision models, particularly when aiming for real-time performance in the magnetic tile industry.

In image-based surface detection for magnetic tiles, the focus is primarily on minimizing the interference of product texture. While embedding denoising techniques like wavelet, curvelet, and shearlet transformations can effectively extract desirable features, they also lead to extended processing times, far from meeting real-time standards. Notably, the performance of experienced maintenance workers still surpasses machinery in terms of detection accuracy and efficiency. Transitioning from manual detection to automated surface inspection (ASI) requires substantial improvements in computer vision model performance.

For the dataset and evaluation metrics, a total of 1344 images are captured, and the region of interest (ROI) of magnetic tiles is cropped and classified into six datasets based on defect types: blowhole, crack, fray, break, uneven (caused by the grinding process), and free (no defects), each with pixel-level labels. To simulate real assembly line conditions, images of a given magnetic tile are collected under multiple illumination conditions.