Abstract
With the development of cognitive computing, machine learning techniques, and big data analytics, cognitive support is crucial for automated industrial production. The real-time automated visual inspection in industrial production is a challenging task. Speed and accuracy are crucial factors for the process of automating the defect detection. Many statistical and spectrum analysis approaches have been introduced; however, they suffer from high computational cost with average performance. This paper proposes a neighborhood-maintaining approach, which is based on the minimum ratio for fast and reliable inspection of industrial products. The minimum ratio between local neighborhood sliding windows is used as a similarity measure for localizing defection. Extreme learning machine is then adapted to classify surfaces to defect or normal. A defect detection accuracy on textile fabrics has achieved 98.07% with 91.29% sensitivity and 99.67% specificity. The minimum ratio shows highly discriminant power to distinguish between normal and abnormal surfaces. A defective region produces a smaller value of minimum ratio than that of a defect-free region. Experimental results show superior speed and accuracy performance over many existing defect detection methods.