A novel multiscale-multidirectional autocorrelation approach for defect detection in homogeneous flat surfaces

Defect detection in flat web surface products is a challenging task. Reliable vision-based systems for detection of defects require the suitable selection of a huge set of parameters which highly impact the performance of these systems such as image resolution/scale, size of the scanning window, feature extraction, direction of scanning, classifier type and parameters and system performance evaluation measures. This paper addresses these issues and introduces a novel multi-scale and multi-directional (MSMD) autocorrelation function (ACF)-based approach for reliable defect detection and localization in homogeneous web surfaces. The proposed approach has been experimentally tested on samples from the well-known TILDA textiles database and wallboards. Performance evaluation using the system Precision, Recall (Sensitivity), Specificity, Accuracy, Youden’s index, F-measure and Matthews correlation coefficient has shown that the MSMD ACF approach outperforms the state-of-the-art approaches like MSMD Log-Gabor filters. The MSMD ACFs approach results in better performance indicators for defect detection than the Log-Gabor based approach in addition to being about 2–6 times faster in defect detection.