Single class classifier using FMCD based non-metric distance for timber defect detection

Abstract

In this work, we propose a robust Mahalanobis one class classifier with Fast Minimum Covariance Determinant estimator (MC-FMCD) for species independent timber defect detection. Having known in timber inspection research that there is a lack of defect samples compared to defect-free samples (imbalanced data), this unsupervised approach applies outlier detection concept with no training samples required. We employ a non-segmenting approach where a timber image will be divided into non-overlapping local regions and the statistical texture features will then be extracted from each of the region. The defect detection works by calculating the Mahalanobis distance (MD) between the features and the distribution average estimate. The distance distribution is approximated using chi-square distribution to determine outlier (defects). The approach is further improved by proposing a robust distribution estimator derived from FMCD algorithm which enhances the defect detection performance. The MC-FMCD is found to perform well in detecting various types of defects across various defect ratios and over multiple timber species. However, blue stain evidently shows poor performance consistently across all timber species. Moreover, the MC-FMCD performs significantly better than the classical MD which confirms that using the robust estimator clearly improved the timber defect detection over using the conventional mean as the average estimator.