Perceptually Relevant Pattern Recognition Applied to Cork Quality Detection

This paper demonstrates significant improvement in the performance of a computer vision system by incorporating the results of an experiment on human visual perception. This system was designed to solve a problem existing in the cork industry: the automatic classification of cork samples according to their quality. This is a difficult problem because cork is a natural and heterogeneous material. An eye-tracker was used to analyze the gaze patterns of a human expert trained in cork classification, and the results identified visual features of cork samples used by the expert in making decisions. Variations in lightness of the cork surface proved to be a key feature, and this finding was used to select the features included in the final system: defects in the sample (thresholding), size of the biggest defect (morphological operations), and four Laws textural features, all working on a Neuro-Fuzzy classifier. The results obtained from the final system show lower error rates than previous systems designed for this application.