Feature extraction for defect classification

This research project aims to determine the best method of image classification of the different defect types with high accuracy using Support Vector Machine (SVM) for classification. In the first experiment, features are extracted from defect images by convolving the image with Root Filter Set (RFS) filter bank to generate filter responses and processed using texture descriptor called Local Binary Patterns (LBP). By applying Discrete Fourier Transform to the LBP histograms of different training images, Local Binary Patterns - Histogram Fourier (LBP-HF) features are computed such that the features are rotationally invariant and less noisy. The LBP-HF features are fed into Support Vector Machine (SVM) to be trained and classified. In the second experiment, a different approach is taken. Instead of processing the filter responses with LBP, the filter responses are processed using Complete Local Binary Patterns (CLBP) to retrieve the magnitude histogram of the filter response of the training defect image. The CLBP magnitude histograms of different training images are fed into Support Vector Machine (SVM) to be trained and classified. The classification rate in both experiments are based on the number of accurate predictions over total number of test images. Confusion matrix are generated to observe the percentage of each type of defect misclassified into other classes. Other than the two experiments that have positive results, this report also briefly discuss about different experiments tried in this period of FYP.Bachelor of Engineering (Computer Science