The images of steel surfaces are generally textural images. There are different texture analysis methods to extract features from these images. In those methods using multi-scale/multi-directional analysis, Gabor filters are used for feature extraction. In this paper, w More
The images of steel surfaces are generally textural images. There are different texture analysis methods to extract features from these images. In those methods using multi-scale/multi-directional analysis, Gabor filters are used for feature extraction. In this paper, we extract texture features using the optimum Gabor filter bank. This filter bank is designed in a way that diverse filtering frequency and orientation will allow it to extract considerable amounts of texture information from the input images. We also introduce a new method called Gabor composition for segmentation and defect detection of steel surfaces. In this method, using two different algorithms, the input image is decomposed into detail images using an appropriate Gabor filter bank and then selected detail images are re composed. The created feature map illustrates the defective areas well. By calculating data distribution of detail images and comparing them, the second method of Gabor composition can accomplish segmentation without needing the normal images and the number of detail images to re-compose. Furthermore, we did different tests towards optimizing of segmentation by means of classifiers. Using a K-means classifier and adding gray levels to the extracted features, complete the segmentation procedure. The experimental results show that the Gabor composition method in most of the tests has got better defect detection performance than the ordinary K-means classifier and the standard wavelet method; also the Second method of Gabor composition has got the best performance over all.
Manuscript profile
