Steel surface defect classification using multi-resolution empirical mode decomposition and LBP

In this work we introduce Multi-Resolution Empirical Mode Decomposition (MREMD) as an image decomposition method that simplifies the implementation of Empirical Mode Decomposition (EMD) for bidimensional data. The proposed method is used in conjunction with the local binary pattern (LBP) to classify the images of six types of defects that can be found on the surface of rolled steel. The process starts by performing MREMD on the training images to obtain the first bidimensional intrinsic mode function (BIMF). Then features are extracted from the images and their first BIMF using the LBP. These features are used to train an artificial neural network (ANN) classifier. After training, given an unknown test image containing a defect, MREMD is applied on it to obtain its first BIMF. Next, LBP features are extracted from the image and its first BIMF and these features are fed to the trained ANN classifier to assign the image to one of the six defect classes. The classification process is carried out on 900 test images of the NEU database of six types of surface defects. The approach achieves an overall accuracy that is better than the result obtained using the LBP features alone. The main contribution of this paper is the introduction of multi resolution envelope interpolation using downsampling and upsampling with a fixed window size that reduces the execution time and decrease the sensitivity of the resulting BIMFs to the positions and number of extrema in the input image. (paper)