Comparison of Texture Feature Extraction Methods for Machine Condition Monitoring and Failure Analysis

The diagnosis of worn and damaged surfaces is an important issue in machine failure analysis and condition monitoring. Of many approaches used, image classification based on feature parameters has often proven to be particularly useful. Accurate classification can, however, be limited by the fact that feature parameters vary with scale and orientation. Hence, it is essential to determine which feature parameters are both scale and rotation invariant. This paper presents a performance evaluation of feature extraction methods currently used in pattern recognition. A comparison of six methods is conducted, in order to find the method that provides the most consistent results over a large range of image sizes and rotations. The methods analysed are: co-occurrence matrix, discrete wavelet transform, combination of wavelet and co-occurrence features, Gabor filter, circular Gaussian Markov random field and local binary patterns. For the comparison, four datasets of images with different scales and rotations are used, i.e. Brodatz textures, artificially generated isotropic fractal images and Talysurf images of sandblasted and abraded steel surfaces. The performance of each method is evaluated on the datasets using k-nearest neighbours and linear based normal densities classifiers. The results showed that the combined feature extraction method produced the most robust and accurate results for each of the datasets, and appears to be suitable for the classification of tribological surfaces.