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. However, large image databases can be computationally costly to analyse, and the datasets are susceptible to noise. Hence, it is essential to determine which feature parameters hold the most useful information, in order to improve the classification rate and computation time. This paper presents a performance evaluation of dimension reduction techniques currently used in pattern recognition. A comparison of three methods is conducted, in order to determine which is able to produce the best results over a large range of image datasets. The methods analysed are: Non-Linear Fisher, Principal Component Analysis and kernel Principal Component Analysis. These are then tested against four different classifiers to obtain the best combination. These classifiers are: Linear Discriminant Classifier, Quadratic Discriminant Classifier, k-Nearest Neighbour and Support Vector Machine Classifiers. For further analysis, two combined dimension reduction and classification methods are tested: Minimum Classification Error and reduced feature space Support Vectors. 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 results showed that a combination of the Non-Linear Fisher dimension reduction technique and a Linear Support Vector Machine Classifier gave the best performance overall and are the most promising for the application in automated machine condition monitoring and expert free failure analysis. Further improvement can be achieved by performing a step-wise dimension reduction by first reducing the features using the Principal Component Analysis method, then further reduction with the Non-Linear Fisher technique.