Support vector machines (SVM) are learning algorithms, frequently used to solve classification and regression tasks, and supporting the linear separation of a binary class of data. SVM techniques have been extensively used to solve a large variety of pattern recognition problems, including isolated digit handwritten recognition, object recognition, speaker identification, face detection in images, and text categorization, as well as for regression estimation purposes. Comparative investigations of their corresponding efficiencies have proven that SVMs’ performance on generalization tasks either matches, or is significantly better than, that of competing methods.
Briefly, the SVM algorithm determines which vectors, from a finite binary labeled input pattern, support the hyperplane that gives the largest margin of separation between the classes. The computation of the coefficients corresponding to the separating hyperplane is carried out by solving a constrained quadratic programming problem.
This paper introduces a perturbation method, used to obtain a sensitivity measure for trained solutions from a set of input patterns, using SVM learning. The authors seek to extend the SVM technique to allow for the evaluation of the quality of image features resulting from trained samples taken from a noisy background. The main contribution of the paper is to present a new method to alter the training data for the SVM algorithm, to extract pixel-wise class data from a trained solution. The extraction of pixel features from a trained SVM problem is performed using the inhibitory perturbation method, and, for image classification, two additional sensitivity measures are considered to provide a direct correspondence to each pixel.