CREATING A NEW INPUT VECTOR TO IMPROVE THE PREDICTION PERFORMANCE OF SVM CLASSIFICATION

Support Vector Machine (SVM) classification, which is based on statistical learning theory, has been increasingly applied to many problems, because of its remarkable performance in prediction accuracy. As in neural network models, the input vectors for learning on SVMs is very important, because they affect predictive performance. Even though the performance of the learning model is different according to the attributes of input vectors, there is little research on determining the best input vectors. Normally in the application of neural network models and SVMs, only a single property, such as pixel gray or edge information, has been used for training input vectors. However, the composition of the input vectors may affect predictive performance and two (or three) kinds of information can be combined and used simultaneously as input vectors for learning.

In this paper, two different types of information, edge and pixel gray information, which are important and have often been used for image recognition with single information, have been combined and used simultaneously for learning in SVMs. The experimental results of this study show that the combined input vectors with two different types of images, edge detection images and gray-scale images, can provide better performance in predictive accuracy. Experiments are performed on real world image data of traffic scenes, with SVMs serving as the classifier. By using the proposed method, the accuracy of the SVM classifier can be improved and the data fusion for input vectors can be a new research topic to improve the prediction accuracy in machine learning systems.