Study On The Object Recognition Based On Least Squares Support Vector Machine

The study on robots focuses on the Auto Mobile Robot whose vision is the key tool for the robot to perceive the external environment. One of the aims of robots'vision is to categorize the objects contained in the images. Image recognition is one aspect of the pattern recognition. And in the region of pattern recognition, support vector machine (SVM), as a new theory, has attracted many researchers'attentions and has succeeded in the pattern classification. When performing the pattern recognition procedure with SVM, the main calculation is the dot product of samples, so SVMs are very efficient to deal with high dimensional samples, which makes it easy for us to accomplish the aspect-based image recognition.The aspect-based image recognition method treats every single pixel of the image remaining to be recognized as a dimension of the training sample, so in this method the dimensionality of the training sample is very high. When training the traditional SVM, we need to do the procedure of quadratic programming (QP). Because this procedure takes up too much storage and spends too much time, especially in the high dimensionality case, the improved version of SVM– least squares support vector machine (LS-SVM)– is introduced here to accomplish the aspect-based image recognition. The reason why we use the LS-SVM is that it has some very attractive properties: first, the constrain conditions are made up of a set of linear equations instead of the inequations, so the calculation procedure becomes much simpler and more efficient; Secondly, the storage space needed in the LS-SVM procedure is much less than that in the traditional case. So it is very effective and suitable to perform the aspect-based image recognition using the LS-SVM.The experimental results show that the method proposed in this thesis can obtain very high accuracy of recognition rate under some kinds of disturbance (some kinds of noises, shifts, both of them or occlusions) with much less storage and much higher calculation efficiency compared to the traditional SVM. When the disturbance is not very big, all the object images are categorized without error. So when we finally apply the proposed method to the practical system, it obtains fine results.