An Improved SVDD Algorithm for Fast Real-time Response

Support Vector Data Description (SVDD) is a one-class classification algorithm. It can distinguish target samples from non-target samples, and is usually used in anomaly detection and fault detection. However, by expanding the kernel of the traditional SVDD decision function, it is found that the running time complexity is linearly related to the number of support vectors. Therefore, in order to achieve fast real-time response, it is necessary to speed up the running speed of its decision function. In this paper, by using the direct image searching method we can find the pre-image of the spherical center which is no longer an expansion expression of support vectors. Finally, the decision function of quick SVDD (QSVDD) is obtained, so that the running time complexity is a constant in the support number. Therefore, regardless of the size of the training set, the kernel extension in the traditional SVDD decision function can be naturally reduced obviously, which solves the problem of reducing the SVDD test time complexity to a great extent. We believe that QSVDD can be applied in the process of product quality inspection, which will greatly improve the speed of this process.