A deeper Newton descent direction with generalised Hessian matrix for SVMs: an application to face detection. (10th February 2021)

• Record Type: Journal Article
• Authors: Amir, Abdessamad; Amine, Korichi Mokhtar El
• Is Part Of: International journal of mathematical modelling and numerical optimisation. Volume 11:Number 2(2021)
• Page Start: 196
• Page End: 208
• Publication Date: 2021-02-10
• Subjects: support vector machines -- quadratic programming -- Newton method -- face detection; Mathematical models -- Periodicals; Mathematical optimization -- Periodicals; 519.605
• Languages: English
• ISSNs: 2040-3607

Abstract

By formulating the generalised Newton descent direction according to the parameter resulting from the calculation of the subgradient of the max function, a new version of NSVM (Fung and Mangasarian, 2004) is presented in this paper. This descent direction ensures even more the precision of the solution in a fast time. Associated with a good features extraction technique like Gabor's wavelets, the application of the proposed method in the context of facial detection shows that either the direction is calculated as an optimal solution of a one-dimensional problem or by a heuristic approach, manages to detect faces not detected by advanced methods.