Damped Newton based Iterative Non-negative Matrix Factorization for Intelligent Wood Defects Detection

The Non-negative matrix factorization (NMF) can be formulated as a minimization problem with bound constraints. NMF is capable to produce a regionor partbased representations of the wood images. We present an extension to the NMF and discuss the development as well as the use of damped Newton optimization approach for update matrices W and H called iterative DNNMF with good convergence property for wood defects detection by adding a diagonal correction to the stiffness matrix and employing a Newton direction in the line search until any constraints become active. We also provide algorithms for computing these new factorizations and the supporting theoretical analysis. DNNMF is tested with color wood images based on the statistical features extracted by local binary pattern (LBP) from the feature spaces. Finally, we present experimental results that explore the properties of the proposed method. After many comparative experiments, the test results show DNNMF is effectual and practical with good research values and potential applications.