An Incremental Feature Extraction Method for Weld Surface Defect Recognition

Abstract

To meet the requirements of online feature extraction for weld defect recognition, an Incremental 2DPCA algorithm (I2DPCA) without estimating the covariance matrix of weld surface images is firstly proposed. A set of weld features with multiple orthogonal principal components can be obtained by a recursive estimation method, instead of the eigenvalue decomposition method or the singular value decomposition method. The computational complexity will be greatly reduced and the online feature extraction speed will be improved. To extract more structure features and reduce the dimension of weld images, an Incremental row-column Bidirectional PCA (IBDPCA) algorithm is presented based on I2DPCA, which preserves the weld features in the horizontal and vertical directions of weld images, and realizes the incremental feature extraction and data dimensionality reduction in the row and column directions. Finally, a series of comparative experiments are performed on the collected weld defect dataset. The results show that the proposed algorithm is superior to other PCA algorithms in convergence rate, classification rate, and complexity. The convergence rate is more than 99%, the classification rate can reach 96%, which can meet the online feature extraction requirements for weld defect recognition.