Search

A noise filtration technique for fabric defects image using curvelet transform domain Filters is proposed in this paper. Firstly, we used FDCT_WARPING to decompose image into five scales curvelet coefficients. Secondly, the proposed algorithm distinguished major edges from noise background at the third scale. Thirdly, the possible lost edges in the procedure above were detected according to the decaying lever of the coefficients. Fourthly, the edges of the defect at the second scale were detected by four correlation coefficients in the two directions at the third scale. Fifthly, the curvelet coefficients at the fourth scale are filtered by the decaying lever. Sixthly, the curvelet coefficients at the fifth scale are filtered by hard threshing. Finally, the processed coefficients are reconstructed. The tests on the developed algorithms were performed with images from TILDA's Textile Texture Database, and suggest that the new approach outperforms wavelet methods in image denoising.

It is a challenging problem to overcome shift and rotation and nonlinearity in fingerprint images. By analyzing the shortcoming of fingerprint recognition algorithm on shift or rotation images at present, manifold learning algorithm is introduced. A fingerprint recognition algorithm has been proposed based on locally linear embedding of variable neighbourhood k (VK-LLE). Firstly, approximate geodesic distance between any two points is computed by ISOMAP ( isometric feature mapping) and then the neighborhood is determined for each point by the relationship between its local estimated geodesic distance matrix and local Euclidean distance matrix. Secondly, the dimension of fingerprint image is reduced by nonlinear dimension-reduction method. And the best projected features of original fingerprint data of large dimension are acquired. By analyzing the changes of recognition accuracy with the neighborhood and embedding dimension, the neighborhood and embedding dimension is determined at last. Finally, fingerprint recognition is accomplished by Euclidean distance Classifier. The experimental results based on standard fingerprint datasets have verified the proposed algorithm had a better robustness to those fingerprint images of shift or rotation or nonlinearity than the algorithm using LLE, thus this method has some values in practice.

Aiming at the change of battery location, environment light or camera location in Li/MnO2 automatic inspection process, a novel WT-FEBFNN (Wavelet Transform Fuzzy Ellipsoidal Basis Function Neural Network) approach to battery defect inspection is proposed. Firstly, WT is applied on original battery image, and low-frequency signal and de-noised signal is obtained, respectively, by setting different thresholds on different scale WT decomposition. Secondly, signal only containing defect (nick) is obtained by subtracting low-frequency signal from de-noised signal. Finally, model of FEBFNN is established and defect recognition is accomplished on 1000 battery images. Experiments have shown the proposed algorithm had a better robustness to the change of battery location, or environment light or camera location than multilayer perception(MLP), and shown that the reason for the high recognition accuracy in battery defect inspection is due to the information contents of the features as well as to proper classifier.