Research On Detection Method Of Pulp Fiber And Paper Defect Based On Computer Vision

Pulp fiber and paper detect detection is an important way to ensure welding product quality. The traditional defect detection methods mainly rely on manual judgment with low efficiency and high mistake rate. In recent years, the use of computer vision for automatic detection of pulp fiber and paper detect has been paid more attention by researchers. On the basis of previous research results, researches on pulp fiber and paper detect detection involving image noise suppression, image enhancement, image segmentation, edge detection and classification methods have been done in this thesis, and are described as follows:Firstly, a denoising method based on pre-classification non-local means(NLM) filter for pulp fiber and paper detect images is proposed. On the basis of pre-classification, we draw useful lessons from the idea of median filter to improve non-local means filter, thus a new non-local filter for mixed noise is obtained. A large number of experimental results show that, compared with related methods of removing Gaussian noise, salt-pepper noise and their mixture, respective. The proposed method has great improvement in subjective visual effect, and also the objective quantitative evaluation indicators. It shows obvious advantages on removing not only Gaussian noise, salt-pepper noise but also their mixture.Then, an image enhancement method based on non-subsampled contourlet transform(NSCT) and Retinex function for pulp fiber and paper detect images is studied. A low-frequency and high-frequency components are produced after decomposition of paper detect and pulp fiber images through NSCT, then low-frequency component are adjusted by Retinex function, the threshold and gain function are adapted by high-frequency components which will be processed by the adapted threshold and gain function. Finally, enhancement for pulp fiber and paper detect images is realized by inverse NSCT. The experimental results show that compared with the bidirectional histogram method, enhancement method in wavelet domain, enhancement method in contourlet domain based on fuzzy theory, enhancement method in NSCT domain, the studied method can improve the contrast of enhanced image, and also enhance the details of pulp fiber and paper detect images.And then, an edge detection method based on nonsubsampled contourlet transform and kernel fuzzy C-means clustering is proposed. Firstly, an original image is decomposed into a low frequency component and high frequency components through nonsubsampled contourlet transform. Then edge information is extracted from low frequency component with less noise, kernel fuzzy C-means algorithm is used for clustering to obtain the edge image of low frequency component. While the method of nonsubsampled contourlet modulus maxima is applied to the high frequency components with more edges and details. Finally, the whole edges of image are obtained by fusing above-mentioned two parts. The experimental results show that, compared with four methods proposed recently, the proposed method has obvious superiority and better effect of edge detection, with accurate edge localization, continuous and complete edges, and abundant details.Next, a two-dimensional Arimoto gray entropy thresholding method for pulp fiber and paper detect images based on bee colony optimization or decomposition is proposed. Firstly, Arimoto gray entropy is definite. Two-dimensional Arimoto gray entropy thresholding method is derived, and fast recursive formulae for the intermediate variable are given. Then a modified artificial bee colony(MABC) optimization algorithm is adopted to find the optimal threshold of the two-dimensional Arimoto gray entropy method, and also to reduce the amount of computation. Finally, Arimoto gray entropy thresholding method are converted into two one-dimensional spaces, as a result, the computational complexity is further reduced. The experimental results show that, compared with three methods proposed recently, the proposed methods have better image segmentation performance and a good anti-noise performance.Finally, a recognition method for paper detect based on Krawtchouk moment invariants and wavelet support vector machine is proposed. The Krawtchouk moment and Krawtchouk moment invariants of paper detect images are calculated to construct feature vectors of flame images. Then a support vector machine is constructed according to the feature vectors of training samples. And the kernel parameter and penalty factor of support vector machine were optimized by chaos niche particle swarm optimization algorithm to obtain best recognition performance. A large number of experimental results show that, compared with the method based on Hu moment and support vector machine, the method based on Zernike moments and support vector machine, using Krawtchouk moment invariants as features of flame image can better recognize the paper detect, and the recognition rate is improved.