Font Size: a A A

Research On Wood Identificantion Based On Semi-supervised Learning

Posted on:2016-12-20Degree:DoctorType:Dissertation
Country:ChinaCandidate:J L LvFull Text:PDF
GTID:1108330461985488Subject:Mechanical and electrical engineering
Abstract/Summary:Request the full-text of this thesis
Wood is tightly related with our life. However, there are so many kinds of wood with different performance and large price gap. The right identification of wood is of great significance for wood study, rational utilization and market circulation as well as wood management.In this paper, the core algorithm and key technology for wood automatic and accurate identification based on machine vision and semi-supervised learning are studied. On the basis of analyzing the macroscopic and microscopic characteristics of wood and wood identification methods, a total of 86 kinds of visual features of wood such as colors and textures are extracted, the method of principal component analysis (PCA) is adopted to reducing the dimension of these characteristics. To make full use of abundant and cheap unlabeled wood images and improve the identification accuracy of wood, five kinds of classifiers based on semi-supervised learning for wood identification are studied, such as Laplacian regularization support vector machine (LapSVM) in single view and Hessian regularization support vector machine (HesSVM) in single view, Laplacian regularization support vector machine in multi-view (mLapSVM) and Hessian regularization support vector machine in multi-view (mHesSVM), as well as Hessian regularization support vector machine in multi-view based on sparse coding (mHesSCSVM).The paper is mainly focus on the following aspects.(1) Based on the analysis of characteristics of common wood, a total of 86 kinds of color features and texture features are extracted according to the traits of wood. Color histogram and moment are calculated respectively in channels of H, S and V in HSV color space. The characteristics like energy, entropy, moment of inertia and correlation of wood texture are extracted based on gray-level co-occurrence matrix and the influence of different direction and growth step angle on extraction of wood features are analyzed. Concerning people can feel wood organically, roughness, contrast, direction, linelikeness are also extracted. The extracted features are as many as 86, thus the method of principal component analysis (PCA) is applied to reduce the dimensions in prevent of reducing operation efficiency. From the analysis of the results,99% information of wood characteristics can be well obtained with the use of former 10 principal components which corresponding to former 10 eigenvalues.(2) On the base of analysis on support vector machine (SVM), semi-supervised learning wood identification under single view is studied. SVM is supervised learning, it cannot make full use of unlabeled samples of wood and its generalization ability is poor. Due to those problems of SVM, Laplacian regularization support vector machine and Hessian regularization support vector machine are put forward. Laplacian regularization support vector machine introduces Laplacian regular terms into SVM objective functions through timber feature adjacencies which expresses the manifold distribution of training samples effectively and are able to improve the classification performance of SVM substantially making use of sample relations of local structure. Samples besides training field tend to be constant identification in Laplacian regularization; HesSVM replaces Laplacian regularization with Hessian regularization and brings it to SVM objective functions in case of the shortcoming in Laplacian regularization. Hessian regularization possesses more ample null space and is able to make a better linear estimation to the samples beside the field of training samples; Hessian regularization has a better identification performance. The experimental results show that, the algorithm which is put forward in this paper is more precise in wood identification.(3) The method of multi-view semi-supervised learning is proposed concerning wood has a transverse, tangential section and longitudinal section and a variety of micro characteristics. Multi-view Laplacian regularization SVM and Hessian regularization SVM are studied. Wood samples are represented with multi-dimension characteristics and each feature is regarded as an angle of view and a learner as well. Multi-view learning framework which integrated with multi-core learning and the overall figure learning is raised and though learning optimized-weights under various angles of view. Combined with multi-view and SVM, it is able to carry on the reasonable optimization on the different characteristics of the samples and has a better identification result compared with single view. The experimental results show that, in wood identification multi-view learning method has better identification ability and more effective, especially training samples are relatively less.(4) Multi-view Hessian regularization sparse coding SVM is studied to identify wood. In this way, wood labels are regarded as additional perspective characteristics and integrates multi-view features in sparse coding SVM. After that, it coded the adjacent map using Hessian regularization and keeps the local suppleness geometrically. Drive solutions change smoothly along streamlined geodesic, integrate Hessian regularization and discriminant function seamlessly. Complementarity under various views is synthesized and precision of wood identification, discriminant ability is improved without the increase of computational complexity.The experiment of wood identification results showed that the proposed wood identification methods based on semi-supervised learning in this paper are able to realize the goal of automatic and accurate wood identification. People are able to get rid of the overdependence on wood experts. This research will provide a new technical means and corresponding theory basis for wood automatic identification.
Keywords/Search Tags:Wood identification, Semi-supervised learning, Multi-view, SVM, Manifold regularization
Request the full-text of this thesis
Related items