CLASSIFICATION OF HARDWOOD SPECIES USING MULTIRESOLUTION FEATURE EXTRACTION TECHNQUES

Abstract

Identification of wood is an important and difficult issue to deal with because of its complex biological structure. Correct identification of wood species is necessary for price fixation, fraudulent checking, protection of threatened plant and tree species at risk, and helping custom officials in proper assessment of wood species and implementation of tariffs accordingly. Wood is by and large classified into hardwood (HW) and softwood (SW) species. Softwood trees have limited number of cell types, which makes discrimination of softwood species a difficult task. On the other hand, hardwood species have a complex cellular structure and are easy to distinguish amongst the similar species. The wood identification task is accomplished by using 1) traditional, and 2) machine vision based methods. In, traditional approaches, wood experts usually identify the wood species by examining surface of the wood specimen at two different stages, first with the naked eye, then with a magnifier. Recognition of large volumes of wood species, employing traditional approach is prolonged, erroneous and unfeasible sometimes. To overcome the problems associated with the traditional methods of wood identification, machine vision technology was employed for wood identification. It works on the principle of obtaining certain statistical parameter of the wood species; thus, helpful in minimizing the error in wood identification. The machine vision based systems outperform experienced officers when large volumes of wood species are to be identified repeatedly with utmost accuracy, without getting fatigued. The macroscopic, microscopic and stereogram images of wood species have been used by the researchers for forest/wood species classification. A comprehensive analysis of the state-of-the-art work published in the area of forest/wood species classification shows that different texture features and classification algorithms are used for wood species identification. Further, microscopic images of wood carries significant information useful in its precise identification as compared to limited information possessed by macroscopic images. Hence, current work utilizes microscopic images for hardwood species identification. The classification accuracy of hardwood species can be improved either by getting appropriate discernible texture features from the image using suitable feature extraction technique and/or by using suitable classifier. However, performance of the classifiers highly depends on the quality of its input features. This has been the motivation behind the proposal of using multiresolution texture feature extraction techniques to extract the significant features of the hardwood specie images for their efficient classification. In the light of the above background, the objectives of the present research work have been broadly categorized as design and development of suitable feature extraction techniques, and determination of vessel elements of hardwood species. Thus, the key objectives of this research work are formulated as: (1) Design and development of multiresolution texture feature ii extraction techniques to acquire substantial information of the microscopic images of hardwood species, useful in their classification; (2) Employing feature dimensionality and feature selection techniques to investigate their effect in improving the classification of hardwood species; (3) Selection of appropriate multiclass classification algorithms to evaluate the effectiveness of multiresolution feature extraction techniques for classification of hardwood species, and determining the best combination of feature extraction and classification algorithms for better discrimination of hardwood species; and (4) Segmentation and determination of vessel elements to compute their hydraulic conductivity and lumen resistivity. To accomplish the above objectives, initially the state-of-the-art texture feature extraction techniques have been investigated for the classification of hardwood species. The texture descriptors produced by the state-of-the-art texture feature extraction techniques are based on spatial interactions over a fixed neighbourhood size on single scale image, which is appropriate for micro-texture analysis only. But, the microscopic images of hardwood species have four key elements namely vessels, rays, parenchyma and fibres, that too of varied shapes and sizes. In order to identify these images efficiently, it must be analysed at several scales of resolution as referred above. The smaller objects are the candidates to be examined at higher resolutions; whereas large size objects need to be examined at coarse view (lower) resolutions. To achieve multiresolution features, the images are decomposed at several levels of resolution; wherein each of the subimages coefficients contain varied and valuable information. Thereafter, the texture feature extraction techniques namely variants of local binary pattern (LBP), local configuration pattern (LCP), local ternary pattern (LTP), completed local binary pattern (CLBP) and local phase quantization (LPQ) are chosen to extract significant information from multiresolution images. The feature vector data obtained from each of the subimages are concatenated to increase the significant information of the images. The multiresolution based texture feature extraction techniques have tendency to produce large number of complex features and many of the features may not be significant. Keeping this aspect in mind, feature selection and feature dimensionality reduction techniques have been employed not only to improve the classification accuracy but at the same time to minimize the computation time needed by classification algorithms. The principal component analysis (PCA) has been incorporated to reduce feature vector data dimensions by computing a few orthogonal linear combinations of the original dataset features with maximal variance. In addition, the minimal redundancy maximal relevance (mRMR) feature selection technique based on mutual information quotient has been chosen to eliminate the irrelevant features, and retain a subset of features that efficiently describes the observed input data. The classification algorithms do play important role in improving the overall classification accuracy. Thus, four widely used classification algorithms, namely, linear discriminant analysis (LDA), random forest (RF), linear SVM and radial basis function (RBF) kernel SVM have been iii employed to see the effect of features produced by proposed texture feature extraction techniques on the classification accuracy. Further, the performance of proposed feature extraction techniques, have been assessed with two different approaches; namely, 1) The 10-fold cross validation, and 2) Randomly divided database (RDD). Furthermore, under the individual approaches the results have further been categorized under full feature vector data (FFVD), PCA reduced feature vector data and mRMR feature selection based reduced feature vector data. The present research work initially, examines the efficiency of several state-of-the-art texture feature extraction techniques for the classification of hardwood species. The selection of the feature extraction techniques have been based on their performance achieved in the various areas of image processing applications. In the objectives mentioned in the beginning, the transform domain techniques have been opted here due to their multiresolution capability for analysing images at different frequencies for several levels of resolutions. The different frequency sub-band images provide substantial information about the various objects of the images compared to the information obtained from spatial domain grayscale images. In, multiresolution feature extraction technique category, first approach has been design of the binary wavelet transform (BWT) based texture feature extraction techniques. By and large, the grayscale images have been utilized for extraction of texture features. It has been observed that the most significant bit (MSB), bit-plane (b7) of grayscale image contributes ample amount of information to the overall image. Thus, the BWT based LBP variants texture feature extraction techniques have been proposed to classify the hardwood species. The performance of BWT based LBP variants texture feature extraction technique has been found comparatively superior or at par with most of the state-of-the-art texture feature extraction techniques for hardwood species’ classification. The BWT based LBP variants have used only the MSB bit to extract the texture features of images; thus, the classification accuracy cannot be improved beyond a specific limit as the MSB bit does not correspond to 100% information about the image. To further enrich the quality of texture features, texture feature extraction techniques based on the Gaussian image pyramid (GP) model has been proposed. The selection of GP has been based on less computational efforts required in its implementation. Amongst, the GP based texture feature extraction techniques, the Gaussian image pyramid based local phase quantization (GPLQP) technique has produced the best classification accuracy, better than the techniques discussed in the preceding paragraphs. In addition, the discrete wavelet transform (DWT) based LBP (DWTLBP) variants texture feature extraction techniques have been chosen for hardwood species classification as DWT has the property to emphasize the directional information of the images. The texture features iv obtained from these directional subimages further help to the enrichment of feature vector data; which in turn help in better discrimination of the hardwood species. Amongst the proposed DWT based LBP variants texture feature extraction techniques, the DWT based uniform completed local binary pattern (DWTCLBPu2) texture features processed by mRMR feature selection technique has yielded the best classification accuracy. Further, to improve the classification accuracy of hardwood species, DWT based hybrid texture feature extraction techniques have been proposed; where, DWT decomposed subimages have been profound to be used with LBP and first-order statistics (FOS) techniques to get the tentative features. Moreover, it is proposed to fuse the features obtained from LBP and FOS methods. This technique thereafter is investigated to extract the texture features of both, the grayscale and color (RGB) images. It is observed that the mRMR feature selection based texture features of discrete wavelet transform based first-order statistics and local binary pattern histogram Fourier features (DWTFOSLBP-HF) technique produces the best classification accuracy for both types of images. Also, the texture features acquired by DWTFOSLBP-HF texture feature extraction technique for hardwood species are of excellent quality and no significant information loss is observed when grayscale image is employed for the classification in place of RGB image. For the purpose of evaluating the performance of the proposed features extraction techniques with the help of classifiers, an open access database of hardwood species consisting of 75 different categories has been selected. These microscopic images of hardwood species are correctly labelled by the experts in the laboratory of wood anatomy at Federal University of Parana, Curitiba, Brazil. In addition to the above classification work, a platform independent tool based on simple digital image processing technique has been developed to quantify the vessel elements of hardwood species. A prototype model has been developed and has been tested on several microscopic images prepared at the Xylarium (DDw) of the Wood Anatomy Discipline of the Forest Research Institute, Dehradun. The analysis of the experimental work suggests that with the help of appropriate parameter selection, the vessel elements are being extracted for most of the images. Further, along with the extraction of vessel elements, the proposed model is capable of computing the hydraulic conductivity and lumen resistivity of the vessel elements, which in turn provides helping hand to the wood anatomist in characterizing the woods. In nutshell, the significant contribution of this thesis work can be summed up as the proposed multiresolution feature extraction techniques, which help in to extract the discernible features of the microscopic images of hardwood species and the improvement in the classification accuracy because of them. The determination of vessel elements while using segmentation approach can be considered as the further contribution.