The Study Of Recognition Methods Of Texture-based Wood Images

Wood recognition is a work involoved in how to identify the names of wood species, which is based on wood structure, color, texture, smell and other characteristics. In these wood classification features, wood recognition is mainly relied on the macroscopic and microscopic characteristics of wood. As we know, the correct recognition of wood is an important task to wood science development, reasonable use and management of timber resources, timber trade circulation, timber import and export management, and wood archaeology, etc. This dissertation conducts the research on the wood recognition based on the macroscopic and microscopic images of woods from the two aspects of the semantic features and texture features of woods. Specifically, the main works for this dissertation are summarized as follows:(1) We collected the wood samples from the main import wood species in East China regions, and constructed the database which contains the wood species name, origin, structures, properties, uses, macroscopic and microscopic images of wood specimen; then, established the web based wood herbarium information management system, which can research and retrieve the corresponding information by wood name and wood properties. To supply the services for research, teaching, production and business trade of related wood profession and industry, this information management system provides reliable data for correctly recognizing and understanding the nature of a variety of wood, and for taking more rational and efficient use of the timber resources. At present, we have collected more than500kinds of wood images, where the extracted information for their various application properties include physical, chemical, etc. At the same time, we selected some typical woods to obtain their stereogram images, and take the microcosmic images from the wood pieces after wood cut-section has been made. After pretreating these images, these pictures respectively form two data sets:ZAFU WS24and ZAFU W M22. We took these two datasets as the samples to verify the effectiveness of the proposed wood recognition methods.(2) We proposed a segmentation method for wood microscopic image, which is used a mixed level sets. Through introducing the local image information, we reduced the segmentation difficulty of wood microscopic image caused by the local uneven sections in the image. And with the used area histogram based on the circular regions, we can obtain the best threshold parameter for the closed area computation automatically. By using the area threshold, we can solve the impurities problems in the wood images, such as blisters. Then, according to the domain knowledge of wood science, the second objective function is introduced, which is the average area, a parameter about all the closed area in wood image, so the image segmentation task converts the one from the single object problem to a multi-objective one. The corresponding experimental results show that we can acquire a satisfying segmentation result under the different circumstances of many kinds of hardwood by using multi-objective function. Finally, according to the pores that we have got in the last step, we can judge their adjacencies by the circumcircle of pores and its neighboring regions; then, the adjacency degrees can be used to determine the combination of the pores. The combination mode of the pores, which we got from our segmentation experiments, is consistent with practical.(3) Based on higher order local auto-correlation (HLAC), we proposed a new concept:Mask Matching Image (MMI), to make up the deficiencies of HLAC and its expansion methods. This method can retain all the information about the template image by calculating the MMI of HLAC. From the MMI of images, we can not only obtain the local statistical characteristc, but also the geometrical characteristc information about the whole images, obtained effectively by the length histogram (LH) feature we proposed, which can be taken as a useful tool for the texture analysis of image. In addition, Max-Min Sorted HLAC (MMS HLAC) we proposed can be used to generate the features simply, quickly, and effectivly by sorting, and at the same time, avoid the problem of the template cumulative sum too big when the previouse HLAC based methods are used to calculate their features. The experimental results show that MMI and MMS HLAC can obtain a higher recognition rate in our wood datasets.(4) This dissertation proposed a method based on the modified blocked Gabor wavelet and Greedy Sort Search (GSS) feature selection algorithm, to conduct the investigation on the automatic wood classification by wood stereogram images. Firstly, we used Gabor wavelet to extract the texture features of wood. After analyzing the best parameters selection of the scale and orientation, we took a suitable means of partitioning on Gabor filter group according to the characteristics of wood stereography. In addition to the mean and standard deviation, more statistics features including entropy, contrast and other statistics are used to extract more effective features. Moreover, the search method for how to find the best subset, which is called Greedy Sorting Search (GSS), is used to reduce the feature dimension. So we can obtain the classification features which have the best capability of distinguishing different woods. The experimental results of wood classification show that the method we proposed can well improve the wood recognition rate and its working efficiency.