03-12-2012, 02:34 PM
Multiple Representations of Perceptual Features for Texture Classification
[attachment=42255]
ABSTRACT
Texture Classification plays a vital role in medical image, remote sensing image, pattern analysis for the past three decades. Eventhough it is three decades problem, still having a lot of scope in pattern analysis. Textural features corresponding to visual properties of texture are highly desirable for two reasons; they will be optimum in terms of feature selection and will be applicable to all kinds of textures. Some of the perceptual features are coarseness, contrast, direction and busyness. The aim of this paper is to present a new method to estimate these perceptual features. The proposal based on two representations: Original Image Representation and Autocorrelation Function Representation. These estimated perceptual features measures are applied to classification on large image data set, the well-known Brodatz database using k-nearest neighborhood classifier.
INTRODUCTION
Texture is an important item of information that human use in analyzing a scene. Literally, texture refers to the spatial distribution of grey-levels and can be defined as the deterministic or random repetition of one or several primitives in an image.
A number of texture analysis methods have been proposed [1]-[9]. Haralick [10] categorized texture analysis methods into statistical methods, structural methods, and hybrid methods [11]-[15]. The drawbacks of almost all of these approaches are that they do not have general applicability and computational cost involved, either in terms of memory requirement, computation time or implementational complexity. In comparison, human visual perception seems to work perfectly for almost all types of textures. The reason for this mismatch between computational methods and human vision is, the majority of the computational methods use mathematical features that have no perceptual meaning easily comprehensible by users.
Coarseness Estimation
When considering the autocorrelation function, we can notice that coarseness is saved in the corresponding autocorrelation function. Therefore, number of extrema in the autocorrelation function determines coarseness of a texture.
Coarseness, denoted Cs, is estimated as the average number of maxima in the autocorrelated images and original images. A coarse texture will have a small number of maxima and a fine texture will have a large number of maxima.
APPLICATION TO TEXTURE CLASSIFICATION
This experiment was carried out to assess the performance of the features in an actual classification task. Computational features presented in this paper are applied in texture classification on Brodatz database. The samples used are D13, D57, D77, D6, D75, D49, D30, and D15. Each of these images of Brodatz database was divided into 9 nonoverlapping tiles to obtain 72 subimages of size 128 x 128 (8 images x 9 titles per image). K-Nearest Neighbourhood Classifier is used to identify the best matching class based on these estimated perceptual features. In the classification, the technique of training on the data was employed, in this case leaving out four samples for each category at a time and training the classifier on the remaining five. After that four untrained samples for each class were presented to the classifier to identify. Thus, training set and testing set contains 40 (8 images x 5 titles per image) and 32 (8 images x 4 titles per image) subimages respectively.
CONCLUSION
In our work, an attempt has been made to develop measures that correspond to some textural properties. Four basic properties of texture, namely: coarseness, contrast, directionality, busyness were conceptually defined. The conceptual expressions were put into computational forms. In this approach, autocorrelation function was computed for a given image, and the features were derived from these autocorrelated and original images. Finally improved results in terms of classification accuracy were obtained for multiple representation using these developed features. The immediate prospect related to this work is consideration of other perceptual features such as regularity and complexity.
[attachment=42255]
ABSTRACT
Texture Classification plays a vital role in medical image, remote sensing image, pattern analysis for the past three decades. Eventhough it is three decades problem, still having a lot of scope in pattern analysis. Textural features corresponding to visual properties of texture are highly desirable for two reasons; they will be optimum in terms of feature selection and will be applicable to all kinds of textures. Some of the perceptual features are coarseness, contrast, direction and busyness. The aim of this paper is to present a new method to estimate these perceptual features. The proposal based on two representations: Original Image Representation and Autocorrelation Function Representation. These estimated perceptual features measures are applied to classification on large image data set, the well-known Brodatz database using k-nearest neighborhood classifier.
INTRODUCTION
Texture is an important item of information that human use in analyzing a scene. Literally, texture refers to the spatial distribution of grey-levels and can be defined as the deterministic or random repetition of one or several primitives in an image.
A number of texture analysis methods have been proposed [1]-[9]. Haralick [10] categorized texture analysis methods into statistical methods, structural methods, and hybrid methods [11]-[15]. The drawbacks of almost all of these approaches are that they do not have general applicability and computational cost involved, either in terms of memory requirement, computation time or implementational complexity. In comparison, human visual perception seems to work perfectly for almost all types of textures. The reason for this mismatch between computational methods and human vision is, the majority of the computational methods use mathematical features that have no perceptual meaning easily comprehensible by users.
Coarseness Estimation
When considering the autocorrelation function, we can notice that coarseness is saved in the corresponding autocorrelation function. Therefore, number of extrema in the autocorrelation function determines coarseness of a texture.
Coarseness, denoted Cs, is estimated as the average number of maxima in the autocorrelated images and original images. A coarse texture will have a small number of maxima and a fine texture will have a large number of maxima.
APPLICATION TO TEXTURE CLASSIFICATION
This experiment was carried out to assess the performance of the features in an actual classification task. Computational features presented in this paper are applied in texture classification on Brodatz database. The samples used are D13, D57, D77, D6, D75, D49, D30, and D15. Each of these images of Brodatz database was divided into 9 nonoverlapping tiles to obtain 72 subimages of size 128 x 128 (8 images x 9 titles per image). K-Nearest Neighbourhood Classifier is used to identify the best matching class based on these estimated perceptual features. In the classification, the technique of training on the data was employed, in this case leaving out four samples for each category at a time and training the classifier on the remaining five. After that four untrained samples for each class were presented to the classifier to identify. Thus, training set and testing set contains 40 (8 images x 5 titles per image) and 32 (8 images x 4 titles per image) subimages respectively.
CONCLUSION
In our work, an attempt has been made to develop measures that correspond to some textural properties. Four basic properties of texture, namely: coarseness, contrast, directionality, busyness were conceptually defined. The conceptual expressions were put into computational forms. In this approach, autocorrelation function was computed for a given image, and the features were derived from these autocorrelated and original images. Finally improved results in terms of classification accuracy were obtained for multiple representation using these developed features. The immediate prospect related to this work is consideration of other perceptual features such as regularity and complexity.