Tanongchai Sookawat. The surface curvedness operator and its applications to image classification and image retrieval. Master's Degree(Computer Science). Mahidol University. : Mahidol University, 2003.
The surface curvedness operator and its applications to image classification and image retrieval
Abstract:
This research proposed a method to define a feature by using the
Surface Curvedness Operator or SC-operator. This operator can be used to describe
the altitude of surface curvature by using the concept of topographic structure, which
can be used to view pixel-based image as 3D surface for use in conjunction with
features derived from the Surface-Shape operator for image classification and image
retrieval applications. The SC-operator is related to the curvedness of the image
surface. The SC-operator was defined by using the relation of two eigenvalues of the
Hessian Matrix, which is the distance of the coordinate of eigenvalues from origin on
the eigenvalue plane. With the SC-operator, the feature called the Roughness Index
was derived to use together with the clumpiness and miscibility values, which are
derived from the Surface Shape Operator and used to describe the shape of image
surfaces. The Fractal Theory and the Scale Space Theory were used to develop the
roughness index in a multiscale approach. Euclidean distance and Mahalanobis
distance were used as classifiers and similarity measures.
In our experiments, the clumpiness and miscibility were calculated
from 4 different scales. While the roughness index was calculated from 5 different
scales. Images used in the experiments were obtained from the Brodatz Texture Image
Database, which consists of 112 categories. For each category, 18 images were
selected, ten of which were used as a training set and eight others as a testing set for
image classification. For image retrieval, 896 (112x8) images were used, as the
testing database.
Results showed that the roughness index could be used to enhance the
results of image classification and image retrieval applications. For image
classification, the percentage of correct classification was improved from 86.38% to
87.95%. For the image retrieval, the roughness index could be also used to enhance
the accuracy rate from 77% to 80% when considering the top 8 retrieved images and
96% to 98% when considering the top 100 retrieved images. In addition, it could be
concluded that Mahalanobis distance provides better results than Euclidean distance.