Power-law transform based spectral features for texture image retrieval

Abstract

Searching speci c digital images from large resources is an area of wide interest. E cient access to digital images requires the development of techniques to search and organize the visual information. The traditional text based retrieval approach is not e ective because it is not only expensive and time consuming process but also very subjective. Content based image retrieval (CBIR) is another image retrieval technique which is very e ective. It automatically extracts image contents using low-level image features. Among the low-level image features, texture is the most important and prominent visual feature of an image. After extensive studying on existing texture feature descriptors, we nd existing texture image retrieval systems are still incapable of providing retrieval result with high retrieval accuracy. To improve the CBIR performance, it is very important to nd e ective and e cient texture feature that can represent images more accurately. In this research, we propose two new texture features named Power-law transform based Gabor feature and Power-law transform based curvelet feature. These features are used to represent images and to measure the similarity among them. The retrieval outcome shows, the proposed Power-law based Gabor and curvelet texture features outperform the conventional Gabor and curvelet features in terms of retrieval precision. To obtain highest retrieval outcome, di erent level transform-power values are also investigated. To observe the retrieval performance of the proposed texture features, retrieval tests are performed using four di erent types of image databases. The experimental evaluation of the system is based on 2800, 11200, 5264 and 5600 texture images of original Brodatz database, scale distorted database, rotation distorted database and illumination distorted database, respectively. From the experimental results, we nd that the Power-law based Gabor and curvelet texture features are robust to di erent types of distortion and signi cantly better than conventional Gabor and curvelet features.