Rotationally Invariant Color, Texture and Shape Feature Descriptors for Image Retrieval

Abstract

The incredible development in the number of digital images has aggravated the demand for step up in image retrieval. The prime challenge in image retrieval is retrieving akin images from huge image repositories with highest precision and least time and storage cost. Selecting apt feature descriptors plays noteworthy role in increasing the competence of retrieval systems for digital images. In this paper, integration of dominant  rotation local binary pattern (DRLBP), dominant edgels (DE) and moments based on DEs, and dominant color descriptors (DCD) are proposed for image indexing and retrieval and are extracting texture, shape and color details respectively. Self-organizing map (SOM) is employed for classification which increases accuracy and decreases the search space considerably. In SOM, Manhattan distance is chosen as discriminant function. The competence of proposed system is evaluated by precision and retrieval time. The proposed system is tested on benchmark databases. The results obtained in the experiments provide evidence that there is a considerable improvement with proposed system when compare to existing systems for image retrievals. Manhattan measure is used to estimate the likeness between the images.