Color Image Retrieval Using Statistically Compacted Features of DFT Transformed Color Images

Abstract

Feature extraction of images are crucial in image retrieval systems. Many approaches are stated and proved by researchers for image feature extraction and processing. Research is being done from low-level feature extraction toward high-level feature extraction. This paper discusses the feature extraction from the DFT transformed color images in multiple color planes. DFT image transform provides effective way to differentiate the image textures. For dimensionality reduction statistical parameters such as kurtosis, standard deviation, and variance are used for feature vector generation. Euclidian distance is used in the proposed approach. Four different types of feature vectors are created and tested for each image class. The images are retrieved based on the image pixel values of DFT phase information and DFT magnitude information of different color spaces like RGB, YIQ, HSV, and YCbCr similar to that of image class. Image retrieval performance of the proposed approach is compared for database of 1000 images of ten different categories. Precision of image retrieval is above 60% for all classes and more than 80% for some of the image classes.