Integrating Pixel Cluster Indexing, Histogram Intersection And Discrete Wavelet Transform Methods For Color Images Content Based Image Retrieval System
Pijush Kanti Bhattacharjee · International Journal of Computer and Electrical Engineering · 2010
This paper proposes a new intelligent content based image retrieval (CBIR) system for integrate, disseminate, retrieval, visualize and correlate color images.Three different methods are integrated for searching out the similar images of the query image from a large image databases like consisting of 1798 numbers different images.Then a semantic approach is implemented to minimize the image database i.e. the images are grouped or subdivided.The color images are first converted hue (H), saturation (S), value (V) from R, G, B values, because the HSV model is more authentic than the RGB model.The first method "Pixel Cluster Indexing of the color images" is a completely innovative method.In this, certain range of pixel values say 100-255 either from gray level for monochromatic image or color image in RGB or HSV values are selected and is converted to an equivalent image of the original image i.e. in the equivalent image, the pixel value below than 100 becomes 0 and all other pixels remain same position in the equivalent image as it is in the original image.Hence the equivalent image represents the original image.By this method, the query and the database images are converted to their equivalent images.The original image can be taken in smaller size either by cropping i.e. taking some portion of the image or reducing size by bilinear or nearest neighbor interpolation method for minimizing mathematical complexity.Then the statistical parameters like std, mad of the equivalent images are computed and are combined std and mad in normalized form.Thereafter, this combined value std and mad of the equivalent images derived from the query and database images respectively are compared by Euclidean distance method and the nearest matching images to the query image are ranked accordingly.In the second method, histogram of colors in HSV model is computed for the query as well as the database images respectively.Then this histogram of each H, S, V separately between the query and the database images are compared.The minimum numbers of pixel values present in both the comparing histograms is taken for each position and are summed for the whole comparing image sizes.The total number of pixel values present is normalized and the best matching images for the query image are sorted out from a large image database of 1798 images.