Improving performance of content based image retrieval system with color features
Aleš Hladnik, Ante Poljičak · University of Zagreb University Computing Centre (SRCE) · 2016
Content based image retrieval (CBIR) encompasses a variety of techniques with a goal to solve the problem of searching for digital images in a large database by their visual content.Applications where the retrieval of similar images plays a crucial role include personal photo and art collections, medical imaging, multimedia publications and video surveillance.Main objective of our study was to try to improve the performance of the query-by-example image retrieval system based on texture features -Gabor wavelet and wavelet transform -by augmenting it with color information about the images, in particular color histogram, color autocorrelogram and color moments.Wang image database comprising 1000 natural color images grouped into 10 categories with 100 images was used for testing individual algorithms.Each image in the database served as a query image and the retrieval performance was evaluated by means of the precision and recall.The number of retrieved images ranged from 10 to 80.The best CBIR performance was obtained when implementing a combination of all 190 texture-and color features.Only slightly worse were the average precision and recall for the texture-and color histogram-based system.This result was somewhat surprising, since color histogram features provide no color spatial information.We observed a 23% increase in average precision when comparing the system containing a combination of texture-and all color features with the one consisting of exclusively texture descriptors when using Euclidean distance measure and 20 retrieved images.Addition of the color autocorrelogram features to the texture descriptors had virtually no effect on the performance, while only minor improvement was detected when adding first two color moments -the mean and the standard deviation.Similar to what was found in the previous studies with the same image database, average precision was very high in case of dinosaurs and flowers and very low with beach, food, monuments and mountains images.