A novel color image retrieval method based on texture and deep features

W Wei, W Wang, Y Yang, Y Wang - Multimedia Tools and Applications, 2022 - Springer
W Wei, W Wang, Y Yang, Y Wang
Multimedia Tools and Applications, 2022Springer
Enlightened by the currently prevalent great complementarity of traditional features and
deep features in image content expression, we propose a novel color image retrieval
method based on texture and deep features which represent the image with the combination
of HOG histogram, Hu invariable moments and deep features. Firstly, HOG features based
on local binary pattern are extracted from the image using the proposed second-order full-
directional derivative, which can express more gradient information through simplification …
Abstract
Enlightened by the currently prevalent great complementarity of traditional features and deep features in image content expression, we propose a novel color image retrieval method based on texture and deep features which represent the image with the combination of HOG histogram, Hu invariable moments and deep features. Firstly, HOG features based on local binary pattern are extracted from the image using the proposed second-order full-directional derivative, which can express more gradient information through simplification the expression method of full-directional derivative. Meanwhile, considering that the combination of information entropy and color can better represent the image content, we propose a novel quaternion expression method for color image and calculate its Hu moment features, which represent color image in a simple way by combination of color and texture information. Secondly, we extract deep features from an improved VGG network structure. Finally, the hybrid features combining HOG histogram, the new Hu moments and deep information are used to represent a color image and to perform retrieve task. In order to prove the effectiveness of our method, three common databases (INRIA Holidays, Oxford 5 K and UKB) are used to prove the proposed algorithm. Experimental results show that the proposed scheme has better performance on the basis of lower feature dimension.
Springer
Showing the best result for this search. See all results