Abstract
The progress of image search engines still proceeds, but there are some challenges yet in complex queries. In this paper, we present a new semantic image search system, which is capable of multiple object retrieval using only visual content of the images. We have used the state-of-the-art image processing methods prior to the search, such as Fisher-vector and C-SVC classifier, in order to semantically classify images containing multiple objects. The results of this offline classification are stored for the latter search task. We have elaborated more search methods for combining the results of binary classifiers of objects in images. Our search methods use confidence values of object classifiers and after the evaluation, the best method is selected for thorough analysis. Our solution is compared with the famous web images search engines (Google, Bing and Flickr), and there is a comparison of their Mean Average Precision (MAP) values. It can be concluded that our system reaches the benchmark; moreover, in most cases our method outperforms the others, especially in the cases of queries with many objects.
Deep Root Memory Optimized Indexing Methodology for Image Search Engines
