Image Retrieval With Learning Low-level Features And Semantics

In general, there are mainly three popular image retrieval methods, text-based, content-basedand semantic-based. Text-based image retrieval employs key words to describe images and matchimages accurately based on the text description around them. Content-based image retrieval(CBIR) uses low-level visual features such as color, texture or shape to represent images. Uses cansubmit one or more query images to search for those which have the larger similarity to the queryexample visually in the database. So it is also called query by example (QBE). However the gapbetween low-level image features and high level semantics leads to the performance of CBIR faraway from users’expectations. Semantic-based image retrieval emerges as the times require.This paper summarized those key techniques in image retrieval and put forward a newmethod by improving existing methods. Experimental results demonstrated that our method had ahigher accuracy rate and recall rate of image retrieval. Details of our work are as follows:(1) In the process of image retrieval, a technique which automatically transformed textualqueries into visual representations was presented. First, we mined the relationships between textsand images and employed the relationships to construct a cross-media dictionary to automaticallytransform textual queries into visual ones. Then, we combine the retrieval results of textual andvisual query as our final results.(2) A key technique in image retrieval is image clustering. K-means algorithm and fuzzyc-means algorithm are two usual methods in image clustering. This paper we proposed improvedfuzzy c-means algorithm called weighting inter-class separating fuzzy clustering (WISFC)algorithm and employed it to content-based image retrieval. Firstly, the training data set wasclustered by k-means algorithm. Then the output cluster centroid was taken as the initial centroidof WISFC and the data set was clustered once again. When submit a query image, it could get itsmembership degree for each class using WISFC and the class which had the maximummembership value was what the query example belonged to. Afterwards we retrieve images whichhad the larger similarity to the query image only in the class it belonged to, which greatly reducedthe time complexity.(3) As a widely used classifier, decision tree can classify unknown data efficiently. This paperproposed a method of semantic-based image retrieval using decision tree learning. The methodincluded two innovation points. The first one was that it could retrieve images with high-levelsemantics. Firstly we extracted low-level features of the training images and obtained theirsemantic labels using an improved decision tree learning. So the continuous-valued features were converted into discrete ones avoiding complicated feature discretization problems. Then weretrieved those images whose semantic label was consistent with that of the query example. Thesecond innovation was that we conducted a process of feature fusion using CCA after color andtexture features extraction, in order to strengthen the correlations between the features.As the size of digital images is increasing rapidly, user requirement for image retrieval ishigher and higher. It has always been a hot pot to how to retrieve proper images from masses ofinformation.