Image Retrieval Based On New Color Similarity And Adaptive Neighbor

Content-based image retrieval is a new kind of image retrieval technologies. Different from traditional tag-based image retrieval, it utilizes some image as query to obtain the relevant images. Therefore, on one hand, this technology saves labor cost for annotation; on the other hand, it diversifies image retrieval mechanism. Feature extraction and similarity measurement are two"bases for content-based image retrieval, in which image content is described by features produced by feature extraction module whereas similarity measurement module calculates the similarity between images based on these feature. Much work has been done on these two fields. However, traditional feature extraction only considers the statistical characteristic of color or texture content in image and ignores the spatial distribution information of these features so that their distinguish ability is very limited. Similarity measurements are not designed based on human perception of images which has the disadvantage of inaccurate similarity value and high time complexity. This dissertation aims to conduct the research on the above mentioned problems and propose new methods on feature extraction and similarity measurement. The main innovations are described as follows:1. A new pulse-coupled neural network (PCNN) based feature extraction method is proposed. To overcome the disadvantages of lack of spatial information in existing feature extraction methods, we introduce the unit-linking PCNN, a simplified model of PCNN, to the feature extraction. Because of its emitting mechanism and propagation effect, this model not only takes into account the color statistical property but also incorporates the spatial distribution information. We use this model to extract both color and texture features and combine them with distribution entropy to further describe the color distribution. Experimental results show that our method can improve the retrieval precision and has less storage cost.2. Enlightened by the human perception of images, we propose three conditions to design a similarity measurement and derived a new color-based similarity measurement. The new method endows colors of different proportions with different priorities for the match with colors from another image. Meanwhile, it also considers similarity both between colors and proportions and adopted weights fashion to decide the final similarity between images. Because this method distinguish the contributions of different colors, experimental results demonstrates that with a few color features extracted the new method can obtain higher retrieval precision and lower time complexity compared to other traditional similarity measurements.3. To further improve the effect of image retrieval, we introduce the transductive learning into image retrieval and propose a new adaptive nearest-neighbor graph method for image retrieval. This method utilizes a new graph structure to organize the images in the database and both global and regional features are adopted to determine the weights in the graph. Through this graph, we utilize the transductive learning to do the image retrieval and also incorporate the relevance feedback strategies including short-term and long-term ones and an extension to queries outside database to make our method more efficient. Experimental results show that our method has vast improvement on retrieval performance and relevance feedback compared with other state-of-the-art algorithms.