Representative Image Selection In Large-scale Image Dataset

Representative image selection is a technology of image summarization toselect the most representative or typical images from the web albums. According tothe query keywords by users, it returns the image collection containing a variety ofsemantic theme related for users to quickly browse and obtain information aboutthe query. Representative image selection technology is used widely in recent years,it becomes a hot topic in many fields such as image processing and analysis, patternrecognition, artificial intelligence and so on. In this paper, the main work andinnovations are as follows:1. The overview of the conception and basic procedure of representative imageselection is given, including the image feature extraction, the image clusteringbased on the features, image cluster ranking and typical image selecting. This paperanalyzes some disadvantages of the existing methods.2. A viable method of representative image selection is proposed for theproblem of the search results show a lack of diversity in semantic theme by thekeyword in the traditional image retrieval system. The representative images aredefined as those with diverse contents in the semantic meaning. First, mutualnearest neighbor consistency is used to adjust the similarity between images whichis as the input for the AP clustering. Then representative clusters are selected basedon the cluster ranking and finally the images of the cluster center fromrepresentative clusters are as the summary of the image dataset. The results showedthat the performance of our method is better than the K-means based method andthe Greedy K-means based method. The selected images can summarize the contentof the original image dataset intuitively and effectively, and they are diverse insemantic meaning as well.3. A kind of adaptive weights allocation method based on particle swarmoptimization algorithm is given for the problem of the weights assignment betweendifferent features in the process of representative image selection. The problem ofweights assignment can be converted to the optimization of the objective function.The particle swarm optimization algorithm is used to optimize the total quality Q ofthe clustering result which implements an adaptive weights allocation methodbased on the character of the image dataset itself to distribute the weight of each feature, and thus produces more effective representation of the similarity measure.The results showed that the performance of our method is better than the K-meansbased method and the Greedy K-means based method. Compared with KMNC-APbased method, it saves a lot of manpower and time.