Image annotation based on bi-coded chromosome genetic algorithm for feature selection

In image annotation system, performance is highly desired. We use a bi-coded chromosome-based genetic algorithm to optimize the weights of multimedia content description interface (MPEG-7) feature descriptors and select optimal feature descriptor subset simultaneously. Two genetic codes are used: a real code represents the weights corresponding to MPEG-7 descriptors; a binary one denotes the presence or absence of feature descriptors in the optimal descriptor subset. The genetic algorithm fitness function takes into account support vector machinepsilas classification accuracy and the number of selected feature descriptors. The result of experiments over 2000 classified Corel images shows that the approach selects 4 of 25 MPEG-7 feature descriptors as optimal feature descriptor subset as well as corresponding optimized weights. With the selected optimal feature descriptor subset and weights, the accuracy and efficiency of image annotation system can be improved.