Research Of Image Annotation Base On Regional Segmentation

Image annotation modeled by the image characteristics, and then use the generated model for the image content, for reflecting the image content accurately and fully. It has broad application prospects, including medical image annotation, the establishment of the digital library, robot visual scene understanding, retrieval and management of digital photos, etc. In this paper, use region feature extraction, super pixel bag of words model and multi-scale sparse coding research for the super pixel area and merging region based on image segmentation, to improve the segmentation performance and labeling accuracy, specific research work is as follows:(1)In view of the over-segmentation problem of super pixel segmentation algorithm segmentation result, put forward a image segmentation algorithm based on the similarity and statistical consistency check of super pixel, for improving the over-segmentation problem.The algorithm use the super-pixel regional segmentation result as the basic unit for region merging, using the basic unit of color similarity, the size of the space distance, and statistical characteristics of consistency, for solving the over-segmentation problem.Simulation results show that the algorithm can solve the problem of super-pixel over-segmentation, and the index PRI and VoI for evaluating segmentation algorithms are improved.(2)In view of the single label problem of the bag of words model and SVM classifier,put forward a image annotation algorithm, which bases on super-pixel bag of words model and SVM classification.The algorithm firstly uses super-pixel segmentation results as regional basic unit of bag of words model, and the use the bag of word model and SVM classifier generated by the image databases for classifying super pixel areas, through the statistics classification results of the super pixel areas for multiple labels annotation. Simulation results show that the algorithm can improve the single label problem of bag of words model and SVM classifier, and classification accuracy improved.(3)In view of the fixed grid problem of the pyramid and the problem for ignoring the spatial location of vector quantization code, put forward a image annotation algorithm,which based on local constraint sparse coding in multi-scale space. The algorithm use super-pixel segmentation and region merging, which based on similarity and statistical standards, to divide the multi-scale space, for keeping the target area and making full use of image space context information. In order to improve the coding performance, use the similarity in the base vector weighted for coding image feature, to position constraint and sparse expression of image characteristics. Simulation results show that the algorithm not only can take advantage of the multi-scale information, but also can improve coding performance for improve the accuracy of the annotation results.