Automatic Image Annotation Algorithm Based On Texture Compressed Measurement

Effectively management and retrieval of the massive growing images have become an urgent problem needed to be addressed. To manually annotation images with key words for image retrieval is time-consuming and person-exhausted. Automatic image annotation(AIA) technology provides a feasible solution. However, the low level visual features used by AIA are complicated for scene description. Thus based on Compressed Sensing(CS) observation, compressed texton is applied as a new feature used to describe image in this thesis. The Bags of Words(Bo W) model of texton is used to substitute the complicated contextual calculation, and then combining semantic co-occurrence with origin co-occurrence model for predicting key words. The image texton distribution’s similarity and words Estimation is used as to weight the origin PLSA-Words Model to improve the model’s ability to differentiate the parameter estimated.1. A method combined with co-occurrence model and key words co-occurrence information is proposed. To construct CS texton by CS measurement principle, and then using Bag of Words model to create the image’s histogram of CS texton, finally combining the co-occurrence model to build the map relationship of between images and semantic key words, finally combined with the key words co-occurrence information to predict key words. When the cluster numbers is 20, Compared with co-occurrence method based on HSV feature, the method based on CS texton get higher annotation accuracy with 3.2% and used less time with 45.7%, which show that the CS texton can use as an effective descriptive feature. A weighted coefficient is constructed by using the words’ co-occurrence and then combining the traditional co-occurrence to annotate image. Results show that the new method has better performance with 1.7%.2. A PLSA-Words improved model base on CS texton feature similarity with word weighted is proposed. When the numbers of key words is too less but the number of topics is many in the PLSA-Words model, learning parameters from the model, both the key words distribution of latent topics and the visual feature distribution of latent topics are weakly to be differentiate. Aiming at this problem, to define a new CS texton feature similarity and then combined with the word’s estimation from annotated image to create a weighted coefficient. This coefficient is word as a weighted for the result of PLSA-Words model to improve the model’s ability to differentiate the annotation words. Experiments show that when measurement rate is 0.4 and the cluster number is 45, the improved model get better result than the origin model which precision relatively improve 5.1%, recall 3.1% and the harmonic average relatively improve 4.9%.