Application And Realization On Localization Of An Object By Using Hidden Markov Model

Localization of an object in an image is always an interesting and challenging research task in the fields of image processing and computer vision. The technology of object localization has a shining perspective and plays an important role in the area of medicine, military and computer science, such as computer-aided diagnose, precise targeting weapons and human-machine-communication systems. In this paper we apply 1-dimensional Hidden Markov Model to localization of a linear structure object (spine) in medical images. A novel method of overlap feature extraction technology is be proposed to pick-up feature vector which has rich information. With the analysis and understanding of the classical Viterbi algorithm and the consideration of actual condition, a best search localization algorithm derived from Viterbi algorithm is presented for determining a feature sequence in a test image, from which the position of a linear structure object could be obtained. Based on the best search localization algorithm , we propose a under-control best search localization algorithm which can balance the time and the effect of localization. To verify the validity of the novel approach, many localization experiments have been carried out with actual X-ray medical images of human spines as the experimental sets. The performance of localization is improved due to use of overlap feature extraction method and the under-control best search localization algorithm. Through experiments, we compare the effect for different size of sampling windows and their overlap on localization performance and the effect for different number of control step on localization performance. According to the experimental results, the effect of different parameterisations of the model on the localization performance is investigated and rules are drawn for establishing HMM prototype better. Finally, some thoughts for improving localization performance are proposed, which are worth in promoting the approach to object localization based on HMM for practicality.