| With the development of modern medical technology, a variety of imaging techniques and the information science and technology have been applied to the medical treatment process, in which the Computer-assisted surgery technology(CAS)is been widely used in clinical diagnosis and treatment. In order to avoid suffering from the diagnosis and show the changes of the lesions in the target area, CAS has concentrated a variety of technology like medical imaging technology, precision machinery technology, computer science and technology, which has made surgeons clinical operation much more convenient and improved the efficiency of diagnosis and treatment. In the process of surgery navigation, as an important part of CAS, image registration technology has been one of the national research focus and this technology is also playing an important supporting role in the key link of the surgery navigation such as the surface reconstruction of the surgery scene, real-time accurate tracking of surgical instruments and the lesion area, improve the visual effect of traditional surgery navigation and so on. So how to improve the matching speed and accuracy is becoming an important problem in the CAS technology.Taking laparoscopic monocular visual image sequences as the research object,this paper puts forward a kind of dimension reduction feature describing vector based on Radon transform and proposes a weighted descriptors similarity measure algorithm based on the Euclidean distance, which is aiming at reducing matching time and improving registration accuracy. The main research content of the article are as follows:1. We analyze the theory basis and the common methods of image registration techniques in detail and describes the registration process such as feature extraction,similarity measure criteria in feature matching process, the searching space and strategy, removing false matching points, setting up space transformation model,sampling interpolation arithmetic and so on. Then we introduce some classic feature extraction operators like Moravec, Harris, SUSAN, SIFT, and also make contrastive analysis based on the existing experimental data. At last, we concluded that the SIFT operator performed better with scale 、 illumination and rotation transform, which it also had good robustness and resistance to the noise.2. Aiming at the problem that the traditional SIFT algorithm’s matching time istoo long, we put forward a dimension reduction feature descriptor based on Radon transform. The descriptor first determines the integral direction based on the main direction of the gradients in feature point neighborhood area, then equivalent substitutes the change of Gaussian blur weights in feature point neighborhood area by the proper integral function. In the end, it builds the new feature vector instead of primary 128-d descriptor by the integration along 30 different directions. This processing procedure obviously reduces the computational complexity of the algorithm.3. Aiming at the problem that the matching accuracy of the primary algorithm remains to be improved, we come up with a new descriptor similarity measure method which is based on Euclidean distance and feature information and gray level statistics information. This method first establishes 2d image entropy to represent the spatial characteristics of gray level distribution in the feature points neighborhood, then provides the relative Euclidean distance formula between feature vectors based on Radon transform as well as determines the weights of feature and gray statistics information by the optimization of distance parameter ?. After the above processing,we construct a new feature vector distance measure method and at the final step, we determine the matching point pairs by using the BBF tree strategy as well as bilateral matching strategy. The article tests our new registration algorithm on the test images that varies on illumination、scale and rotation as well as makes the comparison with the original algorithm and other several commonly used algorithms, We also make analysis between the original algorithm and our improved algorithm based on the test images which in the condition of translation and shade. The experiments result indicates that on the condition of these transformation, out new registration algorithm based on SIFT performs better on speed and accuracy of the registration. |
