Author + information
Medical image fusion technology is an image post-processing technology that can construct an image using anatomical information and functional information from various imaging sources. The image fusion quality depends on the matching accuracy of the image matching technology. The scale-invariant feature transform (SIFT) algorithm is a matching method based on local feature point detection. It is robust against interference from the external environment, such as change in size, angle, mapping, and noise. However, mismatching occasionally occurs when the SIFT algorithm is used to match images. This study improves the SIFT algorithm to reduce the number of mismatches in the image matching process.
First, the graph context of the image was improved. For this purpose, the shape point set of the image was taken as an elliptic inertia, and then, it was normalized to obtain the affine invariant.
Second, the interchangeability of the new graphics translation context with translation, rotation, and scale transformation was analyzed to determine whether it satisfied the invariant property of the SIFT algorithm. The similarity in the improved graphics context process was calculated by using different norms, including the row norm, column norm, 2-norm, L-norm, and F-norm, and the vector angle of the feature point. The threshold values T0, T1, T2, and T3 were respectively one-half, the same, two times, and five times the average of all similarities in a subset.
The appropriate norm and threshold were selected to detect the accuracy of image feature matching by the optimized SIFT algorithm.
Finally, the image matching result obtained using the optimized SIFT algorithm was compared with that obtained using the conventional SIFT algorithm. For this purpose, Matlab R2014a, Microsoft Visual Studio 2010, and OpenCV 2.4.9 software were used. A human brain CT with 277×245 pixel resolution was used to perform image matching. The matching effects of rotation, scale, and both rotation and scale transformations as well as noise were observed.
2-norm showed the best matching accuracy of 95.68% and shortest required time of 4.72s. Furthermore, the matching accuracy obtained using T1 was the highest. Therefore, this study chose 2-norm and the threshold for matching. The optimized SIFT algorithm could improve the matching accuracy even though the number of matching points decreased owing to rotation, scale, and both rotation and scale transformations, and it could reduce the mismatch. It was also robust against rotation, scale, and both rotation and scale transformations. The matching accuracy decreased with the severity of noise when using both the optimized and the conventional SIFT algorithms. With increased noise pollution, the number of feature points also increased. The greater the noise, the lower was the matching accuracy. However, the optimized SIFT algorithm afforded better matching and lesser matching errors compared to the conventional one.
The optimized SIFT method reduces mismatching and is robust to various external transformations of the image, such as rotation, scale, and both rotation and scale transformations.