The Method Research Of Image Detection And Registration

Image detection and registration are the basis of image analysis and pattern recognition, related to many techniques such as image processing, signal processing, automatic control and so on. The kernel of image detection and registration lie in the constructing suitable image analysis algorithm. Some method and application of image detection and registration are introduced in the paper, in instance of face detection and medical image registration.There are mainly two parts in the paper. The first part is face detection and the second part is medical image registration.Based on the lighting condition compensation and the nonlinear color transformation technique which are proposed by Rein-Lien Hsu, Mohamed Abdel-Mottaleb and Anil K Jain(denoted as Hsu's method simply), an improved algorithm of face detection is proposed for color images. Three improved aspects are presented as follows:Firstly, the original images are preprocessed by the lighting compensation (color bias can be corrected by using the technique in the original images.) and nonlinear color transformation technique. In order to reduce the computational complexity of Hsu's method, the parametric ellipse mode is replaced by the Gaussian distribution model, which leads to the fast face detection without affecting on the detection accuracy;Secondly, a method of selecting the skin threshold is suggested based on the combination of self-adaptive process and histogram which is more accurate;Thirdly, in order to extract the face areas accurately, the candidate areas, which are got by skin tone detection and coarse detection, are mapped to the gray images of the original images where the sub-images are extracted. Then, according to the intensity and geometry information of eyes which are got by statistical method, the eyes detection are in the previously extracted sub-images. Finally, the sub-images corresponding to the face are mapped to the original images.Experimental results demonstrate successful face detection over a wide range of facial variations of lighting condition, skin color and 3D poses. The proposed method is much faster than Hsu's method.In order to improve the accuracy of image registration and reduce the computation complexity, an image registration algorithm based on characteristic points is proposed. First of all, the idea of multi-scale and fuzzy theory are used in the characteristic points detection and the characteristic points of images are got; then the suited points are got; finally, the registration images are got using thin-plate spline function. Experimental results demonstrate that the proposed algorithm has high accuracy.