Neural Network Image Matching Based on Variable Weight Smooth Constraint

This paper proposes a global image-matching algorithm based on variable weight smooth constraint. The traditional image-matching algorithm can be divided into area-based, feature-based and relational image-matching algorithm. The area-based image-matching algorithm could obtain accurate and dense disparity surface, but this kind of algorithm depends on the texture information of the images, so it always obtains poor results in Poor-texture areas of images. However the feature-based image-matching algorithm can obtain good results. So we should integrate both algorithms in the practical image-matching works. In the algorithm we integrate the area-based and feature-based image-matching algorithms by use of a well-defined objection function, by minimizing this function we get global image-matching algorithm based on variable weight smooth constraint. The algorithm evaluates smooth constraint and the discontinuance on linear feature. In poor texture area or texture area of the image, the weight of smooth constraint can vary based on the contents of image. In poor-textured area of the image, we impose stronger smooth constraint,and get the smooth disparity surface. Near the texture area, the strength of smooth constraint varies on the strength and orientation of the linear features. So it can correctly use the gray and feature information of image. Later we also give a direct mapping method between global image matching and Hopfield neural network. By comparing the energy function of Hopfield network with the so-called global compatible objection function,we get the bias input and the weight between neurons. So the image matching process can be completed efficiently, accurately and robustly because of the parallel processing of the Hopfield neural network. At the last portion of this paper,we provide some experimental results obtained by processing the practical image data including the images of rough terrain areas and the large-scale urban areas. The results show that the algorithm is effective.