Abstract:

For the image template matching task in which the image to be matched has spatial transformations such as rotation transformation, scale transformation, and translation, the existing algorithms are time-consuming and have low accuracy. In this paper, an image matching algorithm with high accuracy is proposed. Firstly, oriented fast feature detection algorithm is used to find the feature points according to the pixel difference between the center point and the neighboring point in order to detect the feature points. With these feature points taken as the center, patches of a certain size are cropped according to the rotation angle calculated by the fast feauture detection algorithm. These image patches are fed into the feature descriptor extraction network of the spatial transform attention module. Finally, K-nearest neighbor (KNN) algorithm is used to calculate the matched features in the feature descriptors of the two images to be matched. The spatial transformation attention module is introduced into the feature descriptor extraction network. The network focuses on learning spatial information during training, so the entire algorithm improves the accuracy of image matching tasks with large spatial changes. In terms of matching time, the matching algorithm proposed in this paper is second only to the method that uses fast feauture detection algorithm algorithmfor both detection and matching tasks. In terms of matching accuracy, the matching accuracy of the algorithm in this paper is far superior to other algorithms compared in the experiment.

Key words: computer vision, image matching, space transformation invariance, convolutional neural network (CNN), attention mechanism