Unsupervised learning method for shape matching templates based on gradient saliency estimation

In order to improve the accuracy and robustness of the shape-based matching algorithm in practical industrial vision applications, we proposed an unsupervised learning algorithm, which under a given initial shape, can capture more potential shape features of the target in the training set through a comprehensive estimation of both the saliency of gradient orientation and gradient amplitude, consequently, learn a better template for matching. The experiment shows that when the number of valid training images reaches about 30, the point number difference between the learned shape and the ideal shape does not exceed 8%. In a comparative experiment on matching precision and recall of four different shapes, we found that the data curves of the learned shape were almost identical to those of the ideal shape, which proves that the algorithm has effective self-learning ability and thus, can improve the performance of shape-based template matching.