End-to-End Blurry Template Matching Method Based on Siamese Networks

Template matching is a classic problem in computer vision, but most matching methods simply assume the ideal inputs without real-world degradation, such as blur. For blurry template matching, existing methods either have low accuracy or low time efficiency, this paper proposes an end-to-end network BMNet based on siamese networks. When the blurry template image and the clear reference image are input in pairs, the siamese networks are used to extract features respectively, then the two feature maps are transformed and sent to two branches for the classification of the target scene/background and the regression of the coordinates’ offsets. Through the multi-task learning, a trained BMNet can accurately output the coordinates of the template image in the reference image. Extensive experiments demonstrate that our method significantly outperforms state-of-the art on accuracy, speed and robustness.