Research On Blurry Image Matching For Imaging Guidance

Image matching is a classic problem in computer vision and can be widely used in vision-based navigation systems.Image matching for imaging guidance mainly studies how to calculate the position coordinates of small-size real-time images from large-size reference images.The complex imaging environment often leads to a certain degree of blur degradation in real-time images,which brings challenges to image matching.In this paper,three blurry image matching algorithms are proposed for the problem of blurry image matching.Different from the traditional blurry image matching algorithm that first restores and then matches,the joint image restoration and matching algorithms not only use the restoration results to improve the matching performance,but also uses the matching results as a priori information to improve the restoration quality,let them promote each other during the alternate iteration.Among which the algorithm based on distance weighted sparse representation JRM-DSR has good matching performance,but having the problem of low efficiency.Therefore,this paper proposes an joint image restoration and matching algorithm JRM-HSR based on hierarchical sparse representation,which uses PCA and K-means to cluster the dictionary,and generates two-layer dictionaries that are much smaller in size than the original and responsible for the coarse matching and the fine matching respectively.JRM-HSR effectively solves the time efficiency problem while ensuring excellent performance.When the blur degree of the real-time image is high,the matching performance of JRM-HSR has a serious decline.Therefore,this paper proposes an image matching algorithm BINet based on blur invariant feature metric network,which obtains a network that can extract blur invariant feature through metric learning.Using this feature for matching,BINet's matching performance greatly leads the traditional algorithms under a large degree of blur.Although BINet has good anti-blurry robustness,it needs to use CNN to extract features from many highly overlapping candidate images,which has the disadvantage of redundant calculation.In response to this problem,this paper proposes a siamese networks-based blurry image matching network BMNet,which introduces the idea of??cross-correlation information and target detection.Through multi-task learning,BMNet does not need to prepare the reference image in advance and achieves end-to-end matching in higher level of matching performance and efficiency.For the problem of blurry image matching,this paper proposes three different blurry image matching algorithms,which show excellent matching performance,algorithm efficiency and algorithm robustness,and have important practical significance.