Research On Deep Learning-Based Image Matching And Fusion Algorithms

Image matching and fusion are important research contents in image processing,which can make images more informative by synthesizing multiple images.Deep learning has been widely used in many fields and has been very successful,especially in image processing.Deep learning-based image matching and fusion algorithms are discussed in this thesis.The theory of image matching and image fusion,as well as the theory of convolutional neural network and deconvolutional networks are introduced.For the problem of low accuracy of feature points matching,a convolutional neural network-based image matching algorithm is improved.A novel deconvolutional networks-based image fusion algorithm is proposed.The main contents of the thesis are as follows:An improved convolutional neural network-based image matching algorithm is proposed.Firstly,a convolutional neural network is trained,and then it is used to obtain the descriptors of the image features.Finally,the feature-based image matching is implemented according to these descriptors.In this algorithm,the seeds of training samples that used to train the convolutional neural network are image features extracted directly from the images that need to be matched,and smaller seeds are used.Experimental results show that,the accuracy of image feature points matching is improved with the improved convolutional neural network-based image matching algorithm proposed in this thesis compared with the original method,which has more advantage in image matching.A deconvolutional networks-based image fusion algorithm is proposed.In the proposed algorithm,a deconvolutional networks is trained firstly.After that,the images to be fused are applied to the trained network respectively,and their feature maps are obtained by inferencing.Then,the feature maps of the two images are fused according to designed fusion rules to obtain the fused feature maps.Finally,the fused image is obtained with the fused feature maps and the filters of the deconvolutional networks.Experimental results show that,the multi-focus images,medical images and remote sensing images can be effectively fused with the deconvolutional networks-based image fusion algorithm proposed.The performance of the improved convolutional neural network-based image matching algorithm proposed in this paper is better than the original method.Meanwhile,the deconvolutional networks-based image fusion algorithm proposed provides a new fusion idea in addition to the existing methods.