Multispectral Image Matching Based On CNN

Image matching is a basic technology in the field of computer vision,which is widely used in tasks such as image registration,image stitching,and image fusion.In the multispectral image matching task,there are not only significant grayscale differences and illumination differences between images,but also nonlinear intensity differences caused by imaging with different sensors.The hand-designed feature extraction method has limited expressive ability in multi-spectral matching tasks,and cannot accurately extract the common features between images.Therefore,this paper uses the deep learning method,through the powerful feature extraction ability and learning ability of the convolutional neural network.Multi-spectral images are used for feature extraction,and a feature descriptor with excellent expressive ability and discriminating ability is learned.The main research contents of this paper are as follows:1.Design a feature descriptor based on attention mechanism and residual network.Taking advantage of the powerful feature expression ability of deep convolutional network,a feature extraction network based on residual network is designed first,and then considering that although there are significant differences between multi-spectral images,there are often high visual and Similar samples,the network is easy to make wrong judgments.The attention mechanism is introduced into the residual network,focusing on important features,ignoring irrelevant features,and improving the discriminative ability of the network.The experimental results show that,compared with the hand-designed feature descriptors,the method in this paper can well overcome the nonlinear intensity difference between the multi-spectral images in the multispectral matching task,and can accurately extract the multi-spectral images.The common features between the two,the method is effective for multi-spectral image matching.2.Design a feature descriptor based on multi-scale features and residual shrinkage network.Because different scales have different key information,and the data in nature has inherent redundant information and noise interference,and the two images in the multispectral image are quite different,one of the images has too clear texture features or noise.Interference,for multispectral image matching tasks,will adversely affect the matching.Based on the above considerations,it is proposed to combine pyramid convolution and residual shrinking network,use pyramid convolution to extract multiscale information with different kernels,and use the extracted multi-scale feature information as the input of residual shrinking network.The soft thresholding operation eliminates redundant information and noise interference,and further improves the feature extraction and discrimination capabilities of the network.The experimental results show that this method performs better than the previous methods in the multispectral image matching problem.