Deep Feature Description And Robust Matching Of Multispectral Images

Image feature matching is a procedure defined as extracting features from two images taken from different perspectives and determining whether the images contain the same region by the features.As a relatively fundamental problem in computer vision,image matching technology has important applications in the fields of material analysis,medical imaging,and image restoration,etc.Traditional hand-crafted feature extraction algorithms have achieved many good results processing the images captured under the same spectrum,but they do not perform well for multispectral image matching problem.Deep learning,based on deep neural networks,achieved good results in computer vision fields such as object detection,image classification,and image restoration,etc.To tackle with multispectral image matching problem,the algorithms from this paper come with deep description of feature information from multispectral images.Furthermore,the algorithms from this paper can match multispectral images accurately under texture losing or other problems caused by different imaging spectrums.The algorithms are robust towards trans formation.The paper worked out a technology including data processing and neural network model designing for multispectral image matching problem.For details,the paper finished work shown as follows:(1)This paper captured image dataset with infrared thermal imager and digital cameras.And this paper produced the required datasets by pre-processing the image data.(2)Furthermore,this paper proposed a method using deep neural network to match the multispectral images.With experiments on over 10k multispectral image pairs,state-of-the-art deep learning matching solution achieves 78%accuracy,while algorithms in this paper perform more effectively(90%Accuracy)on multispectral image matching problem.The neural network model trained on dataset by this paper worked out well on RGB-NIR dataset,proving the effectiveness of the algorithms.(3)This paper performed augmentation on original multispectral image dataset to simulate several transformations.Then the paper introduced active rotation filter,spatial pyramid pooling,and multi-prediction module to the neural network.State-of-the-art solution achieve 68%matching results,while the model from this paper keeps high effective on the dataset with transformations(85%Accuracy),which proves the technology in this paper is robust for multispectral image matching.