Hyperspectral Image Traget Detection Algorithm Based On Deep Learning

Target detection in hyperspectral image has always been an important research direction in hyperspectral image processing field.Since the 1980 s,target detection methods based on hyperspectral images have been constantly innovating and a series of research results have been obtained.However,because the high-light image data is different from the ordinary two-dimensional image,many target algorithms for ordinary images can not be directly applied to high-light images.In spectral images,the traditional target detection algorithms in hyperspectral images often have the problems of low adaptability,low detection efficiency and slow detection speed.Hyperspectral image is a typical three-dimensional data block,which has the characteristics of high latitude,large amount of data and complex information.Traditional target detection algorithm is difficult to fully mine the spectral information and spatial information of hyperspectral image.In recent years,the theory and methods of deep learning have made great achievements in many fields.However,in the field of hyperspectral image target detection,because some targets contain fewer pixels,insufficient samples and complex data structures,deep learning methods have not been widely used.Deep learning is an important branch of machine learning.Its powerful data analysis ability and feature information mining ability make data processing more efficient.It can extract multi-level features from the bottom to the top of the original input data,and ultimately obtain high-level feature information,which greatly improves the speed and accuracy of target detection.The main contents and achievements of this paper are as follows:Firstly,according to the complex data characteristics of hyperspectral images,hyperspectral data are preprocessed.Hyperspectral image contains not only twodimensional spatial information but also spectral information.It is a typical three-dimensional tensor data.However,in the process of image acquisition,it is often affected by atmospheric noise and spectral interferences,which greatly reduces the amount of information in some bands of the original image.Therefore,this paper uses Frobenius norm method to screen the original hyperspectral data bands,retain the bands rich in target information,and normalize the processed hyperspectral image data to lay the foundation for subsequent target detection.Secondly,aiming at the multi-dimensional data characteristics of hyperspectral images,this paper introduces and establishes the CNN(Convolution Neural Network)target detection model and DBN(Deep Belief Network)target detection model respectively by using the spectral information,spatial information and spatial information of hyperspectral images.The experimental results show that both of the two deep learning target detection models can achieve target detection in hyperspectral images very well,and the results of hyperspectral spatial information model are better than that of spatial information and spectral information.Thirdly,hyperspectral image target detection based on CNN and DBN can effectively detect the target of hyperspectral image,but the target information has not been accurately framed and recognized.Faster-RCNN model framework under indepth learning has achieved good detection results in the field of target detection,but because of the shortage of data characteristics and target samples of hyperspectral image,the model is still available.It has not been successfully applied to target detection and recognition in hyperspectral images,so a Fast-RCNN hyperspectral image target detection algorithm based on composite data sets is proposed in this paper.In this detection model,firstly,the original hyperspectral image data is combined with the image set of Google Earth to make a composite data sample set according to the three-dimensional characteristics of hyperspectral image,and then the target detection model of Faster-RCNN is built by adjusting the detection network parameters according to the composite data set.Finally,the test data are substituted into the detection model after training,and the target is delineated and recognized accurately.Finally,in the research of classification and target detection of hyperspectral images,the shadow areas of hyperspectral images are often directly ignored or simply classified into a group,which does not make full use of all the information of hyperspectral images.In order to solve this problem,this paper proposes a method of target detection in shadow area based on hyperspectral image target classification.This method further realizes target information detection in shadow area of hyperspectral image on the basis of classification-based hyperspectral image target detection method,so as to further excavate the rich information of hyperspectral image.The experimental results show that the method can be used to detect substances in the shadow of hyperspectral images,and has broad application prospects in the detection of hyperspectral targets and the processing of shadow areas of hyperspectral images.