Research On Active Hyperspectral Intelligent Perception Technology Based On Supercontinuum Laser

Hyperspectral imaging(HSI)technology is a combination of the two-dimensional imaging technology and the one-dimensional spectral detection technology.It can capture the spatial information and fine spectrum information of targets simultaneously,and thus has great information abundance.So far,HSI has received widespread attentions from all walks of life,such as the disclosure of disguised targets,medical disease diagnosis,precision agriculture,and criminal forensics.For these applications,HSI classification plays an important role.It can be said that the role of hyperspectral imaging technology is "sense",while the role of hyperspectral classification technology is "cognition",and the purpose of "sense" is to realize the "cognition" of the environment.Only by combining the two technologies can the "perception" of the targets be realized.However,traditional HSI technology is mostly passive perception technologies,which are heavily dependent on solar radiation.In order to overcome the shortcomings of the passive HSI technology,it is necessary to introduce a suitable wide-spectrum light source to replace the sun to illuminate the target scenes.In recent years,with the rapid development of supercontinuum laser technology and deep learning technology,active HSI intelligent perception technology based on supercontinuum laser illumination has become a new research and application trend.although relevant research has been carried out abroad and corresponding explorations have been made in the field of military applications,it is still in the "sense" stage as a whole and there is less research on "cognition".On the other hand,the rise of deep learning technology has avoid the workflow of manually extracting features from raw data,and has achieved better performance in many fields,making the development of intelligence a new technology trend.Combining the advantages of deep learning technology will help realize the intelligent perception of the HSI.In order to promote the further development of this technology,this paper successively explored the imaging results and intelligent classification results of the 90° off-axis parabolic mirror and Powell lens as the laser shaping devices,establishes the active HSI datasets under the illumination of supercontinuum laser,and proposes a series of HSI classification methods based on deep learning technology to solve the practical problems encountered in the application process.The main research content and the work of practical significance in this dissertation are as follows:(1)Traditional CNN mostly uses Soft Max classifiers and Soft Max cross-entropy loss functions to train the classification models,and the learned classification features are linearly distributed.Under this kind of feature representation,there is a phenomenon that the intra-class difference is sometimes even larger than the inter-class difference.To solve the problem,we propose an HSI classification method based on prototype learning mechanism and convolutional neural network(CNN).This method introduces the concept of prototype in the two-dimensional feature space,constructs a class prediction function based on the prototype learning mechanism and a loss function based on the prototype distance,and replaces the traditional Soft Max classifier and Soft Max cross-entropy loss function.This method can improve the intra-class compactness and inter-class dispersion in feature representation,and effectively improves the robustness and classification accuracy of the CNN.In addition,the method also realizes the two-dimensional visualization of the classification features,and the users can intuitively see the distribution of the features learned by the CNN in two-dimensional space.(2)In response to the application requirements of active HSI intelligent perception technology,we build an active HSI system using supercontinuum lasers firstly,which consist of supercontinuum laser,hyperspectral imager,laser shaping system,and electronically controlled turntable.To realize the perception of the targets,we propose an active HSI intelligent perception algorithm called Hybrid DN based on deep learning.The classification model includes two parallel branch structures,i.e.,one-dimensional spectral branch and two-dimensional spatial branch,which can extract the spectral features and spatial features of the targets at the same time.Besides,we successively explored the imaging results and intelligent classification results of the 90° off-axis parabolic mirror and Powell lens as the laser shaping devices.Through experiments,we find that Powell lens can effectively avoid the phenomenon of "same object shows different spectral characteristics " caused by the inhomogeneous distribution of the light intensity and spectra in the laser spot,which is more practicality.(3)When the active HSI intelligent perception technology perceives the target environment,it often encounters the classes which did not appear during training.Faced with these unknown classes,traditional HSI classification methods can only classify them into known classes.To solve the problem,we propose an HSI open-set classification method based on Euclidean distance and deep learning.The classification method first introduces the concept of fixed class center to represent the distribution center of each known class in the label space;then constructs the Euclidean distance-based class prediction functions and loss functions to make the CNN learn the feature vectors which have higher intra-class similarity and more obvious inter-class dissimilarity.Finally,boxplot and Weibull model are used to constrains the decision boundary of each known class,so that the model can realize the unknown rejection without sacrificing the classification accuracy of known classes dramatically.(4)In realistic applications,the lighting condition,atmospheric condition and target attitude may change,which will cause the same kind of targets to show different data distributions in the source domain and the target domain,making the traditional classification method based on the assumption of distribution consistency unable to deal with the problem of "cross-domain classification".In response to this challenge,we use the Powell lens-based active HSI imaging system to collect and produce active HSI datasets of the source and target domains firstly.And then we propose an active HSI classification method based on domain adaptation.By constructing feature extractor,label classifier,and domain classifier,the classification method can learn the class-discrimination and domain-invariance features,and thus achieves the classification of target domain with source domain data training only.The method enhances the robustness of the active HSI classification method to environmental changes and makes active HSI intelligence perception technology own the cross-domain classification capability.