Research On Polarization Hyperspectral Image Processing And Classification Algorithm Based On Generative Countermeasure Network

With the continuous improvement of polarization hyperspectral imaging system,this kind of image which integrates polarization information,spectroscopy and spatial information is developing more and more rapidly in various fields,especially in the field of medicine and bionics today,polarizing hyperspectral is widely used in cancer cell identification,lesion tissue positioning and combat target detection in complex land and underwater environments.However,due to the existence of polarization hyperspectral data with fewer samples,insufficient prior knowledge,expensive equipment,high dimensionality cannot make full use of the problems,seriously hinder the polarization hyperspectral tend to model,quantitative development.Therefore,the researchers put forward a variety of related data processing methods to solve and study the above problems.On the basis of previous generation,this paper puts forward a classification method based on the polarization hyperspectral generated against the network,and in order to further utilize hyperspectral and polarization characteristics,after extracting the features from the polarization spectrum data set,the different features are weighted and fused,which is intended to obtain the global and detailed information of the whole d ata set,and make full use of the known information to obtain more accurate classification results.The main research contents are as follows:First of all,from the application point of view,a set of smoke camouflage,haze weather and other noise data pre-processing scheme for war zone warfare.This paper puts forward the use of polarization and spectral chromatic information,the whole information using data correlation rules for color space conversion,the use of dark primary color prior information,th e establishment of demister model to remove fog interference,the design of polarization hyperspectral filtering and reconstruction algorithm,the acquisition process to carry out deviation correction,to avoid the polarization of changes in the block effect,information loss serious situation,to achieve the overall image pre-processing.Secondly,a polarized hyperspectral feature extraction model based on the generation of anti-network is constructed.The whole model consists of three parts,the first part is two generating networks,using transposing convolution for generating space and spectral/polarization characteristics of pseudo-sample generation,the second part is the identification network,using real samples and pseudo-samples to compare,until the identification network cannot identify the true and false samples,the generation of the network At this time to achieve the best,the third part is the feature extraction network,after the generation and identification network reached Nash equilibrium,the use of maximum pooling in the identification network output layer to extract features and stitching,respectively,extract polarization characteristics,spectral features and spatial features,in order to achieve polarization hyperspectral feature ex traction.Finally,it is proposed to use the deep-weighted fusion algorithm to achieve the fusion classification of features.The main content is to establish the pyramid model of feature weighting,the first layer uses the method of differential correlation analysis to fuse the extracted polarization spectral spatial features,fuse into two features,and then divide into different subsets,carry out corresponding fusion,and finally get the pyramid model that can reflect the global and local information at the same time,and then classify the treated features,make full use of the polarization hyperspectral information,and solve the problem of insufficient utilization of polarization information.