Application Research Of GaborCNN Deep Learning In Hyperspectral Object Recognition

With the continuous development of remote sensing technology,hyperspectral images(HSI)have been widely used in space remote sensing,precision agriculture remote sensing,enemy situation reconnaissance,geological mapping,environmental protection,atmospheric research,extraterrestrial space exploration,development and utilization,target detection and surveillance,etc.The rich spectral information and spatial structure information contained in hyperspectral remote sensing have the characteristics of wide spectral range,large number of bands,and high resolution,which provide more information for the classification of target features.Since the classification effect of machine learning in hyperspectral feature recognition is very significant,it has been widely used in hyperspectral feature recognition.The remote sensing feature classification method directly affects the feature recognition and classification effect.In the current classification methods of remote sensing features,there is a problem that the spectral information and spatial structure information cannot be combined effectively.In this thesis,through in-depth analysis of the characteristics of hyperspectral remote sensing data,based on Fourier transform,Gabor transform,Gabor wavelet,Gabor filter,genetic algorithm,machine learning,deep learning and other fields of new theories and new methods,research on the hyperspectral information space and frequency feature extraction method,and verify the effectiveness of the method on multiple hyperspectral data sets.The main research work carried out on the utilization of space-spectrum information is as follows:(1)Aiming at the problem of frequency domain feature extraction in hyperspectral remote sensing data,the characteristics of Fourier transform and Gabor transform and their advantages and disadvantages in remote sensing feature extraction are analyzed.(2)In view of the low utilization of space-spectrum information of conventional hyperspectral remote sensing and the insufficient combination of spectral and spatial features,the Gabor filter feature extraction method combined with space-spectrum information is used to simultaneously utilize spectral domain information.Extraction of spatial features.(3)In view of the current use of empirical parameter fixed filter banks that cannot be matched with image features,the information cannot be extracted effectively,and some filter cores are redundant.The method of using a single filter core to extract the feature effect evaluation method is used to evaluate the effect of the filter.Preferably,filters that are valuable for target classification are retained.(4)Aiming at the problem that the current Gabor filter bank cannot match the data set,a hyperspectral image classification method based on parameter optimization of the Gabor filter bank is proposed.The genetic algorithm is used to optimize the parameters of the filter bank to realize the accurate characterization and matching of the spatial structure characteristics.(5)Aiming at the problems of slow convergence speed of genetic algorithm,poor local search ability,many variables to be controlled,and difficulty in determining termination criteria,combined with the convolutional neural network(CNN)method,an end-to-end depth based on Gabor feature extraction is proposed.Learning methods,while reducing the time and space complexity of the algorithm.(6)Optimizing,optimizing and learning the Gabor filter bank,using random forest,genetic algorithm,and convolutional neural network to achieve the purpose of characterizing features.