Research On Hyperspectral Known Target Recognition And Unknown Target Detection Algorithm

Hyperspectral images describe two-dimensional spatial image information of target distribution and one-dimensional spectral information of target spectral characteristics.The traditional hyperspectral recognition method based on feature extraction has low recognition accuracy due to the weakness of feature loss.The neural network-based hyperspectral image recognition method can obtain high recognition accuracy due to the powerful representation ability of the neural network.However,the neural network can only recognize targets of the same type as the training set called known classes.Outside the target called the unknown class,the neural network cannot detect them.In view of the above problems,this paper studies the hyperspectral known target recognition and unknown target detection algorithms.The main work and innovations of this article include:(1)Aiming at the recognition of hyperspectral known targets,this paper proposes a new type of 3D convolutional network.This network has two important characteristics.One is the residual structure.The characteristics of the residual structure can increase the number of network layers a lot.The increase of the network layers is helpful for extracting high-level data features.At the same time,the residual structure can avoid training difficulties caused by too many network layers.The second feature of this network is the use of three-dimensional convolution operations.Unlike one-dimensional and two-dimensional convolutions,three-dimensional convolution operations can simultaneously extract spectral and spatial joint information.Experimental tests show that the network has high recognition accuracy for hyperspectral known targets.(2)This paper proposes and implements an OpenMax-based hyperspectral unknown target detection algorithm.OpenMax refers to the output of the penultimate layer of the neural network as the activation vector.During training,a Weibull model is built using the correct activation vector of each known class.During the test,input data and calculation of the output activation vector using the established N Weibull probability models can obtain the probability that the input belongs to a known class and the probability that it belongs to an unknown class.Experimental tests show that the algorithm has a certain detection effect on the detection of hyperspectral unknown targets.(3)This paper proposes and implements a DOC-based hyperspectral unknown detection algorithm.DOC uses a 1-vs-rest layer instead of the Soft Max layer.The1-vs-rest layer has two functions,sigmoid mapping and decision.During training,the N component numbers output from the penultimate layer of the neural network are used to build N Gaussian models and calculate N decision thresholds for the 1-vs-rest layer.During the test,the data from the penultimate layer of the neural network is input to the1-vs-rest layer,and the decision-maker of the 1-vs-rest layer decides whether the input belongs to a known class or an unknown class according to the sigmoid mapping output value.Experimental tests show that the algorithm has a certain detection effect on the detection of hyperspectral unknown targets.(4)Finally,this paper proposes an FPGA implementation method for convolutional neural networks for hyperspectral known target recognition.Using this method,a hyperspectral known target recognition network was successfully established on the FPGA platform.Tests using public data sets show that the network on the FPGA and the network on the GPU platform have consistent recognition accuracy for hyperspectral known targets.