A Research Of Hyperspectral Target Detection Algorithm Based On Deep Learning

With the rapid development of spectral camera imaging technology,the obtained hyperspectral images contain more fine and high-resolution spectral characteristic curves,which can better reflect the subtle differences between different objects.However,the large amount of data not only brings a serious burden to the satellite transmission and storage system but also poses a great challenge to the fast and accurate target detection of hyperspectral images.Based on the constrained energy minimization algorithm,this thesis explores the methods that can improve the detection performance of hyperspectral images.Finally,the proposed methods are objectively evaluated by real hyperspectral remote sensing images.The main achievement of this thesis includes:(1)The existing target detection algorithm based on spectral matching cannot extract the deep spectral features to reduce the redundancy between the spectra,so the detection speed is slow and cannot meet the requirements of real-time detection.In order to solve this problem,the thesis deeply analyzes the advantage of neural network automatic learning feature and pioneers proposing a spectral feature extraction model based on deep learning to reduce spectral dimension.In fact,the method can extract the distinguishable bands to remove the useless information between the bands,and the reduction of the amount of data involved in the target detection task can speed up the detection of the target.The results of feature extraction method are input into three different detectors,and then the detection performance is compared from the aspects of detection accuracy and speed.At the same time,the influence of improved algorithm and PCA band selection dimensionality reduction algorithm on processing time is compared.Experiment results demonstrate that the proposed method can accurately detect the target of different position,angle and size.The average detection speed of CEM,SBS-CEM and DPBD-CEM detector is 12.76 and 6.27 times faster than the original detection algorithm and the optimization algorithm based on band selection,respectively.(2)Research on the problem of joint extraction of spatial spectrum features to improve the accuracy of target detection.Spectral reflectance is affected by the external environment such as spectral camera and illumination,resulting in the different material with identical spectral signatures.Therefore,only relying on a single spectral characteristic curve to detect the target,the detection results fluctuate greatly and the anti-interference ability is poor.In order to make full use of the abundant spatial feature of HSI,a 3D CNN is proposed to extract spatial spectral features.To solve the problem of insufficient training samples in the process of network training,a deep convolutional feature learning method based on transfer learning is proposed,which pre-train the deep network to learn the low level visual features by using the large scale image dataset,and then a small number of training samples of the target data set is used to fine-tunes the network parameters,so as to learn the high-level features suitable for the target task.The experimental results under the same conditions show that the optimized detection algorithm of 3D CNN based on migration learning can eliminate noise interference.Compared with SAE and PCA optimization algorithms,the detection accuracy is improved by 2.45% and 2.85% on average,and the detection speed is increased by 24.29 times on average compared with the original detection algorithm,which achieves the effect of real-time processing.It lays a foundation for the application of hyperspectral target detection technology to satellites later.