Research On Hyperspectral Image Target Detection Algorithm Based On Deep Convolutional Network

At present,deep learning is one of the most popular research directions under these circumstances of big data age.Due to its powerful learning ability,convolutional neural networks and their variants are widely used in hyperspectral image processing,and the potential distribution of images is mined through deep neural networks.Combined with classification and detection,deep neural networks have achieved good results.In view of the effectiveness of convolutional neural networks in hyperspectral remote sensing data processing,this paper studies the hyperspectral image target detection algorithms based on deep convolutional neural network are considered,they are from two aspects: anomaly detection and target detection.Firstly,from the imaging mechanism of hyperspectral image,the spectral characteristics and influencing factors of the ground object are analyzed.The existing hyperspectral target detection algorithm and anomaly detection algorithm are summarized.The problems and challenges faced by the classical algorithm are mainly discussed.At the same time,because a wide range of convolutional neural networks are applied in hyperspectral image processing,the convolutional neural networks are introduced and its advantages and significance for target detection are expounded.It provides a theoretical basis for determining target detection algorithms based on deep convolutional networks.Secondly,aiming at the problem that the sample size in hyperspectral image target detection is small and it is not enough to train multi-layer convolutional neural networks,a subtractive pixel pairing model is proposed.In the algorithm,the homogeneous object samples and the different types of ground objects are paired,and the spectral dimension is subtracted.A new data set is generated by the method,and the original data is generated in the new data set.The new samples generated by pairing samples between different classes are marked as the target class in the new data set,and the new samples generated by the pairing samples between background and target are marked as the background class in the new data set.In this way,the original less sample is expanded by about 4,000 times,and the target detection problem based on the original hyperspectral data set is transformed into a classification problem based on the new data set.By comparing the classical target detection algorithm,it is verified that the CNN-based detection method can achieve high detection accuracy without eliminating the noise band,which is beneficial to save the pre-processing time,and can effectively improve the detection rate and reduce the false alarm rate..Thirdly,for the anomaly detection in which target spectra information is unknown,a combined measure vector that can characterize the difference between two spectral vectors is proposed.In this strategy,the paired samples are no longer subtracted,but the Canberra distance,spectral gradient angle,spectral curve shape and spectral information divergence of the corresponding bands of the two spectral vectors are first obtained,and the four factors are compared.By multiplying the combined measure vector,the combined measure vector has the excellent characteristics of four spectral similarity measures,and contains features that can reflect the difference in amplitude,direction,curve shape and information amount of the two spectral curves.The neural network learns the nonlinear deep features contained in the mixed vector,and then uses the learned features with recognition ability to detect the anomaly pixels.By comparing the traditional anomaly detection algorithm,the validity and excellent performance of the convolutional neural network anomaly detection algorithm based on the combined measure vector is verified.