Research On Concealed Target Detection Technology Based On Hyperspectral Fusio

In the military field,the detection interference caused by the scene of the target is particularly obvious.In military camouflage operations,when the band of target is adjacent to the band of the scene,the detection method of multispectral images is difficult to distinguish the target from scene at the level of image spatial information.The continuous spectral information of each pixel in the hyperspectral image is well preserved,the spectral characteristics of the pixels are more abundant,containing a large amount of deep dimensional information,and in the camouflage scene,the target detection according to the spectral characteristic curve is more accurate than the detection using spatial information.However,the huge amount of data in hyperspectral images is very easy to cause data disasters,and it is necessary to correctly and effectively process the data and select the appropriate pixels for the hidden target to analyze its spectral characteristics.In view of the research of hidden target detection technology,through data processing and feature fusion of hyperspectral/multispectral images in the same scene,the hidden target detection experiment under environmental interference is completed.Main contents are as follows:(1)Introduced the characteristics of hyperspectral/multispectral images,and formulated algorithms and hardware implementation schemes for hidden target detection research(2)An initial correction model is constructed by using VGG-16 convolutional network which extracted spatial features from multispectral images.According to the correction model,a single-band PAM clustering method and Coefficient of variation discriminant are used to screen out the bands containing the main spectral features of the target.At the same time dimensionality reduction of hyperspectral images is completed.(3)Comprehensive use of the spatial and spectrum features of the two images,HSV fusion based on multispectral fusion and supervised spatial spectrum fusion network fusion are two methods to complete the deep fusion of the two image features after extraction.(4)On the basis of the feature fusion results,a decision-level fusion module for hidden object detection is added,and the hidden object detection algorithm based on context semantic level segmentation is used to achieve hidden object detection.The fusion and object detection algorithms are ported and optimized in Hi3559 to improve hardware utilization and portability.