Research On Target Detection Method Based On Filter-Type Spectral Imaging

As hyperspectral imaging technology contains rich two-dimensional spatial information and one-dimensional spectral information,which can provide more information about ground objects,it has been widely used in military target recognition,precision agriculture,environmental survey,disaster assessment and other fields,and hyperspectral imaging technology has also made continuous breakthroughs in recent years.Therefore,conducting application research based on hyperspectral imaging technology has important research value and academic significance.This article designs and constructs a miniaturized and convenient hyperspectral imaging system with a spectral range of 450-900 nm,using a linear variable bandpass filter as a spectroscopic element.The system and reconstruction processing are used to obtain hyperspectral data.Based on two classic object detection algorithms,the detection of camouflaged and hidden targets in complex backgrounds is completed.The detection results show that good detection results can be obtained for targets of different sizes,a combination of SAM and CEM was used to optimize the situation of multiple missed detections in classical algorithm detection,effectively improving detection efficiency.The main content is as follows:(1)Based on the principle of spectral imaging technology,a scheme of linear variable bandpass filter type imaging spectral system was designed.The optical path structure of the hyperspectral imaging system was built by calculating the optimal distance of each component of the system and the optical system parameters,and the system debugging and system calibration in the laboratory were carried out to determine the Spectral resolution of the system and the central wavelength of each channel.By reconstructing the collected data from the system,a complete hyperspectral dataset was obtained.Reflectivity correction and denoising level processing were used to preprocess the hyperspectral data and obtain hyperspectral data for subsequent target detection.(2)Two classic object detection algorithm models,Constrained Energy Minimization(CEM)and Adaptive Cosine Estimator(ACE),were established on the MATLAB platform for object detection on standard datasets.ROC curves and AUC values were used for performance evaluation.The detection results show that the CEM algorithm detected AUC values of 0.9907 and the ACE algorithm detected AUC values of 0.9795,respectively,It can be seen that the CEM algorithm has better detection performance than the ACE algorithm.On this basis,the CEM algorithm with better performance was selected to perform target detection on 9 sets of hyperspectral datasets corresponding to different targets and target sizes collected by the system.The results show that when the false alarm probability is 0.1492,the detection probabilities of dark false leaves,light false leaves,and camouflage cars with an exposed area of 2/3 accounting for 20% are 0.9851,0.9790,and 0.9098,respectively,The AUC values are 0.9883,0.9866,and 0.9201,indicating good detection results.(3)In the process of object detection using the CEM algorithm on the hyperspectral data collected in this article,there were cases of missed and multiple detections of the target.In order to further improve the efficiency of object detection,this article used the Spectral Angle Mapping(SAM)algorithm to study the similarity between the target and the dataset.Through image segmentation and image filling,the target that needs to be improved for missed and multiple detections was optimized,Joint CEM algorithm for object detection again to improve detection efficiency.The results of using the optimization algorithm show that the detection probabilities corresponding to the same false alarm probability after optimization are 0.9919,0.9827,0.9126,and AUC values are 0.9923,0.9913,and 0.9401 when the target proportion in the above 2/3 scenario is20%.It can be seen that the detection probability and detection efficiency have been improved after optimization.