Research On Spectra Processing And Regconition Algorithm Of Image-spectra-joint Detection System

Mostly,the target detection is achieved by images.However,when the detection distance is far,or there are many pseudo/false/jamming objects,the target's texture,shape and other information are not obvious in the image.At this time,the Image-Spectra-Joint detection will have great advantages,especially for the moving targets.As we all kown,object's spectrum can be used as basis for the identification by their unique feartures.But,the spectrum obtained by the detection system does not have the same response on each band.And the spectral information is affected by the absorption,scattering,radiation,etc.of the atmosphere.So the measured value of the sensor may be inconsistent with the actual spectral radiation value of the target.In addition,the acquired spectrum may be a mixed spectrum of multiple targets beacause of the limited system's spatial resolution.And the feature extraction methods of the spectrum are mostly concentrated on the waveform and dimensionality reduction methods.When the spectral difference is small,these feature extraction methods are not applicable for object's recognition.For the above problems,the spectral characteristics of the Image-Spectra-Joint systems are studied in this paper.The corresponding mathematical model to process the various spectral data acquired in different scenes are established.For the target's infrared spectrum and target recognition task in the Image-Spectra-Joint systems,the studies of unmixing,radiation correction and atmospheric correction on spectrum are used to exclude the impact caused by the measurement conditions.The unmixing method includes the approach based on the spectral Gaussian mixture model and the unmixing method of the Image-Spectra-Joint.The multi-point calibration method is adapted in the radiation correction,and the MODTRAN atmospheric transmission model is used in the atmospheric correction.In order to distinguish various objectives in essence,a method to estimate the target's brightness temperature and emissivity is proposed in this paper.In addition,a method for identifying object's spectra using the stable amplitude/phase frequency features of the spectrum and classification methods based on convolutional neural networks and spectral feature transformation are also proposed.The Image-Spectra-Joint system's recognition accuracy of the target is greatly improved by using the above techniques and reseaches,which provides a more effective guarantee for its target detection and identification reliability.