Research Of The Target Detection Based On The Statistical Distribution Characteristics Of Multispectral Filtering

Different substances have different spectral characteristics. Multispectral system uses thespectral variance to realize objects classification, detection and identification of objects in thebackground. It can use both the spatial and spectral information in the same time, whichovermatches the single-band image. Recently, multispectral technique has been developingrapidly. Multispectral system has a wide range of applications in environmental monitoring,geological survey, mineral detection, flood control, and plant diseases and insect pestsmonitoring. In this thesis, a detection method is studied based on the statistical distributioncharacteristics of multi-spectral filtering results.Output of linear filtering multispectral images is analyzed by using the statistical mixturemodel. Its statistical distribution meets the Gaussian mixture model (GMM). GMM parameterestimation method is studied based on Expectation-maximization (EM) algorithm Theprobability density curve of linear filtering results together with the theoretical curve of theGaussian mixture distribution are relatively analyzed.A target detection scheme is proposed. It linearly filter the obtained image firstly, tosuppress the background information output, and to maintain the target signal output. Thedistribution characteristic parameters are estimated from the filter results via using the EMalgorithm. Then, the estimated parameters are used to judge if the targets exist or not. Last,the detection threshold for judging target is determined. Meanwhile, the correspondingsimulation experiments are carried out. We choose various target numbers and sizes in theresearch processes to test the detection performance and stability of the method, and compareit with the traditional threshold method. The simulation results indicate that our method isadvantageous in the application of testing small dim targets in large quantity.Various backgrounds and targets are selected for shooting the actual surface featuresimages, multispectral images are registered by using mutual information method. The spectralinformation of the target signal is extracted, and the target characteristics are detected. Thetarget detecting results are then compared with that of the actual tracking situation. Theexperiments verify the feasibility of multispectral target detection technique presented in thepaper, and indicate the stability of the detection performance.