Multi-spectral Image Real-time Target Recognition Method Based On GPU Parallel Operation

Hyperspectral data contains both spatial and spectrum information of scene.Because of this characteristic,hyperspectral data can describe the scene more precisely and richly.In the field of target detection and image classification,it shows unique advantages.Due to the lack of prior information of targets,unsupervised classification technology becomes a fine choice for unknown targets and unknown scenes.With the increase of hyperspectral methods,the quality of hyperspectral images is getting higher and higher,and the amount of hyperspectral data is becoming larger and larger.how to detect the unknown target information quickly and accurately in the massive hyperspectral data become a tough problem.Affected by noise,when the target size is small and discrete,traditional unsupervised classification method is difficult to detect.How to efficiently and accurately detect the discrete small target in this case is another problem.Real-time unsupervised classification of hyperspectral images based on Normalized spectral vector.Traditional unsupervised classification which based on clustering algorithm has low precision and takes long time.To improve the efficiency of classification while ensuring clear and accurate,a new algorithm of real-time unsupervised classification based on Normalized spectral vector and GPU parallel optimization is proposed.Classification accuracy is improved via spatial coherence property,while computing speed is improved via GPU parallel processing,the normalized spectral vector is used to simplify the formula for calculating the similarity between pixels.The proposed algorithm outperforms current unsupervised classification algorithms on classification accuracy and computing efficiency.Experimental results demonstrate the effectiveness of the proposed algorithm.The common small targets and abnormal target extraction algorithms usually distinguish the data anomalies from the image dimension information.On the one hand,the utilization rate of hyperspectral data is low;on the other hand,when the dimensionality of hyperspectral images increases,the huge amount of data will cause "computational disaster".To solve this problem,a new algorithm based on spectral dimension is proposed.From the spectral dimension,this method uses the difference of spectral curves to divide the abnormal targets,and improves the utilization ratio of hyperspectral data while reducing the computational complexity,thus effectively overcoming the traditional defects.Experimental results demonstrate the effectiveness of the proposed algorithm.In conclusion,this paper focuses on real time target recognition based on hyperspectral images,bases on the spatial and spectral characteristics of hyperspectral image data,combines with the parallel technology of graphic processing unit(GPU),studies the classification and discrete small target recognition of hyperspectral imagery respectively,tries to solve the above problems.