A Statistical Approach for Multiclass Target Detection

Abstract Fukunaga-Koontz Transform (FKT) is a statistical technique which has many application areas for two-class classification or detection problems. In this paper, we have proposed improved target detection algorithm for hyperspectral imagery (HSI) based on enhanced FKT which gives better results for multi-class target detection problems. Hyperspectral imagery is popular for target detection applications due to its additional properties compare to multispectral images. It presents one dimensional (spectral) and two dimensional (spatial) features of targets for detection problems. For multi-class target detection in hyperspectral images, we have selected each target's 1D features, called signature of targets, to introduce to enhanced FKT during learning stage. After learning stage of FKT, we have applied our operators, obtained by enhanced FKT, to HSI images to detect more than one target simultaneously. The experimental results show that multi-class FKT has satisfactory performance over 95 percent of true detection rate especially on the pixel sized targets. In addition, multi-class processing ability of the proposed enhanced FKT is very important for many applications such as classification, recognition, clustering, tracking problems in the literature.

[1]  Chein-I Chang,et al.  Unsupervised target detection in hyperspectral images using projection pursuit , 2001, IEEE Trans. Geosci. Remote. Sens..

[2]  Mark R. Pickering,et al.  Unsupervised band removal leading to improved classification accuracy of hyperspectral images , 2006, ACSC.

[3]  Gary A. Shaw,et al.  Hyperspectral Image Processing for Automatic Target Detection Applications , 2003 .

[4]  Kenneth A. Brown,et al.  Intelligent Target Detection in Hyperspectral Imagery , 1999 .

[5]  Jie Yang,et al.  Infrared small target detection with kernel Fukunaga–Koontz transform , 2007 .

[6]  Lori M. Bruce,et al.  Automated detection of subpixel hyperspectral targets with continuous and discrete wavelet transforms , 2001, IEEE Trans. Geosci. Remote. Sens..

[7]  A. Plaza,et al.  A new method for target detection in hyperspectral imagery based on extended morphological profiles , 2003, IGARSS 2003. 2003 IEEE International Geoscience and Remote Sensing Symposium. Proceedings (IEEE Cat. No.03CH37477).

[8]  Abdullah Bal,et al.  Hyper-spectral image segmentation using spectral clustering with covariance descriptors , 2009, Electronic Imaging.

[9]  Johannes R. Sveinsson,et al.  Classification of hyperspectral data from urban areas based on extended morphological profiles , 2005, IEEE Transactions on Geoscience and Remote Sensing.

[10]  Bernhard E. Boser,et al.  A training algorithm for optimal margin classifiers , 1992, COLT '92.

[11]  Marios Savvides,et al.  Kernel Fukunaga-Koontz Transform Subspaces For Enhanced Face Recognition , 2007, 2007 IEEE Conference on Computer Vision and Pattern Recognition.

[12]  Keinosuke Fukunaga,et al.  Application of the Karhunen-Loève Expansion to Feature Selection and Ordering , 1970, IEEE Trans. Computers.