Wavelet frame-based feature extraction technique for improving classification accuracy

  • PDF / 469,524 Bytes
  • 9 Pages / 516 x 707.75 pts Page_size
  • 38 Downloads / 227 Views

DOWNLOAD

REPORT


J. Indian Soc. Remote Sens. (September 2009) 37:423–431

RESEARCH ARTICLE

Wavelet Frame-Based Feature Extraction Technique for Improving Classification Accuracy R.A. Alagu Raja . V. Anand . S. Maithani . A. Senthil Kumar . V. Abhai Kumar

Received: 15 April 2009 / Accepted : 8 May 2009

Keywords Wavelet frame transform . Texture classification . Unsupervised clustering . Accuracy assessment . Feature extraction

Abstract Classification of textures in remotelysensed data has received considerable attention during the past decades. One difficulty of texture analysis in the past was lack of adequate tools to characterize different scales of textures effectively.

R.A. Alagu Raja1( ) . V. Anand2 . S. Maithani3 . A.S. Kumar4 . V.A. Kumar5 1 Remote Sensing & GIS Lab, Thiagarajar College of Engineering, Madurai – 625 015, India 2 R & D Division, Tata Consultancy Services, Pune – 411001, India 3 HUSAD, Indian Institute of Remote Sensing, Dehradun – 248 001, India 4 National Remote Sensing Centre, Indian Space Research Organisation (ISRO), Hyderabad – 500 037, India 5 Thiagarajar College of Engineering, Madurai – 625 015, India

email : [email protected]

Recent space-frequency analytical tools like the wavelet transform can effectively characterize the coupling of different scales in texture and helps to overcome the difficulty. This paper presents a wavelet-based texture classification technique that was applied to a Multi-Spectral Scanner (MSS) image of Madurai City, Tamil Nadu, India The feature extraction stage of the technique uses Lemarie-Battle orthogonal wavelets to derive a texture feature vector and this vector is input to a fuzzy-c means classifier, an unsupervised classification procedure. Four indices (user’s accuracy, producer’s accuracy, overall accuracy and Kappa co-efficient) are used to assess the accuracy of the classified data. The experiment results show that the performance of the presented technique is superior to the classical techniques.

Introduction Texture analysis is a fundamental method for many applications in areas like remote sensing, digital

424

imaging, quality inspection and medical imaging. Textures are replications, symmetries and combinations of various basic patterns usually with some random variation (Haralick, 1979). Although texture analysis has a long history, its applications to satellite sensor image data have been limited. Classification of textures in digital images has received considerable attention in the literature and a large number of approaches have been suggested. Texture classification is a process to extract features from a set of texture classes. In order to have an effective classification, features with good discriminatory details have to be obtained. Most of the existing approaches for texture feature extraction make use of statistical techniques in which processing the texture image data requires large storage space and computational capability. The scalar features calculated from the feature matrix may not be efficient to represent the characteristics of t