Synthetic Aperture Radar Automatic Target Recognition -Computer Science DepartmentCalifornia Polytechnic State University, San Luis Obispo Alvin Y. Wang and Chia-Huei Yao Faculty Advisor: Dr. John Saghri Project Sponsor: Raytheon Company Contact Personnel: Jeff Hoffner Agenda Introduction Automatic Target Recognition Synthetic Aperture Radar Problem and Proposed Solutions Feature Extraction Image Matching Conclusion Introduction Usage of image identification Military Medical SAR images MSTAR image database Courtesy of Sandia National Laboratory Synthetic Aperture Radar SAR instruments use pulses of microwaves as an active source of illumination Benefits Independent of light sources Capable to see through clouds Spatial resolution remains the same no matter how far the target area is Automated Target Recognition Five Stages Feature Extraction – Detection Feature Enhancement - Discrimination Image Matching – Classification, Recognition, & Identification Database Input Image Templates Feature Feature Target Extraction Enhancement Classification Noise and Nonfeature Found Not Found Problem and Proposed Solutions Traditional ATR algorithms Problem: Removal of useful target information Solution: Multi-feature ATR techniques Feature Extraction Edge Detection, Topographical Primal Sketch Image Matching Hausdorff Distance Transform Feature Extraction Feature Detection Edge Detection – Sobel Mask Line Detection – Laplacian Mask Topographical Primal Sketch Multiple-feature consideration Wait…Before Feature Detection Reject Noise The target images are full of noise Median filter Edge Detection The box provides little clue for identification Even worse, the edges are affected by different illuminating status and orientation SAR image Extracted Edge (before threshold) T72 Tank in different orientation Topographical Primal Sketch The light intensity variations on an image are caused by an object’s surface orientation, its reflectance, and characteristics of its lighting source Based on the variance of light intensity, we can classify and group the underlying image into some topographical categories Topographical categories includes: peak, pit, ridge, ravine, saddle, flat, hillside, etc. Based on the location of the topographical features, we can reasonably reconstruct the original 3D model. Feature Extraction and Distance Transform Edge Feature Extraction Original Image Distance Transform Peak Ridge Image Matching Database Model templates Problems Scale Rotation Partially obstructed images Distance Transform Image Matching procedure Find contour points of the reference shape and obtain their DT Obtain contour points of the measured shape Compute and superimpose the centroids of the two point sets Rotate and translate the measured point set with respect to the initial pose Select those relative positions that yield the minimum HD value Select the one with the least mean HD. Hausdorff Distance Transform h(A,B) = max {min { d(a,b)} } H(A,B) = max {h(A,B), h(B,A)} Hausdorff Distance Illustration a2 h(A,B) a1 b3 b1 H(A,B) h(B,A) Hausdorff Distance provides a measure of set A and set B’s proximity – it indicates the maximal distance between any points of A to B. b2 Chamfer Distance Transform CDT Provides good approximation to the exact Euclidean distance Distance Trasform converts a binary image to another image in which pixel value is the distance from this pi xel to the nearest nonzero pixel of the binary image. courtesy of IPAN Image Matching procedure Image Matching procedure Find contour points of the reference shape and obtain their DT Obtain contour points of the measured shape Compute and superimpose the centroids of the two point sets Rotate and translate the measured point set with respect to the initial pose Select those relative positions that yield the minimum HD value Select the one with the least mean HD. An image (left) and its distance transform (right) Test image and Target detected when the contours are superimposed courtesy of IPAN Template image Test image Target detected courtesy of Cornell Vision Group Conclusion Current Progress and Future Directions Feature Extraction Feature detection TPS Image Matching Hausdorff Distance Transform Testing Database Actual Matching with test images References Image and Pattern Analysis Group – http://visual.ipan.sztaki.hu/ Cornell Computer Vision Group http://www.cs.cornell.edu/vision Robert M. Haralick, Layne T. Watson, Th omas J. Laffey, The Topographic Primal Sketch. The international Journal of Rob otics Research. Vol. 2, No. 1, Spring 1983 Thank You Questions and Comments Visit our web page Alvin: www.csc.calpoly.edu/~aywang Huey: www.calpoly.edu/~cyao