Image Pattern Recognition The identification of animal species through the classification of hair patterns using image pattern recognition: A case study of identifying cheetah prey. Principal Investigator: Thamsanqa Moyo Supervisors: Dr Greg Foster and Professor Shaun Bangay. Presentation Outline • Problem Statement • Objectives • Approach • Research Done • Conclusion Problem Statement • Hair identification in Zoology and Forensics • Subjectivity Problem Statement • First application of automated image pattern recognition techniques to the problem of classifying African mammalian species using hair patterns. – based on the numerical and statistical analysis of hair patterns. Approach to the Study: • Lack of literature focused on hair recognition • Multi-disciplinary nature • New process designed Approach to the Study: Process Stages Image Capture Sensor Feature Feature Classifier System Generation Selection Design Evaluation • Each stage detailed next Figure Adapted from Theodoris et al (2003:6) Research Done: Image Capture • How can hair pattern images be captured? – Based in Zoology Department – 2 approaches considered Light Microscope SEM Research Done: Image Capture Light Microscope SEM Scale Patterns Cross Section Patterns Research Done: Image Capture • Scale Patterns – Use SEM – Better representation of texture in image Light Microscope SEM Research Done: Image Capture • Cross section patterns – Use Light microscope – 2D shape preferred to a 3D shape Light Microscope SEM Research Done: Image Capture • Decisions affecting design – Scale patterns texture based – Cross section patterns shape based – 2 separate sub-processes – Decision not to combine their results Research Done: Sensor • What image manipulation techniques are applied in a hair pattern recognition process? – Scale Pattern Processing • User defined ROI • Handle RST variations • No need to cater for reflection variations • Convert to greyscale Research Done: Sensor Stage • What image manipulation techniques are applied in a hair pattern recognition process? – Cross section pattern processing • • • User defined ROI Image segmentation and thresholding Challenges Research Done: Sensor Stage Original Thresholding Edge Detection Grab Cut + Thresholding Research Done: Feature Extraction How can features be extracted? • Scale Pattern Processing – Gabor filters – Capture pattern orientation and frequency information – Produces n number of filtered images where n is the size of the Gabor filter-bank Research Done: Feature Extraction Filtered Images from a Gabor Filter of size 4. Images filtered at initial orientation of 0 degrees Images filtered at initial orientation of 180 degrees Research Done: Feature Extraction How can features be extracted? • Cross Section Processing – Hu’s 7 moments – RST invariant shape descriptors – Calculated from central moments – Require black and white image Research Done: Feature Selection What selection of features is necessary • Scale Pattern Processing – Image tessellation – Use of variance or average absolute deviation Research Done: Feature Selection What selection of features is necessary? • Cross section processing – None required for Hu’s moments – Would affect scalability of the process Research Done: Classifier Design • What mechanisms can be used to classify features? – Scale Pattern Processing • • Euclidean distance measure 3 Scale patterns used to train – Cross Section Processing • • Euclidean distance measure or Hamming distance measure 10 cross section patterns used to train Research Done: Results • From implementation using: – ImageJ plugins written in Java 1.4 – 25 scale patterns processed – 50 cross section patterns processed Research Done: Results Scale pattern results (Variance) 60% % Correct Classifications 50% 40% Best Worst Changes 30% 20% 10% 0% 4 Filters 8 Filters Number of Filters 16 Filters Research Done: Results Scale pattern results (AAD) 80% % Correct Classifications 70% 60% 50% Best Case Worst Case Changes 40% 30% 20% 10% 0% 4 Filters 8 Filters Number of Filters 16 Filters Research Done: Results • Summary of scale pattern results: – AAD is a better feature selection method – Results most stable with 8 filters using AAD as feature selector – Explanation of this result Research Done: Results Cross section pattern results 100% 90% % Correct Classifications 80% 70% 60% Euclidean Hamming 50% 40% 30% 20% 10% 0% Blue Wildebeest Impala Jackal Species Springbok Zebra Research Done: Results • Summary of cross section pattern results: – Euclidean distance overall classification rate: 26% – Hamming distance overall classification rate: 40% – Explanation of this result Conclusion • Findings and Contributions – Gabor filters and moments shown to provide hair pattern classification information – AAD performs better feature selection than variance – Hamming distance more suitable classifier of moments than Euclidean distance – First application of hair pattern recognition on African mammalian species hair. Questions • • • • • • Manual Preparation Work Sensor Feature extraction Feature Selection Classifier Design Results