Ashish Uthama EOS 513 Term Paper Presentation Ashish Uthama Biomedical Signal and Image Computing Lab Department of Electrical and Computer Engineering University of British Columbia G. Chen and T.D. Bui "Invariant Fourier-wavelet descriptor for pattern recognition," Pattern Recognition, vol. 32, pp
Ashish Uthama The problem … Pattern recognition: Classifying an object into predetermined categories Applications: Written character recognition Object identification for unmanned vehicles Content based image retrieval ……
Ashish Uthama What’s in it for me? My problem: Try to find if there is a significant difference two groups of 3 dimensional distributions. Quantify this difference. Similarities between the problem domains: Sparse representation of the object Sparse enough to significantly speed up the computations Complete enough to discriminate between important differences Use this representation to classify (differentiate)
Ashish Uthama Solution requirements … Translation and scale invariant representation Rotation invariant representation Noise resistant
Ashish Uthama Translation invariance Achieved by changing the origin to the centroid (Centre of gravity/mass ) of the image
Ashish Uthama Achieved by normalizing in the polar coordinate system ‘N’ concentric circles (radius = d*i/N) Scale invariance
Ashish Uthama Rotational invariance Analyzing the data along polar angle axis Rotation results in circular shift of signals along this axis 1-D Fourier transform results in features that are invariant under rotations
Ashish Uthama Feature extraction Apply wavelet transform along the radial direction (after 1-D Fourier) Multiresolution representation Haar, Daubechies-4, Coiflet-3 and Symmlet-8 basis tried with no much difference in performance Coarse coefficients aggregate at the center
Ashish Uthama Classification Number of coefficients are small in coarse scale and increase with scale Use the wavelet coefficients to locate a match progressively At each scale: If only one match found : STOP (object classified) If none match : STOP (object can not be classified) If more than one match: Repeat at next scale Efficient, Reduces number of entries to search
Ashish Uthama Images from the paper
Ashish Uthama Results Table shows the performance of this approach using Haar wavelet basis.
Ashish Uthama Critique Image parameters and algorithm parameters (N, angular resolution, database size/content) not mentioned in the results Performance under noise not evaluated (Effects all steps) Effect of Quantization/ Re-sampling (while converting to polar) errors not clear Details of comparing coefficients not presented (Distance between coefficients?) Handling of different number of samples along the angular direction not clarified
Ashish Uthama Critique Novel, simple and intuitive! Invariance of extracted features seems plausible (as demonstrated) Computations/Comparisons for classification reduced Easily extensible to 3D!
Ashish Uthama Questions … Comments … ?