Enhanced Local Texture Feature Sets for Face Recognition under Difficult Lighting Conditions
[Link] Maheswari [Link] (DSCE) 09471D0602
The Challenges of difficult lightning conditions
Abstract
Making recognition more reliable under uncontrolled lighting conditions. we present a simple and efficient preprocessing chain.
we introduce local ternary patterns (LTP), a generalization of the local binary pattern (LBP) local texture desciptor.
Kernel principal component analysis (PCA) feature extraction
Traditional approaches
Appearance based Normalization based Feature based
Appearance based
Training examples are collected under different lightning conditions and directly used to learn global model of the possible illumination variations
Direct learning of this kind makes few assumptions but it requires a large number of training images and an expressive feature set.
Normalization based
Normalization based seek to reduce the image to a more canonical form in which illumination variations are suppressed Example: histogram equalization A method of contrast adjustment using the images histogram
Feature based
Feature based approaches extracts illumination-insensitive feature sets directly from the given image
Our Approach
Combine robust preprocessing and robust descriptors
Experiments show that including both is
useful
Belongs to approach II, dont need any training images.
Preprocessing Chain
(Top) the stages of our image preprocessing pipeline, and (bottom) an example of the effect of the three stagesfrom left to right: input image; image after Gamma correction; image after DoG filtering; image after robust contrast normalization.
Preprocessing: what to remove ?
Unwanted illumination effect Non-uniform illumination Shadowing & highlights Aliasing, blurring, noise Useful information Facial features : eyes, nose,etc. Ridges, wrinkles, skin details Local shadowing, shading
Basic Assumption
Unwanted lighting effects lie mainly at low frequencies Useful information lies mainly at middle & high frequencies
PreprocessingStep 1
Lighting is a multiplicative process (Lambertianmodel) observed value = illumination * albedo(local surface reflectance) Log transform corrects this log(grey value) = log(illumination)+log(albedo) In practice log transform amplifies noise in dark regions too much use Gamma transform, gamma = 0.2 0.5
PreprocessingStep2
Bandpassfilter suppresses low frequencies
(lighting variations) and high frequencies (noise) Use difference of Gaussians (DoG) inner filter reduces aliasing, noise < 1 pixel outer filter suppresses low frequency lighting variation without suppressing too much class information several pixels, depending on class
PreprocessingStep3
(Optional) mask out facial regions that are irrelevant or too variable
PreprocessingStep4
Robust contrast normalization normalizes overall range of output values must be resistant to specularities We use simple two stage method based on compressive nonlinearity
Overall Effect
Before Normalization
50 ms per image with 120x120 pixels Before
After Normalization
Local Binary Patterns (LBP)
Threshold local image at centre pixel value Resulting binary patterns are pixel-level descriptors Histogram locally to get local texture descriptors
+ Robust to most lighting changes -Sensitive to noise in near-uniform regions e.g. many facial regions (cheek, forehead)
Local Ternary Patterns (LTP)
3-valued generalization of LBP Adds a threshold / gap to counter noise sensitivity
LBP / LTP Similarity Metric
Past LBP work partition image into regular grid build local histogram of LBP codes in each grid cell use 2 histogram distance to compare images Limitations Arbitrary image partition, (cells) not aligned well with facial features abrupt spatial quantization causes aliasing
Effect of Preprocessing on the LBP Descriptor
LBP Histograms for the marked region before and after illumination normalization
Distance Transform Based Similarity Metric
For a reference image X, a test image Y Find their LBP/LTP codes Transform these into a set of sparse binary images, one for each code Calculate the distance transform image of each binary reference image For each test image position, look up corresponding reference
Distance Transform Based Similarity Metric 2
Distance calculation
Each pixel of dk gives the distance to the nearest image X pixel with code k kY(i,j)is code value of pixel (I,j) of image Y W is a monotonically increasing function
Experiments on data set
Before Normalization
Uncontrolled
controlled After Normalization
Overall Architecture
Results
Results
Conclusion
Simple preprocessing greatly improves face recognition performance under difficult lighting conditions LTP feature generalizes LBP DT based LBP/LTP similarity metric State of the art performance on several major face databases
References
T. Ojala, M. Pietikainen, and T. Maenpaa, Multiresolution gray-scale and rotation invarianat texture classification with local binary patterns, IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 971987,Jul. 2002. Y. Pang, Y. Yuan, and X. Li, Gabor-based region covariance matrices for face recognition, IEEE Trans. Circuits Syst. Video Technol., vol. 18, no. 7, pp. 989993, Jul. 2008. Y. Pang, Y. Yuan, and X. Li, Iterative subspace analysis based on feature line distance, IEEE Trans. Image Process., vol. 18, no. 4, pp. 903907, Apr. 2009. P. J. Phillips, P. J. Flynn, W. T. Scruggs, K. W. Bowyer, J. Chang, K. Hoffman, J. Marques, J. Min, and W. J. Worek, Overview of the face recognition grand challenge, in Proc. CVPR, San Diego, CA, 2005, pp. 947954. L. I. Rudin, S. Osher, and E. Fatemi, Nonlinear total variation based noise removal algorithms, Phys. D, vol. 60, no. 14, pp. 259268, 1992. S. Shan, W. Gao, B. Cao, and D. Zhao, Illumination normalization for robust face recognition against varying lighting conditions, in Proc. AMFG,
THANK U..
