Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18An Adaptive Image An Adaptive Image Enhancement Algorithm for Face Enhancement Algorithm for Face DetectionDetection By Lizuo Jin, Shin’ichi Satoh, and Masao Sakauchi. ECE 738In Young ChungOutlineOutlineMotivation – Problem in face detectionSuggestion Basic idea of suggestionApproachAdaptive Image Enhancement Algorithm < Step 1. ~ Step 6. >Result and ComparisonConclusionProblem in Face Detection?Problem in Face Detection?Almost every face-detection methods depends on the intensity values of image  Face detection under unconstrained condition result in failure because of the drastic variation of pixel intensity in face regions Image enhancement by intensity transformation can reduce this problem, with histogram equalization (HE).  HE applied to images with faces on a very light background, it may produce very dark face regions  face detection failure.HEAfter HistogramEqualizationSuggestion of solutionSuggestion of solutionSolution?An adaptive image enhancement algorithm which is adapt to the intensity distribution within an image.Basic idea Basic idea 1. Why don’t we make it even?Entropy of darker pixels = Entropy of lighter pixels2. Face is made up of many pixelsFace pixels make a cluster in histogram We can histogram ridge analysis techniqueApproaches I.Approaches I.EER (Entropy Error Rate) as an information theoretic measurerepresents the tendency of the information distribution within an imageIf EER is positive and largethe information lies mainly in the darker pixels If EER is negative large the information is lies mainly in the lighter pixels Goal : minimizing the EERSHHEERBD))((4minmaxminminmaxmaxIIIIIIIISmeanmeanmeanIIkDkpkpHmin)(log)(manmeanIIkBkpkpH )(log)(1minmaxIIHHDD1maxmeanBBIIHHWhere, S : estimate the relative position of the mean in histogram HD,HB : information in darker pixels and lighter pixels respectively HD,HB : the average entropy in either sideApproaches II.Approaches II.Histogram Ridges Analysis : suggested in the paper “A fast histogram-clustering approach for multi-level thresholding” by Du-Ming Tsau and Ting-Hsiuing Chen Important parameter: the distance between the leftmost and rightmost ridge because this distance is related with the intensity range of valid content in the image.Adaptive Enhancement Adaptive Enhancement AlgorithmAlgorithmStep 1.Extract Intensity Value in the input imageStep 2.Step 2.Compute the intensity histogram of the input imageStep 3.Step 3.Threshold the intensity histogram Against noise and stretch to [0,255]Smoothing with Gaussian smoothing Operator with variance = 2.0Find valid ridges and distance between the ridges (Dr)this is related with the intensity range of valid content in the image.Step 4.Step 4.Filter the histogram obtained in step 2 with a filtering coefficient to eliminate noise or unimportant detailsStep 5.Step 5.Compress the detail region and expend important region by using entropy in darker and lighter sideStep 6.Step 6.Minimum EER finding processAfter gamma correction with the parameter obtained in minimum EER F.PResultsResultsBefore, Histogram EnhancementAfter Adaptive EnhancementComparison I.Comparison I.Classical histogram equalization (HE)Adaptive histogram enhancement (AE)Comparison II.Comparison II.Original image HEAEImage with very light back groundConclusion and future worksConclusion and future worksImage enhancement is very important technique for face detection, especially in the images acquired in unconstrained illumination conditionUnsuitable enhancement can increase detection-failure rateAE algorithm estimate the image quality base on EER and intensity histogram and select best transform It performs much better than classical HE methodQuestion ?Question