Contrast Image Correction Method Journal of Electronic Imaging, Vol. 19, No. 2, 2010 Raimondo Schettini, Francesca Gasparini, Silvia Corchs, Fabrizio Marini, Alessandro Capra, and Alfio Castorina Presented by Woo-Heon Jang School of Electrical Engineering and Computer Science Kyungpook National Univ.
abstract Proposed method Contrast enhancement Local correction Overexposed region Underexposed region Image dependent exponential correction Using bilateral filter» Prevention halo artifact Preserving treatment» Stretching» Clipping» Saturation 2 / 30
Introduction Contrast image correction Considering scene of room by sunlit landscape Human observer Adaptation of eyes Photograph» Easy recognition of object in room» Easy recognition of feature in outdoor landscape Unaware outside landscape by overexposed window Unaware inside object by underexposed room 3 / 30
Two classes of contrast corrections Global correction Overall adjustment in image Disappointing result when shadow and highlight detail Gamma correction Histogram equalization Local correction Local adjustment in image» Allowing way for simultaneous shadow and highlight adjustment 4 / 30
Different algorithms for local contrast correction Moroney method Using nonlinear masking Simple pixel wise gamma correction of input data Representation of halo artifact» Boundary in scene Shrinking of dynamic range of scene Adaptive histogram equalization method Use of Local image information Modified cumulation function» Added weighting at cumulation function in local region 5 / 30
Retinex model Emphasis of reflectance component Elimination of luminance by acquisition of reflectance component Multiscale retinex with color restoration Rizzi et al. Automatic color equalization Enhancement of simultaneous global and local effect» Satisfactory tone equalization» Color constant correction High computation cost 6 / 30
Fairchild and Johnson Image color appearance model Traditional color appearance capability Chui et al.» Spatial vision attribution» Color difference metric Repeated clipping and low-pass filtering Larson et al Presented tone reproduction operator Preserving visibility in HDR scene Battiato et al Survey of early methods» Application of HDR image 7 / 30
Meylan and Susstrunk Using adaptive filter for halo artifact Durand and Dorsey Using anisotropic diffusion for enhanced edge Pattanik et.al Computational model by tone reproduction Proposed method Local contrast correction Elimination of halo artifact by bilateral filtering Estimation of optimized parameter for contrast correction Histogram clipping for improvement of overall image 8 / 30
Local contrast correction method Local exponential correction in LCC Gamma correction Elaborating input image through exponential γ Gamma correction transform Ii (, j) Oi (, j) = 255 255 where γ is a positive number, Ii (, j) is each pixel of input image, and Oi (, j) is each pixel of output image. γ (1) Satisfactory result for total image Underexposed image Overexposed image 9 / 30
Dissatisfactory result Simultaneous Correction of under and overexposed regions Fig.1. (a) Original image with simultaneous underexposed and overexposed regions. (b) Gamma correction with γ =0.35. 10 / 30
Extension of gamma correction Need of local correction Adjustment of simultaneous shadow and highlight region Ii (, j) Oi (, j) = 255 255 γ [ i, j, N(, i j) ] (2) where Ni (, j) is neighboring pixels. Expression for exponent by Moroney s method γ [ 128 mask ( i, j)/128] [ i, j, N(, i j) ] = 2 (3) where mask(, i j) is an inverted Gaussian low-pass filtered version of the intensity of the input image. 11 / 30
where Greater distance from mean value 128» Stronger correction [ 128 BFmask ( i, j)/128] [ i, j, N(, i j) ] = γ α BFmask(. i j) is an inverted low pass version of intensity of input image, filtered with a bilateral filter, and α is a parameter depending on the image properties. (4) 12 / 30
Bilateral filter in LCC Computation of Gaussian low pass filter Weighted average of pixel values in neighborhood Decreasing weight with distance from center Similar values at near pixels Slow spatial variation at edge Blurring by low pass filtering Bilateral filter Preserving edge Smoothing image Decreasing halo artifact 13 / 30
Local exponential correction by proposed method Ii (, j) Oi (, j) = 255 255 Bilateral filter mask 1 BFmask(, i j) = ki (, j) i+ K j+ K p= i Kq= j K [ 128 BFmask ( i, j )/128] α 1 exp ( i p) + ( j q) 2 2σ 1 2 2 1 exp (, ) (, ) 2 inv inv 2σ 1 I ( pq, ) inv [ I i j I pq] 2 (5) (6) where ki (, j) is the normalization factor. 14 / 30
Normalization factor i+ K j+ K 1 1 ki (, j) = exp ( i p) + ( j q) 2 ki (, j) p= i Kq= j K 2σ 1 1 exp (, ) (, ) 2 inv inv 2σ 2 [ I i j I pq] 2 2 2 (7) where σ 1 σ 2 is standard deviation of Gaussian function in spatial domain, and is standard deviation of Gaussian in intensity domain. Dimension of window Depending shape of spatial Gaussian K = 2.5 σ1 (8) where is integer part. 15 / 30
Fig. 2. Original image to be elaborated. Dimension: 511341. Fig. 3. (a) LCC output of Fig. 2. (b) Bilateral filtered mask used. Fig. 4. (a) Moroney correction of Fig. 2. (b) Gaussian mask used. 16 / 30
α parameter optimization Performance of different contrast correction Depending on characteristic of single shot Adjusting α value Development of automatic tool [ 128 BFmask ( i, j )/128] α Ii (, j) EOi [ (, j) ] = 255 E 255 where E is expected value. (9) 17 / 30
Mean value of gray image equal to 128 [ 128 BFmask ( i, j )/128] α Ii (, j) EOi [ (, j) ] = 255 E = 128 255 (10) Estimated values ln( I / 255) α ln(0.5) when BFmask = 255 ln(0.5) α ln( I / 255) when BFmask = 0 (11) (12) where I is the estimated value or mean value of the input image. 18 / 30
(a) (b) (c) (d) Fig. 5. (a) Original image; dimension 320X240. (b) LCC output with α =1.5. (c) LCC output with α =2. (d) LCC output with α =2.6. 19 / 30
Contrast enhancement chain : stretching, clipping and saturation gain in LCC Stretching and clipping Analysis of intensity histogram Dissatisfaction of overall contrast enhancement Low quality image with compression artifact Enhanced noise in dark zone Domination of undesirable loss in image quality Stretching Clipping Saturation 20 / 30
Entire enhancement procedure of proposed method Fig. 6. Flowchart of the local contrast correction method proposed here. 21 / 30
Color saturation Minimized change of color saturation 1 Y ' R' = ( ) 2 R+ Y + R Y Y 1 Y ' G' = ( G Y) G Y 2 + + Y 1 Y ' B' = ( B+ Y) + B Y 2 Y (13) where Y ' is the corrected luminance obtained after the (LCC + clipping) correction module 22 / 30
Results and Discussion (a) (c) (e) (b) (d) Fig. 7. (a) Original image. (b) Histogram of (a). (c) LCC output with α =2.5. (d) Histogram of (c). (e) LCC+clipping+saturation. (f) Histogram of (e). (f) 23 / 30
(a) (c) (e) (b) (d) (f) Fig. 8. (a) Original image. (b) Histogram of (a). (c) LCC output. (d)histogram of (c). (e) LCC+clipping+saturation. (f) Histogram of (e). 24 / 30
(a) (b) (c) Fig. 9. (a) Original image. (b) Our proposed method. (c) Moroney correction. (d) Retinex. (d) 25 / 30
Quality assessment Objective metric Estimation of brightness and contrast in image Entropy Occupation of all intensity level Evaluation of pleasing image Absolute mean brightness error Absolute difference between input and output mean Measure of enhancement Approximation of average contrast in image Maximum and minimum intensity values in average Subjective evaluation Mean opinion score 26 / 30
(a) (b) (c) Fig. 11. (a) Original image (EME=25.09, H=5.86). (b) Our proposed method (EME=18.09, AMBE=18.89, H=6.01). (c) Moroney correction (EME=23.76, AMBE=23.86, H=6.58). (d) Retinex (EME=31.70, AMBE=48.58, H=7.61). (d) 27 / 30
Fig. 10. (a) Original image. (b) Our proposed method. (c) Moroney correction. (d) Retinex. 28 / 30
Fig. 12. (a) Original image. (b) Our proposed method. (c) Moroney correction. (d) Retinex. 29 / 30
Conclusion Local contrast correction algorithm Prevention of halo artifact Edge preserving filter Bilateral low pass technique Comparison with other solutions Enhancement of dynamic range in image Common quality loss Halo artifact Desaturation Grayish appearance 30 / 30