HDR Image Compression based on HDR Image Compression based on Local - PowerPoint PPT Presentation

HDR Image Compression based on HDR Image Compression based on Local Adaptation for Scene and Local Adaptation for Scene and Display Using Retinal Model Display Using Retinal Model th Color Imaging Conference , 14 th Color Imaging Conference , 14 Lijie Wang, Takahiko Wang, Takahiko Horiuchi Horiuchi, and , and Hiroacki Hiroacki Kotera Kotera Lijie Presented by Tae Hyoung Hyoung Lee Lee Presented by Tae School of Electrical Engineering and Computer Science Kyungpook National Univ.

Abstract Abstract � Simple static local adaptation method for HDR image compression – Recreation of the same sensations between the real scene and the compressed images on displays at steady state local adaptation, respectively – Scene adaptation based of retinal model – Use of bilateral filter for preserving details without banding artifacts 2 / 20

Introduction Introduction � Dynamic range – HVS : 14-order magnitude – HDR images : 6-order magnitude – LDR display : 2 or 3-order magnitude � Tone mapping – Recreation of HDR image to LDR images � Perception of HDR of scene luminance by HVS – Adaptation 3 / 20

� Tone mapping – Global model • Simple and efficient by using single spatially-invariant curve • Problem in local contrast – Local model • Preservation of local visual contrast by using spatially variant operation • Artifacts around high contrast edges 4 / 20

� Proposed method – HDR image compression based on a retinal model • Prediction of the response of eyes at any given adaptation level – Aim • Recreation of the same sensations between the real scene and its range compressed image on the displays at steady state local adaptation, respectively – Computation of adaptations for both the scene and displays – Use of bilateral filter • Suppression of the banding artifacts 5 / 20

Related work Related work � Retinex – Preservation of local contrast and details – Banding artifacts around high contrast edges � MSR – Removal of banding artifacts with good color appearance by using 3-7 SSR images 6 / 20

Retinal Model Retinal Model � Model of retinal cell – S -shaped response n I = R ( I ) + σ n n I Where, is the light intensity, I σ is I value that causes the half-maximum response, and n is a sensitivity control parameter as 0.73(0.7-2.0). σ – Role of • Determination by adaptation to the overall scene intensity • Small value with dark scene – Perception of glares at high light environment • Large value with light scene Fig. 1. The response of retina at adaptation level for overall luminance. Dramatic compression at high and dark shadow. 7 / 20

Tone reproduction framework Tone reproduction framework � Objective – Match between viewed scene appearance and display luminance n I ( x , y ) = = w R R ( x , y ) + σ scene w n n I ( x , y ) ( x , y ) w w n I ( x , y ) = = d R R ( x , y ) + σ display d n n I ( x , y ) ( x , y ) d d (Scene luminance) Appearance vector I w = − Q R ( I ) R blk (Display luminance) Where, is the reference R blk I black response d Fig. 2. Tone reproduction framework of Pattanaik et al. 8 / 20

– Assumption (difference between Pattanaik’ model) [1]Display appearance is considered to be equivalent to scene appearance as shown in Figure 3 [2]Both scene and display adaptations are local [3]Display adaptation is formulated in relation to scene adaptation by retinal model [4]Bilateral filter is applied to mimic the edge preserving of HVS to reduce the banding artifacts Fig. 3. Tone reproduction framework of proposed model. 9 / 20

– Objective • Recreation of the same sensations between the world scene and the display, respectively • Proposed model – Preservation of contrast and omission of n = R ( x , y ) R ( x , y ) w d I ( x , y ) I ( x , y ) Final value Final value = d w σ σ ( x , y ) ( x , y ) d w Unknown values Unknown values 10 / 20

Display adaptation Display adaptation � Computation of display adaptation – Reason of use for display • The same sensation of HVS after own local adaptation – Use of local adaptation process • Characteristic of HVS – Different adaptation process depending on the real world and images on displays • Relation with surround background (a) (b) Fig. 4. The simultaneous contrast. 11 / 20

σ – Determination of ( x , y ) d • Use of the retinal response model – Input as the scene adaptation – Output as the display adaptation σ ( , ) x y σ = w ( x , y ) σ + α d ( x , y ) w • Match of the display adaptation level with similar narrow range – Expression of the very dark and light intensities in real world • Monotonic curve – Corresponding of Light scene adaptation to light display adaptation – Corresponding of dark scene adaptation to dark display adaptation I ( x , y ) I ( x , y ) Known value by using Gaussian filter = Known value by using Gaussian filter d w σ σ ( x , y ) ( x , y ) Scene adaptation half point level d w 12 / 20

– Determination of display luminance I ( x , y ) = w I ( x , y ) σ + α d ( x , y ) w Decision of these values are key point. Decision of these values are key point. 13 / 20

Scene adaptation Scene adaptation � Computation of display adaptation σ – Determination of ( x , y ) w • Surround image in Retinex with luminance channel – Kotera et al. based on Center/Surround method σ = ⊗ S ( x , y , ) G ( x , y ) Y ( x , y ) m m m { } ∫ ∫ = − + σ = 2 2 G ( x , y ) K exp ( x y ) / . G ( x , y ) dxdy 1 m m m » Banding artifact by using single Gaussian filter 14 / 20

– Substitution of Gaussian filter by bilateral filter 1 ∑ σ = − − ( s ) f ( p s ) g ( I I ) I w ∈ p s p p heigh ( s ) k ( s ) ∑ = − − k ( s ) f ( p s ) g ( I I ) ∈ p s p heigh ( s ) − ( I I ) Decrease of weight of pixels with Decrease of weight of pixels with − p s σ 2 − = 2 large luminance differences large luminance differences g ( I I ) e g p s over center pixel I s over center pixel I s =>avoiding banding or haloring =>avoiding banding or haloring Standard Gaussian filter Standard Gaussian filter 15 / 20

Experiment Experiment � Reproduction of RGB image γ ⎛ ⎞ RGB ( x , y ) ⎜ ⎟ = ⋅ w RGB ( x , y ) I ( x , y ) ⎜ ⎟ d d ⎝ ⎠ I ( x , y ) w γ Where, is a gamma parameter (0.5 to 1.0) γ – Resulting images with different γ = 0 . 5 , 0 . 8 , 1 . 0 Fig. 4. The results of Memorial Church for 16 / 20

α � Set of α = – with larger mean value than 0.5 1 α = – 0 . 1 with smaller mean value than 0.5 Problem with Fig. 6. The result of groveC by proposed method dark area 17 / 20 Fig. 7. The result of cathedralby proposed method

� Improvement of details in dark shadows – Gain parameter • Improvement of details and preservation of the light regions σ – As increase of , c ( x , y ) tends to be 1 w σ ( x , y ) = + ⋅ − 2 w c ( x , y ) 1 a exp( ( ) ) b I ( x , y ) = ⋅ w I ( x , y ) c ( x , y ) σ + α d ( x , y ) w Where, is an interactive constant by (0,0.1), and b a is set to 1 for HDR image, otherwise to 0. Fig. 8. The result of rosette proposed method. Better performance in dark area for (b). 18 / 20

� Comparison (a) (b) (c) (d) Fig. 9. (a) proposed method by Gaussian filter, (b) proposed method by bilateral filter, (c) Larson et al., and (d) Durand et al. Fig. 10. Comparison of small office. 19 / 20

Conclusion Conclusion � Simple static local adaptation method for HDR image compression – Basis on retinal model – Recreation of the same visual sensations between the real scene and the image on display – Use of bilateral filter to remove banding and halo effects 20 / 20