1 Resolution Loss without Optical Blur Tali Treibitz Alex Golts Yoav Y. Schechner Technion , Israel 14 total intensity I L object airlight A D direct transmission scattering D I ( x, y) Lobject ( x, y) t ( z) A( x, y) A 1 e z 0 object radiance 1 e z z Schechner, Narasimhan, Nayar 0 z 15 Haze I ( x) l object object Schechner et al., Applied Optics ‘03 ( x)t ( x) A( x) transmittance airlight 17 Pointwise Degradations noise I ( x ) l object t( x) A( x ) n( x ) object pointwise attenuation: additive component: vignetting reflection atmosphere attenuation glare path radiance Treibitz & Schechner, Recovery Limits in Pointwise Degradations 18 Pointwise Degradations reduce SNR even if known I ( x) l object t( x) A( x ) noise n( x ) object pointwise attenuation additive (positive) bias Treibitz & Schechner, Recovery Limits in Pointwise Degradations 19 Noise: Object size matters? 20 Noise: Object size matters? 0.5 21 Noise: Object size matters? 1 22 Noise: Object size matters? 1.5 Visibility Under Noise depends on: noise level object background intensity difference object size quantify this dependency ! Prior art: resolution limits due to optical blur here: no optical blur 23 Previous criteria • Is there something there? • Is it a tank? • What type is it? Johnson charts: detection tank 0.75 orientation recognition 1.2 3.5 minimum line pairs for 50% success identification 7 NIIRS- National Image Interpretability Rating Scales 0 Interpretability of the imagery is precluded by obscuration, degradation, or very poor resolution 1 Detect a medium-sized port facility and/or distinguish between taxi-ways and runways at a large airfield . 2 Detect large hangars at airfields. Detect large static radars (e.g., AN/FPS-85, COBRA DANE, PECHORA, HENHOUSE), Detect military training areas... 3 Identify the wing configuration (e.g., straight, swept, delta) of all large aircraft (e.g., 707, CONCORD, BEAR, BLACKJACK) ... 9 Identify small light-toned ceramic insulators that connect wires of an antenna. Identify vehicle registration numbers (VRN) on trucks . Identify screws and bolts on missile components ... 24 pattern visible Treibitz & Schechner, Recovery Limits in Pointwise Degradations 25 Where is the Cutoff? pattern visible calculated analytically! pattern invisible 0.5 Input SNR | S (u ) | noise 10 0.1 u (frequency) 0.5 Treibitz & Schechner, Recovery Limits in Pointwise Degradations 26 Cutoff Per Success Rate success rate 50% 0.5 Input SNR | S (u ) | noise 1 10 0.1 ucutoff u (frequency) 0.5 Treibitz & Schechner, Recovery Limits in Pointwise Degradations 28 Noise Suppression in the HVS response of receptive field Theoretical Neuroscience, Dayan & Abbott low noise high noise frequency (cycles/degree) We derive: fundamental analytical model Model: simple linear denoising not a denoising method 29 SNR Improvement by Averaging signal HW SNR output - SNR change after averaging C(u,W) = input SNR noise Treibitz & Schechner, Recovery Limits in Pointwise Degradations /W 30 Different Sizes of Windows too big for signal too small for noise Treibitz & Schechner, Recovery Limits in Pointwise Degradations 31 Averaging by Optimal Window Treibitz & Schechner, Recovery Limits in Pointwise Degradations 32 SNR Improvement by Averaging C max (u ) output SNRmax (u) SNRinput depends on frequency! u same plot for a Gaussian filter Treibitz & Schechner, Recovery Limits in Pointwise Degradations 33 Output SNR SNRoutput 1 SNRoutput 0.5 0.32 Input SNR | S (u ) | noise 1 6.5 0.1 ucutoff u (frequency) Treibitz & Schechner, Recovery Limits in Pointwise Degradations 0.5 34 Cutoff Per Success Rate success rate 70% success rate 40% 0.32 Input SNR | S (u ) | noise 6.5 0.1 ucutoff u (frequency) Treibitz & Schechner, Recovery Limits in Pointwise Degradations 0.5 35 Vision Success is Probabilistic visible invisible SNR=2/3 SNR determines chances of visibility Treibitz & Schechner, Recovery Limits in Pointwise Degradations 36 Success within a Confidence Interval Object is visible if N ( x) S N S prob N ( x ) S …depends on SNR and.. - success rate 68% 0.68 2 0 2 N ( x) SNR Treibitz & Schechner, Recovery Limits in Pointwise Degradations 25 Success within a Confidence Interval visibility is kept if edge keeps sign prob N background ( x ) N object ( x ) S depends on SNR noisy clear background object pixel pixel what is the probability for correct detection? - success rate %(sign kept) - %(wrong sign) 0.5 SNR Treibitz & Schechner, Recovery Limits in Pointwise Degradations 37 Determining Resolution Limits cutoff for ρ=70% success 0.32 system input SNR | S (u ) | noise 6.5 0.1 ucutoff frequency 0.5 Treibitz & Schechner, Recovery Limits in Pointwise Degradations 38 Pointwise Degradations noise I ( x ) l object t( x) A( x ) n( x ) pointwise attenuation: additive component: vignetting reflection atmosphere attenuation glare haze Treibitz & Schechner, Recovery Limits in Pointwise Degradations 39 Noise Model 2 2 I / g 2 Nikon D100 2 200 I photon noise dominates Treibitz & Schechner, Recovery Limits in Pointwise Degradations 9 Photon (shot) Noise Photon either { Electrons e or nothing Schechner 10 Photon (shot) Noise Photons either { Electrons e e e nothing Schechner 41 SNR per size (frequency) l l object l background background l object SNR = |S| , S l Treibitz & Schechner, Recovery Limits in Pointwise Degradations object l background t ( x) 42 Resolution Limits in Haze minimal visible object size[m] reciprocal to ucutoff limit due to atmosphere limit due to pixel size distance [km] Treibitz & Schechner, Recovery Limits in Pointwise Degradations 43 Cutoff Per Success Rate success rate 50% 0.5 Input SNR | S (u ) | noise 1 10 0.1 ucutoff u (frequency) 0.5 Treibitz & Schechner, Recovery Limits in Pointwise Degradations 44 Haze in the Galilee raw frame average of 50 frames limit due to noise and not blur Treibitz & Schechner, Recovery Limits in Pointwise Degradations 45 What now? What are the reconstruction limits? What is the minimal detectable object size? What camera noise properties are acceptable for detection? … 46 Imaging in Haze Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? Haze Through a Polarizer 47 best polarized image Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? Haze Through a Polarizer 48 best polarized image increased exposure time single frame- used by photographers Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? Dehazing using a Polarizer 49 best polarized image worst polarized image post-processing 2 frames two frames- Schechner et al. Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? Is it worth using a polarizer? 50 rarely! under the constraint of equal acquisition time unpolarized image best polarized image goal: object detection local contrast stretch- OK Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? post-processing 2 frames 51 Using a Single Polarized Image I min Best polarized image Imin (x, y ) D ( x, y ) 1 p A( x, y ) 2 2 Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? p 0,1 degree of polarization 52 SNR Comparison I unpolarized Ibest Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? 53 A Single Saturated Frame SNRpolarized SNRunpolarized p maximal value in nature Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? 54 SNR Comparison I unpolarized Ibest equal acquisition time acquisition time = exposure time X number of frames 55 Same Total Acquisition Time SNRpolarized SNRunpolarized p p<0.4 in our experiments Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? maximal value in nature 56 Experiment Wide field of view I best I unpolarized average of 2 frames same total acquisition time Treibitz & Schechner, Polarization- Beneficial for Visibility Enhancement? 57 SNR Comparison I unpolarized dehazing 1 p p 1 p p equal acquisition time optimal exposures tbest tworst 58 Advantages of Polarization • contrast stretch in non-uniform distances • restoring color • compensating for attenuation distance map 59 Limits in Pointwise Degradations • Freq cutoff – due to noise – without imaging blur • Relation between cutoff and success rate • Application: limits in pointwise degradations • Case study of performance trade-offs