IMAGE COMPRESSION Joseph RONSIN The Challenges 2 To increase compression performances To increase transmission channel capacity
Classical Coding Scheme 3 PREDICTION TRANSFORMATION QUANTIZATION CODING Spatial: DCT : Temporal: 01100101110 t Image Compression and Transmission 4
OUTLINE 5 Introduction Redundancy / Irrevelancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion IMAGE COMPRESSION Starting date for interest : 70's. 6 Digital T.V. Color - SD 720*576*25*8+2*(360*576*25*8)=19.8Mb/s High Definition Television HDTV 1920*1080*8*3*25 =1.2Gb/s ( = 3mn bluray) Fax : 2100 x 1728 pixels (of 1 bit) on a transmission line at 4.8 K bit/s 12 minutes Landsat : 6 spectral bands 6000 x 6000 x 8 bits 1.7 Gbits
7 Screen size Image format product Bit rate 2 QCIF [176 x 144 lines] Mobile phone 64 kb/s 4 CIF [352 x 288 lines] Mobile phone PDA 256 kb/s 8 CIF [352 x 288 lines] Mobile DVD 512 kb/s 27 SDTV [720 x 288 lines] Standard Television 2 Mb/s 40 HDTV [1920 x 1080 lines] Hi. Def. Television 8 Mb/s Products and compression with MPEG-4 AVC / H.264 OUTLINE 8 Introduction Redundancy / irrelevancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion
Redundancy and Irrelevancy 9 What is the value of the missing pixel? how critical is its exact reproduction??? 39 42 40 Image compression Because image data are often highly redundant and/or irrelevant then compression can be achieved 10 Redundancy relates to the statistical properties of the image we can use neighbours for reconstruction of this pixel Irrelevancy relates to an observer viewing an image observer will not perceive some reconstruction errors The redundancy or irrelevancy exists in spatial, spectral, and temporal forms.
Image compression 11 Definition: Compression or coding: Process of reducing the amount of data required to represent a given image. Lossless coding: using redundancies Lossy coding: using redundancies and irrelevancies Image compression Compression using 12 Coding redundancy solution: variable length coding. Interpixel redundancy (spatial, temporal, spectral). solution: transformation in a more efficient format that 2-D pixel. Psychovisual redundancy - Irrelevancy solution: quantization or mapping of a broad range of input values to a limited number of output values.
Statistical redundancies Example: observing the output of a predictor transmitting codes for prediction errors 13 Code for 0 Spatial redundancies Example: observing differences between 2 neighbours 14
Spectral redundancies Example: observing components 15 Temporal redundancies Example: observing difference pixels between 2 consecutive images 16
Psychovisual redundancy Human vision: Contrast sensitivity function-csf 17 MTF: Modulation Transfert Function - Spatial frequency + 18 Contrast + Maximum 2 5 cycles/degree
Contrast Sensitivity 19 The response of the eye to changes in the intensity of illumination is nonlinear: Weber s law If I = 100 then I=+/- 2 If I = 200 then I=+/-4 Contrast Sensitivity 20 0% 1% 2% 3% 4% Circle + I constant - I Background constant Just noticeable difference (JND) at 2%
OUTLINE Introduction Redundancy / irrevelancy 21 Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion 22 Data redundancy can be mathematically quantized: If for each pixel level m i with probability p i we associate a codeword n i Then mean lenght of code: n Efficiency N p i 1 i n i Redundancy n H min < 1 n n log q 1
Image compression compression ratio: original image with n 1 denoting the number of information carrying units (bits), compressed image with n 2... C R is called compression ratio: 23 n 1 C R n 2 range [ 2-10 000 ] OUTLINE 24 Introduction Redundancy / irrevelancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion
Image Compression Models Encoder: 25 source encoder: removing input redundancies. channel encoder: increasing noise immunity, protecting messages. It can be omitted Source encoder Channel encoder Channel Channel decoder Source decoder encoder decoder Source Encoder and Decoder Each operation is designed to reduce: 26 interpixel redundancies : Mapper, psychovisual redundancies : Quantizer, coding redundancies : Symbol encoder. I(x,y) Mapper Quantizer Symbol Encoder Channel Source encoder Channel Symbol Decoder Inverse Mapper I(x,y) decoder
OUTLINE 27 Introduction Redundancy / irrevelancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion Choosing a coding technique Lossy / Lossless technique 28 Sensitivity to input image types Operational bit rate Constant bit rate / constant quality Implementation PC based technology DSP ASIC's Encoder decoder asymetry broadcasting
Choosing a coding technique Channel error tolerance Artifacts 29 quality or impairments on edges, textures, blocking effect Effect of multiple coding accumulation of coding degradation Progressive transmission capability approximate image at low bit rate browsing in data bases details incrementally transmitted if desired System compatibility Standards Functionalities Choosing a coding technique Scalability 30 images obtained at different resolutions (spatial, temporal, SNR) using only part of coded bit stream. Bit stream composed with basic bit stream corresponding to minimal reconstruction quality enhancement bit streams using basic one as reference and offering incremental quality. C low resolution : basic bit stream C -1 + - C mean resolution : enhancement bit stream 1 C -1 + - high resolution : enhancement Image bit Compression stream 2CH 2
Choosing a coding technique 31 Embedded bitstream Embedded bitstream 32 Non embedded non-embedded non-embedded
Embedded bitstream 33 embedded Flux embedded Flux embedded OUTLINE 34 Introduction Redundancy / irrevelancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion
Image Quality Objective quality criteria: 35 MSE = 1 M N X(m,n) X(m,n) n1m 1 MN 2 PSNR = 10 log 10 2 S MSE S : maximum pixel value = 2 8-1 = 255 Image Quality 36 Subjective fidelity criterion Ratings by a number of human observers, based on typical decompressed images, are averaged to obtain this subjective fidelity criterion. Example: on a side-by-side comparison of the original image and the decompressed image: value -3-2 -1 0 1 2 3 ratings Much worse worse Slightly worse same Slightly better better Much better
Image Quality 37 CCIR Recommendation500-3 (CCIR, 1986). Note that CCIR is now ITU-R (International Telecommunications Union Recommendations). 38 1. Impairment is not noticeable 2. Impairment is just noticeable 3. Impairment is definitely noticeable, but not objectionable 4. Impairment is objectionable 5. Impairment is extremely objectionable New objective criterion of subjective quality: MOS for Mean Opinion Score,
Image Quality 39 MOS = 1.38 PSNR = 26.14 db original degraded MOS : Mean Opinion Score MOS = 4.51 PSNR = 24.09 db OUTLINE 40 Introduction Redundancy / irrevelancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion
coding tools 41 DPCM Transform Coding DCT P-frames Macroblocks B-frames Video Object Planes Multiple Reference Variable Block-size Motion Compensation Generic 1950 ~1985 1999 B-pictures Huffman Coding Hybrid Coding Block Motion Estimation Scene Adaptive Coder Motion Vector Prediction Interlace Object- Based Scalability Integer Transform Error Resilience Advanced Deblocking Filter OUTLINE 42 Introduction Redundancy / irrevelancy Images Properties of HVS Compression efficiency Compression models Choosing a coding technique Image quality Coding tools Conclusion
Conclusion 43 Image Quality must be observed Performances of a coding technique must be given with their context Scalability Source and channel coding