Understanding and Compression of BIG -Image, -Signal,...

18
SIGNAL PROCESSING LABORATORY / IOAN TABUS 1 Understanding and Compression of BIG -Image, -Signal, and –Data Professors: Ioan Tabus, Jaakko Astola, Jorma Rissanen Best existing compression of DNA sequences Human genome (3 billion bases): GeNML: 1.66 bits/base MAIN ACHIEVEMENTS: 1. Efficient Representations and Coding Lossy and lossless image coding state of the art for: Depth maps, stereo images, plenoptic images Biological sequence compression state of the art compressor: GeNML Lossless audio compression: state of the art compressor: OptimFrog Lossy audio coding (with Nokia Research Center) 2. Image and audio analysis MDL inference with sparse modelling Sensor array processing (with Patria) Image segmentation and object detection Best existing lossless compression of AUDIO files WITMSE 2010 Tampere PRESENT FOCUS AREAS: Image representation, coding, and analysis Image understanding Best existing compression of DEPTH-MAP images Best existing lossless compression of STEREO images Best existing lossless compression of PLENOPTIC images

Transcript of Understanding and Compression of BIG -Image, -Signal,...

SIGNAL PROCESSING LABORATORY / IOAN TABUS

1

Understanding and Compression ofBIG -Image, -Signal, and –Data

Professors:

Ioan Tabus, Jaakko Astola, Jorma Rissanen

Best existing compression of DNA sequencesHuman genome (3 billion bases): GeNML: 1.66 bits/base

MAIN ACHIEVEMENTS:1. Efficient Representations and Coding

Lossy and lossless image codingstate of the art for:

Depth maps, stereo images, plenoptic images

Biological sequence compressionstate of the art compressor: GeNML

Lossless audio compression:state of the art compressor: OptimFrog

Lossy audio coding (with Nokia Research Center)2. Image and audio analysis

MDL inference with sparse modelling

Sensor array processing (with Patria)

Image segmentation and object detection

Best existing lossless compression of AUDIO files

WITMSE 2010 Tampere

PRESENT FOCUS AREAS:Image representation, coding, and analysisImage understanding

Best existing compression of DEPTH-MAP images

Best existing lossless compression of STEREO images

Best existing lossless compression of PLENOPTIC images

SIGNAL PROCESSING LABORATORY / IOAN TABUS

2

SPARSE REPRESENTATIONS AND LOSSLESSCOMPRESSION OF PLENOPTIC IMAGESALLOWING FULL POST-PROCESSING FUNCTIONALITY

SIGNAL PROCESSING LABORATORY / IOAN TABUS1.12.2016

3

Current projects

q Image Representation, Analysis, and CompressionØ Sparse models and context models for image analysis and

compression (together with Prof. B.D. Rao, UCSD)Ø Pekka Astola: stereo imagesØ Petri Helin: plenoptic images

q 3D environment for intelligent manipulators (2015-16)Ø cooperation with Department of Intelligent Hydraulics

AutomationØ Pekka Astola: Scene analysis for intelligent manipulators

q D2I Digile Project (2015-16)Ø Ionut Schiopu: Pothole detection from video in moving cars

q TUT’s graduate school (2016-17)Ø Pekka Astola: Algebraic and combinatorial approaches for

index generating functions

SIGNAL PROCESSING LABORATORY / IOAN TABUS

4

Ø Conditional CERV is the proposed methodØ Compresses left disparity conditional on right disparityØ Gets LOSSLESS compression ratios 100-300 times

Conditional lossless compressionfor stereo disparity images

I.Tabus ”Conditional CERV”Signal Processing Letters, 2013

SIGNAL PROCESSING LABORATORY / IOAN TABUS

5

Ø Localizing for grasping non-planar objects without strong texture

Object pose estimation from stereo images (1)

P. Astola et. al. “Precise Outline Matching Criteriafor Target Pose Estimation and Odometry fromStereo Video”, 2016

SIGNAL PROCESSING LABORATORY / IOAN TABUS

6

Ø Estimation of the pose by outline matching algorithms

Object pose estimation from stereo images (2)

P. Astola et. al. “Precise Outline Matching Criteria for Target PoseEstimation and Odometry from Stereo Video”, 2016

SIGNAL PROCESSING LABORATORY / IOAN TABUS1.12.2016

7

PRESENT FOCUS AREA:Image representation, coding, and analysis

• P. Helin, P. Astola, B. Rao, I. Tabus “Sparse modelling and predictive coding of subaperture images forlossless plenoptic image compression”. 3DTV-Conference, pp1-4, Hamburg, Germany, 4-6 July 2016.Received the Best paper award of the conference.

• I.Tabus, ”Patch-Based Conditional Context Coding of Stereo Disparity Images” IEEE Signal ProcessingLetters, 21:10, pp. 1220-1224, Oct. 2014.

• I. Tabus, I. Schiopu, J. Astola, “Context coding of depth map images under the piecewise-constant imagemodel representation”. IEEE Trans. Image Processing, 22:11, pp. 4195–4210, Nov. 2013.

• I. Schiopu, I. Tabus, ”Lossy depth image compression using greedy rate-distortion slope optimization”.IEEE Signal Processing Letters, 20:11, pp. 1066-1069, Nov. 2013.

• I. Tabus and J. Rissanen, ”Time- and space-varying context tree models for image coding and analysis”,6th Work. Information Theoretic Methods in Science and Engineering, Tokyo, Aug. 26-29, 2013.

SIGNAL PROCESSING LABORATORY / IOAN TABUS

8

1. The main ingredients for modelling and compression

Ø Coding distributions tracked in the nodes of context trees (the Context algorithmintroduced by Rissanen in 1983)

Ø Arithmetic coding using the coding distributions (arithmetic coding introduced in1978)

Ø Context tree arithmetic coding – at the center of major standards for image coding(JPEG-2000) and all video coding standards

2. Statistical inference

Ø Minimum description length inference (introduced By Rissanen in the 1978Automatica paper “Modeling By Shortest Data Description”)

Ø Normalized Maximum Likelihood models (J. Rissanen, “Fisher information andstochastic complexity” 1996)

Theoretical background: Classical tools inInformation Theoretic Modelling(originating in work of Jorma Rissanen)

SIGNAL PROCESSING LABORATORY / IOAN TABUS1.12.2016

9

Stereo image warping and compression (1)

SIGNAL PROCESSING LABORATORY / IOAN TABUS

10

Stereo image warping and compression (2)

SIGNAL PROCESSING LABORATORY / IOAN TABUS1.12.201611

Depth image analysis

• I. Tabus and J. Rissanen, ”Time- and space-varying context tree models forimage coding and analysis”, 6th Workshop on Information TheoreticMethods in Science and Engineering, Tokyo, Aug. 26-29, 2013.

SIGNAL PROCESSING LABORATORY / IOAN TABUS1.12.201612

Lossless compression of depth images

Ø Track and encode contours of the constant regions in the imageØ Context tree coding using dynamical programming and semi-adaptivecontext treesØ Prediction in a graph of neighbor regionsØ Achieves compressions of 30 to 60 times, for high resolution images

•I. Tabus, I. Schiopu, J. Astola, “Context coding of depth map images under the piecewise-constant image model representation”. IEEE Trans. Image Processing, 22:11, pp. 4195–4210,Nov. 2013.

SIGNAL PROCESSING LABORATORY / IOAN TABUS1.12.201613

Lossy compression (1)

ØMerges in an optimal way the adjacent regionsØ Rate-distortion optimization by greedily maximizing the next slopeØ Absolute winner in the regions above 60 dB, with up to tends of dB overstandard lossy codersØ In the lower ranges differences of up to several dB over competitors

[ST13] I. Schiopu, I. Tabus, ”Lossy depth image compression using greedyrate-distortion slope optimization”. IEEE Signal Processing Letters, 20:11,pp. 1066-1069, Nov. 2013.

SIGNAL PROCESSING LABORATORY / IOAN TABUS

14

Lossy compression (2) :Ø GSO+CERV is the winner (magenta)Ø Four dots indicates images shown in next 4 slides

SIGNAL PROCESSING LABORATORY / IOAN TABUS

15

PSNR = 80dB BitRate = 0.22 bpp Number of Regions = 8460

Depth image with region contours superimposed Constant regions marked in pseudocolor

Lossy compression: reconstruction (3)

SIGNAL PROCESSING LABORATORY / IOAN TABUS

16

PSNR = 45dB BitRate = 0.05 bpp Number of Regions = 1642

Depth image with region contours superimposed Constant regions marked in pseudocolor

Lossy compression: reconstruction (4)

SIGNAL PROCESSING LABORATORY / IOAN TABUS

17

J. Hukkanen, P.Astola, I.Tabus, LOSSLESS COMPRESSION OF REGIONS-OF-INTERESTFROM RETINAL IMAGES, Euvip 2014.

LOSSLESS COMPRESSION OF REGIONS-OF-INTERESTFROM RETINAL IMAGES (1)

SIGNAL PROCESSING LABORATORY / IOAN TABUS

18

J. Hukkanen, P.Astola, I.Tabus, LOSSLESS COMPRESSION OF REGIONS-OF-INTERESTFROM RETINAL IMAGES, Euvip 2014.

LOSSLESS COMPRESSION OF REGIONS-OF-INTERESTFROM RETINAL IMAGES (2)