Energy Efficient Distributed JPEG2000 Image Compression in Multihop - - PowerPoint PPT Presentation

Energy Efficient Distributed JPEG2000 Image Compression in Multihop Wireless Networks

Huaming Wu & Alhussein A. Abouzeid

• Dept. of Electrical, Computer and Systems Engineering

Rensselaer Polytechnic Institute Troy, New York 12180, USA rpi.edu/~abouza/

2

Outline

Motivations Distributed Image Compression Energy Model Simulation Conclusion and Future Work

3

Motivations

Recently, Visual Sensor Network is emerging

for applications such as surveillance, environmental monitoring, security and interactive environments.

It consists of tiny wireless-enabled battery-

• perated cameras.

4

Challenges and Objective

Sensor networks will undergo a transition

similar to the Internet transition from text- based to multimedia.

Visual data incur high computation and

communication energy Sensors will remain relatively resource constrained

“divide and conquer” Distributed image compression enables the

sharing of computation load among sensors.

5

Assumptions

Nodes, some of which are camera-equipped Cluster-based routing mechanism Contention-free and error-free Session: a source sending one image to a

destination, in response to receiving a request from the destination

In the request, Q (bit rate of compressed

image) and L (wavelet decomposition level) are specified

6

Background on Image Compression

Objective: Reduce Redundancy JPEG2000: wavelet-based, error resilience,

progressive, multi-resolution

Wavelet-based image coding:

Forward Wavelet Transform Quantization Entropy Coding Entropy Decoding Dequantization Inverse Wavelet Transform (a) Encoder (b) Decoder

7

Wavelet Decomposition

Octave-band decomposition:

1D-DWT applied to vertical and horizontal

direction line by line: 2D-DWT.

The LL band is recursively decomposed, first

vertically, and then horizontally.

L H LL HL LH HH Image in spatial domain HL LH HH HL LH HH 1 level 2 level HL LH HH HL LH HH 3 level

8

Distributed Image Compression

Wavelet transform consumes most energy in

image compression.

Basic idea: distributing the workload of

wavelet transform to several groups of nodes along the path

Data (raw image or intermediate results

between decomposition levels) exchange is

• f key importance due the incurred wireless

communication energy

9

Data Exchange Method 1

Traditional data partitioning in

parallel wavelet transform

Data is divided in rows/columns Each node applies 1D-DWT No image quality loss, but 2D-

DWT needs twice data exchange

10

C1 distributes Ii to processing nodes C1 collects 1D-DWT results Qi

Example of Method 1

s C1 p11 p12 p13 p14 C2 p21 p22 p23 p24 d p31 p32 p33 p34 C4 C3 Compressed data Control data Raw image data Level 1 data Level 2 data Level 3 data Query and get node set info from cluster head Source distributes compressed rows Ri to processing nodes Sending level1 2D-DWT results Ji to C2 In this way, compressed image reaches d Repeat for LL subband of level 1 data and compress

• ther subbands to next cluster head

Repeat for LL subband of level 2 data and compress

• ther subbands to next cluster head

11

Data Exchange Method 2

Tiling: Node does 2D DWT independently Rate-distortion loss and blocking artifacts

increase with number of tiles

12

Example of Perceptual Image Quality with tiling

Image quality loss and

blocking artifacts are small if

• Number of tiles is

small or

• Not very low bit rate

Still applicable for

distributed image compression

Top left: Without tiling. 0.1bpp,PSNR=29.30dB Top right: Tile 64x64. 0.1bpp,PSNR=25.12dB

• Btm. left: Tile 256x256.

0.1bpp,PSNR=29.12dB

• Btm. right: Tile 64x64.

0.5bpp,PSNR =35.67dB

13

Example of Method 2

s C1 p11 p12 p13 p14 C2 p21 p22 p23 p24 d p31 p32 p33 p34 C4 C3 Compressed data Control data Raw image data Level 1 data Level 2 data Level 3 data S query and get processing node info from C1 S distributes tiles to processing nodes. Running 2D-DWT independently on each node. Send 2D-DWT results of each tile to next cluster head Repeat for LL subband of level 1 and compress

• ther subbands

Repeat for LL subband of level 2 and compress

• ther subbands

14

Other Issues

To save communication energy, entropy

coding is applied before data exchange

Randomly rotation of processing nodes in

each cluster among sessions.

15

Energy Model

Communication:

ETX=ee+eada (Transmission) Joule per bit ERX=ee (Receiving) ee: startup energy parameter ea: amplifier energy parameter a: path loss exponent d: distance between transmitter and receiver

Computation: (Estimated by JouleTrack on Jasper)

EDWT = ? (1 level of 2D-DWT) Joule per raw image bit EENT = d (Quantization and entropy coding)

JouleTrack: http://www-mtl.mit.edu/research/anantha/jouletrack/JouleTrack/index.html JasPer: http://www.ece.uvic.ca/~mdadams/jasper/

16

Metrics

Total energy: includes both computation and

communication energy

System lifetime: time when the first node in

the network fails due to depleted energy.

17

Simulations

500 nodes Transmission radius=10m JPEG2000 code (in C) from Jasper

18

Total Energy Consumption (1)

5 10 15 20 25 200 300 400 500 600 700 800 Distance between source and destination (hop) Normalized total energy consumption per raw image bit (nJ) Method 1 (L=1) Method 2 (L=1) Centralized (L=1) Method 1 (L=5) Method 2 (L=5) Centralized (L=5)

Total (comp.+comm) energy consumption per raw image bit versus distance between source and destination for different desired decomposition level L. Q=1bpp.

19

Total Energy Consumption (2)

5 10 15 20 25 200 300 400 500 600 700 800 Distance between source and destination (hop) Normalized total energy consumption per raw image bit (nJ) Method 1 (1bpp) Method 2 (1bpp) Centralized (1bpp) Method 1 (0.1bpp) Method 2 (0.1bpp) Centralized (0.1bpp)

Normalized total energy dissipation per raw image bit versus distance between source and destination for different Q. L=5.

20

System Lifetime (1)

50 100 150 200 250 300 350 400 450 500 1 2 3 4 5 6 7 8 Number of nodes System lifetime (session) Distributed (L=1) Centralized (L=1) Distributed (L=3) Centralized (L=3) Distributed (L=5) Centralized (L=5)

distributed (method2) versus centralized for different desired decomposition level L. Q=1bpp.

21

System Lifetime (2)

50 100 150 200 250 300 350 400 450 500 1 2 3 4 5 6 7 8 Number of nodes System lifetime (session) Distributed (1bpp) Centralized (1bpp) Distributed (0.5bpp) Centralized (0.5bpp) Distributed (0.1bpp) Centralized (0.1bpp)

System lifetime comparison: distributed versus centralized for different Q. L=5.

22

Conclusion

In terms of total energy consumption:

• Method 1 is much higher than the other two (method 2 and

centralized)

Method 2 is slightly higher than centralized image

compression

Method 2 extends the system lifetime by up to 4

times

Simple and easy to implement

23

Future Work

Impact of wireless link errors Effect of node failure Dynamic number of processing nodes Multipath routing

24

Error Robust Distributed Image Transmission

Sensor networks: error prone. Wireless link

errors and node failures. -> Need mechanisms to provide reliability

Distributed way is preferred for WSN Add spatial redundancy (e.g. FEC, multipath)

not temporal redundancy (e.g. ARQ) to image/video surveillance: real time applications.

25

Network Assumptions

Average wireless channel error probability: Pe Sensor node failure probability: P(off) No failure detection service to predict node

failure

Both can be modeled by a Markov chain:

Good “1” or bad “0” state for wireless channels On “1” or off “0” state for nodes

26

Error Robust Distributed Image Transmission

2 components: FEC-based unequal error

protection and path diversity

Choose Reed-Solomon (RS) code. UEP by

selecting different k for RS(n,k) code

Randomly choose multiple forwarding nodes

in a cluster

Combining multiple copies of coefficients

from different nodes

Information bits Redundancy bits

27

Example

C2 p20 p22 p24 p25 C3 p21 p23

X

Cluster head C3 gets level 2 data of tile 0 from p21 Cluster head C3 combines level 2 data of tile 1 from p24 and p25 Cluster head C2 sends 2 copies of level 1 data of tile 0 to p20 and p21 Cluster head C2 sends 2 copies of level 1 data of tile 1 to p24 and p25 p20 fails

28

Simulations

Image quality: PSNR Overhead: energy consumption per node 4 schemes:

(A) no error protection (B) only FEC code (C) only path diversity (D) our proposed scheme (FEC+multiple nodes)

29

Relative energy consumption

4 8 12 16 20 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5

Distance between source and destination (hop)

Scheme (A) Scheme (B) Scheme (C) Scheme (D)

30

Image quality vs. distance between source and destination

4 6 8 10 12 14 16 18 20 5 10 15 20 25 30 35 Distance between source and destination (hop) PSNR of received image (dB) Pe=0.001, P(off)=0.1 Scheme (A) Scheme (B) Scheme (C) Scheme (D)

31

Image quality vs. Pe

10

• 4

10

• 3

10

• 2

5 10 15 20 25 Average wireless channel error rate Pe PSNR of received image (dB) h=8, P(off)=0.1 Scheme (A) Scheme (B) Scheme (C) Scheme (D)

32

Image quality vs. P(off)

0.05 0.1 0.15 0.2 5 10 15 20 25 30 35 40 Average node failure rate P(off) PSNR of received image (dB) h=8, Pe=0.001 Scheme (A) Scheme (B) Scheme (C) Scheme (D)

33

Comparison of perceptual image quality

Scheme (A), Scheme (B) Scheme (C), Scheme (D) Pe=5 x10-3,P(off)=0.02, h=4.

34

Results

The difference between scheme (A) and (B) is very

• small. As well as the difference between scheme (C)

and (D). -> Impact of RS coding on energy consumption is small.

The normalized total energy consumption decreases

with the increase of h and almost becomes flat for large h. The energy consumed in image compression is distributed into more nodes for large h.

Low energy overhead: about 20% more than scheme

(A)

Image quality improvement: up to 10 dB and better

perceptual image quality

35

Thank You!