comparative performance evaluation of image descriptors over ieee 802.11b noisywireless networks
DESCRIPTION
We evaluate the image retrievalprocedure over an IEEE 802.11b Ad Hoc network, operatingin 2.4GHz, using IEEE Distributed Coordination FunctionCSMA/CA as the multiple access scheme. IEEE 802.11 is awidely used network standard, implemented and supportedby a variety of devices, such as desktops, laptops, notebooks,mobile phones etc., capable of providing a variety of differentservices, such as file transfer, internet access et.al. Therefore,we consider IEEE 802.11b being a suitable technology toinvestigate the case of conducting image retrieval over awireless noisy channel. The model we use to simulate the noisyenvironment is based on the scenario in which the wirelessnetwork is located in an outdoor noisy environment, or inan indoor environment of partial LOS - Line-of-sight power.We used a large number of descriptors reported in literaturein order to evaluate which one has the best performance interms of Mean Average Precision - MAP values under thosecircumstances. Experimental results on known benchmarkingdatabase show that the majority of the descriptors appear tohave decreased performance when transferred and used in suchnoisy environments.TRANSCRIPT
Comparative Performance Evaluation of Image Descriptors Over IEEE 802.11b Noisy Wireless Networks
Presented by Evaggelos Chatzistavros
Evaggelos Chatzistavros, Savvas A. ChatzichristofisGeorgios Stamatellos, Kostantinos Zagoris
Retrieval over noisy channel. Why?
Image retrieval
over noisy wireless link
Image retrieval
Extended use of IEEE 802.11b
networksWireless link
- signal degradation
Content Based Image Retrieval
Image retrieval applications
Image retrieval based on images
stored or taken by a user’s ubiquitous
device
Retrieval performed using an
Internet connection, via a 3G or a wireless network (IEEE
802.11b)
• Any technology that in principle help you to organize digital image archives based on their visual content.
• Anything ranging from a simple similarity matching metric to a robust image annotation engine can fall under the purview of CBIR.
IEEE 802.11 Wireless Networks
Ad- Hoc Mode
Infrastructure Mode
Problems with wireless link
Signal Interference
Signal Fading BER
• Interference: Presence of many different wireless transmitters at the same frequency
• Fading: Multiple copies of the same signal reach the receiver
• BER: Bit Error Rate, affected by the above factors, has an effect on transmitted data
Configuration of experiments
Retrieval
WANG database
9 different descriptors
Experiments conducted using img(Rummager)
Network
IEEE 802.11b, 2.4GHz, 11Mbps
2 wireless static nodes
UDP - User Datagram Protocol as the transportation protocol
Different channel conditions caused by different BER values, based on Rician fading model
Rician fading
Retrieval scenario 1• Two wireless nodes• IEEE 802.11 b, 11Mbps• 200 – 250 m distance between nodes• Transport protocol: U D P – User Datagram Protocol
• BER values: 0, 0.001, 0.01, 0.1, 0.17
• Descriptor is transferred between nodes • Increased BER values lead to increased number of errors (bits) at the descriptor• Retrieval is conducted based on corrupted descriptor
Retrieval scenario 1 (2)D
escr
ipto
rs • CEDD• FCTH• EHD• CLD• SPCD• RGB• Tamura• Autocorrelograms
Met
rics • MAP – Mean
average precision
• Normalized MAP – N MAP
Retrieval scenario 1 – Simulation results
MAP over different BER
All descriptors achieve high scores for BER values up to
0.01
EHD appears to be the most resistant to noise since it
maintains its value for BER 0.17
Autocorrelograms’ performance degrades by
91.66% for BER 0.17, due its large size
Retrieval scenario 2• Two wireless nodes• IEEE 802.11 b, 11Mbps• 200 – 250 m distance between nodes• Transport protocol: U D P – User Datagram Protocol
• BER values: 0, 0.001, 0.01, 0.1, 0.17
• Image is transferred between nodes • Increased BER values lead to increased number of errors (bits) at the image• Retrieval is conducted based on corrupted image
Retrieval scenario 2 (2)D
escr
ipto
rs • CEDD• FCTH• EHD• CLD• SPCD• RGB• Tamura• Autocorrelograms• TOP - SURF
Met
rics • MAP – Mean
average precision
• Normalized MAP – N MAP
Retrieval scenario 2 – Simulation results
MAP over different BER
CLD appears to be the most insensitive to channel errors since its retrieval score is reduced only
by 0.77% for 0.17 BER value
Tamura’s performance degrades by 74.71% for BER 0.17
Top Surf: fairly resistant to noise, since its performance degrades by 13.19% and 5.37% for 10000 and 20000 visual words respectively
N – MAP descriptor valuesN - MAP: MAP per Byte
(Normalized)
All descriptors from scenario 1 and CEDD from
scenario 2
CLD has the best N – MAP value. Each byte carries the
higher amount of valid information per byte
Tradeoff between a descriptor’s size and the degradation of its performance
Conclusions – Future workConclusions
For the first configuration, BER=0.01 is a threshold, after which the majority of descriptors cannot maintain high retrieval scores
For the second configuration, most descriptors maintain their retrieval scores up to BER=0.17, with CLD, Autocorrelograms and TOP-SURF
having the best performance
CLD maintains a higher N-MAP value, even under very noisy channel
conditions
Future Work
Extended scenario, with mobile nodes and Access
Points (connected to a server), where retrieval can
take place
Descriptor evaluation over heterogeneous networks
감사합니다
Thank you
ありがとう
Ευχαριστώ πολύ!