Distributed ContentBased Visual Information Retrieval System on PeertoPeer Networks

1
Distributed Content-Based Visual Information
Retrieval System on Peer-to-Peer Networks

• Irwin Kin, Cheuk Hang Ng, and Ka Cheung Sia
• The Chinese University of Hong Kong
• In ACM Transactions on Information Systems,
  V22N3July04P477-501
• http//www.cse.cuhk.edu.hk/miplab/discovir/
• By Pruet_at_DSSG.CS.UMB

king04distributed
2
Agenda

• Motivation
• DISCOVIR architecture
• Peer Clustering Based on Image Similarity
• Firework Query Model
• Experiments and Results
• Conclusion and Comment

3
Motivation

• Currently, most content-based image retrieval
  (CBIR) systems are centralized, both on
  computational and storage.
• P2P should give some advantages to CBIR, eg,
  larger and more diversify image collection,
  better scalability, better performance and
  responsiveness.
• Problem with using P2P in CBIR
• No centralize feature extraction, so some
  standard has to be applied.
• No centralize image vector storage, so some
  distributed and search mechanism has to be
  applied.
• Image vector feature is large, so some bandwidth
  optimization has to be applied when exchanging
  vector data.
• Because of bandwidth requirement, query flooding
  model for traditional P2P (eg. Gnutella) is not
  suitable.

4
Motivation (cont.)

• Some approaches for solving the query flooding
• CAN/Chord use DHT to distributed index among
  peers, the data is convert to hash and
  distributed in circular space(Chord)/multi-dimensi
  on Cartesian space(CAN).
• Crespo uses routing cache to keep previous query
  result, so the entries in cache will be used to
  assist in forwarding new queries to peers that
  are supposed to contain the target data.

5
DISCOVIR architecture

• The authors propose new P2P architecture targets
  CBIR, called DIStributed Content-based Visual
  Information Retrieval (DISCOVIR).
• Based on modification made on Gnutella network,
  DISCOVIR is compatible with the Gnutella
  protocol, with some additional types of messages.
• Each peer has their own image collection, the
  image feature is extracted from local image
  collection using pluggable feature extraction
  module, and the image feature is kept in local
  database.
• Image query is based on example image (so, QBE
  Query By Example approach). The query peer has to
  extract and send to its neighbors (using Gnutella
  protocol).
• The other peers uses distance measure to find a
  set of similar images and return the result back
  to the query peer likewise, these peers will
  porpage the query to their connecting peers.

6
DISCOVIR architecture (cont.)
7
DISCOVIR architecture (cont.)

• Flow of Operations
• Preprocessing
• Feature extractor module can be loaded from
  DISCOVIR central website by Plug-in Manager and
  installed in local system. Then, feature
  extractor extracts features and pass feature
  vector to Image Indexer which will index the
  feature vector and keep in local index storage.
• Connection Establishment
• Connection Manager asks the Bootstrap Server at
  the first time that this peer joins the network.
  Then, the peer can hooks up to the DISCOVIR
  network via available peers using information
  from Bootstrap Server.

8
DISCOVIR architecture (cont.)

• Flow of Operations (cont.)
• Query Message Routing
• When user submits query of an image, Feature
  Extractor instantly extract feature from that
  image and construct a query message and send out
  through Packet Router.
• When other peers receive the query message, they
  need to perform two operations
• Local Index Look Up - searches for similar images
  from local index using Image Indexer
• Query Message Propagation - Packet Router uses
  Gnutella mechanism for forwarding messages, TTL
  and Replicated message checking.
• Query Result Display
• When the query result returns to query peer, user
  will obtain a list of location and size of
  matched images. Then, user can retrieve images
  via HTTP Agent and the image will be displayed on
  the User Interface.

9
Peer Clustering Based on Image Similarity

• To solve the query flooding problem, or brute
  force search problem, the peers in P2P network
  has to be clustered based on image similarity.
• On top of the P2P network, an overlay network of
  connections, called attractive links, groups
  similar peers together.
• Instead of using feature vector of every images,
  a signature value of image collection in each
  peer is used to determine the similarity between
  two peers.
• Some definition

10
Peer Clustering Based on Img. Sim. (cont.)

• Definition 1. is the set
  of n images shared by peer p. Image feature of
  each image in the collection is extracted and map
  into a d-dimensional vector (R) by function f as,
  . Therefore, each peer will contain
  a set of vector
• Definition 2. is defined as
  where and are the mean and variance
  of the vector collection.
• Definition 3. is defined as the
  Cartesian distance between two peers signature
  values and using following
  equation

11
Peer Clustering Based on Image Similarity (cont.)

• Based on the definitions, the attractive link can
  be assigned to group of similar peers using these
  steps
• Signature Value Calculation - every peer
  calculates its signature value, .
• Neighborhood Discovery - After a new peer joins
  the network, it broadcasts a signature query
  message. This broadcasting also be repeated in a
  regular interval.
• Similarity Calculation and Attractive Link
  Establishment - After acquiring the signature
  values of other peers, the peer can find the peer
  other peer with signature value closet to its
  signature value using and make an
  attractive connection to link them up.

12
Peer Clustering Based on Img. Sim. (cont.)
13
Firework Query Model

• In this query routing model, a query message is
  routed selectively according to the content.
• When it reaches its designated cluster, based on
  similarity, the query message is broadcast by
  peers through the attractive connections inside
  the cluster.
• So, when each peer receive a query, it needs to
  carry out two steps
• Shared File Look Up - This will compare query
  feature vector with feature vector of each image
  in local collection, if any image matched, it
  will reply to query peer.
• Route Selection - The peer calculates the
  similarity between the query and its signature
  value. If the similarity is more than threshold,
  it will send the query to the peers connected by
  attractive link (explosion), otherwise, it will
  forward the query to P2P connected peer.

14
Firework Query Model (cont.)
15
Firework Query Model (cont.)

• For preventing infinite query looping, replicated
  message checking rule and TTL are used.
• When a query appears to a peer, it is checked
  against a local cache for duplication, if found,
  the query is dropped.
• Each time the query passes through a peer, the
  TTL is decreased by one. Once the TTL reaches
  zero, the query is dropped. However, if the query
  is passed along an attractive link, the TTL value
  may not be decreased based on a probability
  called Chance-To-Survive (CTS).

16
Experiments and Results

• Performance Metrics
• Recall
• Recall RA/RT RA retrieved relevant images,
  RT total relevant images in the network. Higher
  is better.
• Query scope
• Visited Vpeer/Tpeer Vpeer number of peers
  that received and handled the query and Tpeer
  total number of peers in network. Lower is better
• Query efficiency
• Efficiency Recall/Visited
• The experimental result will compare with query
  flooding algorithm (Breath-First Search (BFS)) .
• Platform Sun Blade 1000 2GB Ram, Solaris v.8
  C, for simulate 20,000 peers and TTL 7 with 10
  iterations (queries), it took 3 Hrs.

17
Experiments and Results (cont.)

• Data Set
• Synthetic data - 100 sets with random mean and
  variance. For each set 100 points (images) are
  generated according to Gaussian distribution.
• Real data - 10,000 images from 100 categories in
  Corel Draws Image Collection CD.

18
Experiments and Results (cont.)

• Experiments 1 Performances affected by different
  number of peers.
• Experiments setup
• Number of peers 2,000 - 20,000.
• Network diameters 9 - 11
• Average distance 5.36 - 6.58
• Number of images assigned to each peers 100
  images
• Feature vector dimensions 9
• TTL FQM 5 BFS 7

19
Experiments and results (cont.)
Recall vs. peers
Recall
Number of Peers
20
Experiments and results (cont.)
Query Scope vs. peers
Query Scope
Number of Peers
21
Experiments and results (cont.)
Efficiency vs. peers
Recall/Query Scope
Number of Peers
22
Experiments and results (cont.)

• Experiments 2 Performances affected by TTL
• Experiments setup
• Number of peers 10,000
• Network diameters 10
• Average distance 6.2
• Number of images assigned to each peers 100
  images
• Feature vector dimensions 9
• TTL 4-9

23
Experiments and results (cont.)
Recall vs. TTL
Recall
TTL value of query message
24
Experiments and results (cont.)
Query Scope vs. TTL
Query Scope
TTL value of query message
25
Experiments and results (cont.)
Efficiency vs. TTL
Recall/Query Scope
TTL value of query message
26
Conclusion and Comment

• FQM outperforms BFS in all tests.
• FQM can reduce the network traffic cost (query
  scope) while able to maintain high query
  efficiency.
• Comment
• Random query routing, kind of BFS.
• Broadcasting when join, update, and explosion.
• Costly for dynamic network
• Complete feature vector is sent, so more traffic
  when using high-dimensional feature vector.
• Still log(n)