Transcript Lecture 6: Multimedia IR: Indexing and Searching

Special Topics in Computer Science
Advanced Topics in Information Retrieval
Lecture 6 (book chapter 12):
Multimedia IR:
Indexing and Searching
Alexander Gelbukh
www.Gelbukh.com
Previous Chapter: Conclusions
 Basically, images are handled as text described them
 Namely, feature vectors (or feature hierarchies)
 Context can be used when available to determine features
 Also, queries by example are common
 From the point of view of DBMS, integration with IR
and multimedia-specific techniques is needed
 Object-oriented technology is adequate
2
Previous Chapter: Research topics
 How similarity function can be defined?
 What features of images (video, sound) there are?
 How to better specify the importance of individual
features? (Give me similar houses: similar = size?
color? strructure? Architectural style?)
 How to determine the objects in an image?
 Integration with DBMSs and SQL for fast access and
rich semantics
 Integration with XML
 Ranking: by similarity, taking into account history, profile
3
The problem
 Data examples:
 2D/3D color/grayscale images: e.g., brain scans, scientific
databases of vector fields
 (2D) video,
 (1D) voice/music; (1D) time series: e.g.,
financial/marketing time series; DNA/genomic databases
 Query examples:
 find photographs with the same color distribution as this
 find companies whose stock prices move as this one
 find brain scans with a texture of a tumor
 Applications: search; data mining
4
Solution
 Reduce the problem to search for multi-dimensional
points (feature vectors, but vector space is not used)
 Define a distance measure
 for time series: e.g., Euclidean distance between vectors
 for images: e.g., color distribution (Euclidean distance);
another approach: mathematical morphology
 Other features as vectors
 For search within distance, the vectors are organized
in R-trees
 Clustering plays important role
5
Types of queries
 All within given distance
 Find all images that are within 0.05 distance from this one
 Nearest-neighbor
 Find 5 stocks most similar to IBM
 All pairs within given distance
 Further: clustering
 Whole object vs. sub-pattern match
 Find parts of image that are...
 E.g., in 512  512 brain scans, find pieces similar to the
given 16  16 typical X-ray of a tumor
 Like passage retrieval for text documents
6
Neighbor and pairs types of queries




The objects are organized in R-trees
For neighbor queries: branch-and-bound algorithm
For pairs: recently discovered algorithms
These types of queries are not discussed here
7
Desiderata for a method
 Fast
 No sequential search with all objects
 Correct
 100% recall
 Precision is less important, though kept low. False alarms
are easy to discard manually
 Little space overhead
 Dynamic
 easy to insert, delete, update
8
Types of methods
 Linear quadtrees
 Complexity = hypersurface of the query region
 Grows exponentially with dimensionality
 grid-files
 Complexity grows exponentially with dimensionality
 R-trees methods, such as R*-trees
 Most used due to lower complexity
9
R-tree
 Objects and parts of images represented as Minimal
Bounding Rectangle (MBR)
 Can overlap for different objects
 Larger objects contain smaller objects
 MBRs are nested
 MBRs are arranged into a tree
 In storage, an index of disk blocks is maintained
 Disk blocks are fetched at once at hardware level
 For better insertion/deletion, tight MBRs are needed
 Good clustering is needed
10
File structure of R-tree
 Corresponds to disk blocks
 Fanout = 3: number of parts to group
11
R-tree
R-tree
12
Search in R-tree
Range queries:
find objects within distance  from query object




= Find MBRs that intersect with query’s MBR
Determine MBR of the query
Descend the tree
Discarding all MBRs that do not intersect with the
query’s MBR
Many variations of R-tree method have been proposed
13
Indexing
Only consider here whole match queries


Given collection of objects and distance function
Find objects within given distance  from given object Q
 Problems:
1. Slow comparison of two objects
2. Huge database
 GEMINI approach


GEneric Multimedia object INdexIng
Attempts to solve both problems
14
GEMINI indexing
 Quick-and-dirty test to quickly discard bad objects
 Uses clusters to avoid sequential search
 Quick test
 Single-valued feature, e.g., average for series.
Averages differ much  objects differ much
 Not vice-versa. False alarms are OK
 Several features, but fewer than all data. E.g., deviation
for series
15
Algorithm
 Map the actual objects into f-dimensional feature
space
 Use clusters (e.g., R-trees) to search
 Retrieve objects, compute the actual distances, and
discard false alarms
16
17
Feature selection
 Features should reflect distances
 Allow no misses (100% recall)
 features should make things look closer
 Lower Bound lemma:




If distance in feature space  actual distance
then 100% recall
(we speak about whole-match queries)
Holds for distance search, nearest-neighbor, pair search
18
Algorithm (more detail)





Determine distance
Choose features
Prove that distance in feature space  for actual objects
Use quick method (R-tree) to search in feature space
For found objects, compute the actual distances (this
can be expensive)
 Discard false alarms
 objects with greater actual distances, even if in feature space
the distance is OK
 Example: similar averages, but different series
19
Discussion
 The method does NOT improve quality
 Provides SAME quality as sequential search, but faster
 Distance definition requires domain/application expert
 How much do the two images differ?
 What is important/unimportant for the specific application?
 Feature selection requires a good knowledge engineer
 Choose the most characteristic feature: discriminative
 If needed, choose the second best, etc.
 Good features should be orthogonal: combination adds info
20
Example: Time series
 In yearly stock movements, find ones similar to IBM
 Distance: Euclidean (365-D vectors); others exist
 Features:
 First feature is average.
 If needed, Discrete Fourier Transform (DFT) coefficients
 Or, Discrete Cosine Transform, waivelet Transform, etc.
 Lower-bound lemma:




Parseval theorem: DFT preserves distances (DCT, WT too)
First several coefficients give  distance
Transforms “concentrate energy” in the first coefficients
Thus, the more realistic prediction of distance
21
Time series: Applications
 Such feature selection is effective for many skewed
spectrum distributions
 Colored noises: the energy decreases as F–b




b = 0: white spectrum: unpredictable. Method useless.
b = 1: pink noise: works of art
b = 2: brown noise: stock movements
b > 2: black noise: river levels, rainfall patterns
 The greater b the better the first coefficients of the
transform predict the actual distance
 Some other n-D signals show similar properties
 JPEG compression ignores higher coefficients
22
Time series: Performance
 Fewer features  more false alarms  time lost
 More features  more complex computation
 Optimal number of features proves to be about 1..3
 for skewed enough distributions
 JPEG compression shows that photographs have it
23
Time series: Sub-pattern search
 Use sliding window
 Encode each window with few features
24
Example: Color images
 Give me images with a texture of tumor like this one
 Give me images with blue at top and red at bottom
 Handles color, texture, shape, position, dominant
edges
25
Color images: Color representation
 Compute color histogram
 Distance: use color similarity matrix
 Very expensive computationally: cross-talk between
features (compare all to all features)
26
27
Color images: Feature mapping
 The GEMINI question again: What single feature is
the most representative?
 Take average R, G, B
 Lower-bound?
 Yes: Quadratic Distance Bounding theorem
28
Automatic feature selection





Features can be selected automatically
In texts: Latent semantic indexing (LSI)
Many methods
Principle components analysis (= LSI), ...
In fact, they can reduce features, but not define them
 Of colors, one can select characteristic combinations
 But not classify into faces and flowers
 So description of the objects is still on human researchers
29
Research topics




Object detection (pattern and image recognition)
Automatic feature selection
Spatial indexing data structures (more than 1D)
New types of data.
 What features to select? How to determine them?
 Mixed-type data (e.g., webpages, or images with
sound and description)
 What clustering/IR methods are better suited for
what features? (What features for what methods?)
 Similar methods in data mining, ...
30
Conclusions
 How to accelerate search? Same results as sequential
 Ideas:
 Quick-and-dirty rejection of bad objects, 100% recall
 Fast data structure for search (based on clustering)
 Careful check of all found candidates
 Solution: mapping into fewer-D feature space
 Condition: lower-bounding of the distance
 Assumption: skewed spectrum distribution
 Few coefficients concentrate energy, rest are less important
31
Thank you!
Till Tuesday 11, 6 pm
32