## Using Discriminant Eigenfeatures for Image Retrieval (1996)

Citations: | 392 - 13 self |

### BibTeX

@MISC{Swets96usingdiscriminant,

author = {Daniel L. Swets and John Weng},

title = { Using Discriminant Eigenfeatures for Image Retrieval},

year = {1996}

}

### Years of Citing Articles

### OpenURL

### Abstract

This paper describes the automatic selection of features from an image training set using the theories of multi-dimensional linear discriminant analysis and the associated optimal linear projection. We demonstrate the effectiveness of these Most Discriminating Features for view-based class retrieval from a large database of widely varying real-world objects presented as "well-framed" views, and compare it with that of the principal component analysis.

### Citations

2785 | Eigenfaces for recognition
- Turk, Pentland
- 1991
(Show Context)
Citation Context ...s type of framework, a training phase finds salient features to use in the subsequent recognition phase of the system. These types of approaches can deal directly with complex, real-world images [14] =-=[20] [21]-=- because the system is general and adaptive. The efficient selection of good features, however, is an important issue to consider [2]. A well-known problem in pattern recognition is called "the c... |

2649 |
Introduction to Statistical Pattern Recognition”, 2nd edition
- Fukunaga
- 1990
(Show Context)
Citation Context ...= P m i=1 y i v i ; where the v i 's are the column vectors of V . Let the effectiveness of the approximation be defined as the mean-square error kX \GammasX(m)k 2 . Then we can use the proven result =-=[6]-=- [10] [12] that the best vectors v 1 ; v 2 ; \Delta \Delta \Delta vm to use are the unit eigenvectors associated with the m largest eigenvalues of the covariance matrix of X, \Sigma X = E [(X \Gamma M... |

892 |
Nearest neighbor pattern classification
- Cover, Hart
(Show Context)
Citation Context ...o find a the set of k nearest neighbors for retrieval. It has been shown that the probability of error for this nearest neighbor decision rule is bounded above by twice the Bayes probability of error =-=[4]-=- if we have an infinite number of samples. This simple measure of similarity is used because it does not require estimation of the distribution function, which is impractical for our high dimensional ... |

638 | View-based and modular eigenspaces for face recognition
- Pentland, Moghaddam, et al.
- 1994
(Show Context)
Citation Context ...n-Lo`eve projection to produce a set of Most Expressive Features (MEFs), and a subsequent discriminant analysis projection to produce a set of Most Discriminating Features (MDFs). In this work, as in =-=[16] and [14],-=- we require "well-framed" images as input for training and query-by-example test probes. By well-framed images we mean that only a small variation in the size, position, and orientation of t... |

538 |
Sirovich 'Application of the Karhunen-Loeve Procedure for the Charac:erisation of Hunan Face
- Kirby, L
(Show Context)
Citation Context ...ese y i 's give the minimum mean-square error, we call them the Most Expressive Features (MEF) in that they best express the population in the sense of linear transform as evidenced in reconstruction =-=[11]-=-. This projection, also called the Karhunen-Lo`eve projection and principal component analysis [8], has been used to represent [11] and recognize [20] [16] face images, and for planning the illuminati... |

208 | Probabilistic visual learning for object detection. Paper presented at the The - Moghaddam, Pentland - 1995 |

160 |
ªOVID: Design and Implementation of a Video Object Database System,º
- Oomoto, Tanaka
- 1993
(Show Context)
Citation Context ... a host of problems that alphanumeric information systems were never designed to handle [1]. A central task of these multimedia information systems is the storage, retrieval, and management of images =-=[15]-=-. In many cases, the operator would like to base this retrieval on objects contained in the images of the database. As such, content-based image retrieval is fundamentally an object recognition proble... |

141 | Face recognition from one example view
- Beymer, Poggio
(Show Context)
Citation Context ...aches can deal directly with complex, real-world images [14] [20] [21] because the system is general and adaptive. The efficient selection of good features, however, is an important issue to consider =-=[2]. A well-k-=-nown problem in pattern recognition is called "the curse of dimensionality"---more features do not necessarily imply a better classification success rate. For example, principal component an... |

91 |
The statistical utilization of multiple measurements
- Fisher
- 1938
(Show Context)
Citation Context ...ermine the projection matrix W that maximizes the ratio detfS b g detfSwg : In other words, we want to maximize the between-class scatter while minimizing the within-class scatter. It has been proven =-=[5]-=- [22] that this ratio is maximized when the column vectors of projection matrix W are the eigenvectors of S \Gamma1 w S b associated with the largest eigenvalues. Then the scalar components in Z are f... |

80 | Nonlinear manifold learning for visual speech recognition
- Bregler, S
- 1995
(Show Context)
Citation Context ...etter classification success rate. For example, principal component analysis, also known as the Karhunen-Lo`eve projection and "eigenfeatures," has been used for face recognition [20] and li=-=p reading [3]-=-. An eigenfeature, however, may represent aspects of the imaging process which are unrelated to recognition, such as the illumination direction. An increase or decrease in the number of eigenfeatures ... |

73 |
Dubes, Algorithms for Clustering Data, Prentice-Hall, Englewood Cli s
- Jain, C
- 1988
(Show Context)
Citation Context ...hat they best express the population in the sense of linear transform as evidenced in reconstruction [11]. This projection, also called the Karhunen-Lo`eve projection and principal component analysis =-=[8]-=-, has been used to represent [11] and recognize [20] [16] face images, and for planning the illumination of objects for future recognition tasks [14]. To determine m, the number of features to use, we... |

70 |
A Visual Information Management System for the Interactive Retrieval of Faces
- Bach, Paul, et al.
- 1993
(Show Context)
Citation Context ...of digital library technology [9]. The complexity in the very nature of two-dimensional image data gives rise to a host of problems that alphanumeric information systems were never designed to handle =-=[1]-=-. A central task of these multimedia information systems is the storage, retrieval, and management of images [15]. In many cases, the operator would like to base this retrieval on objects contained in... |

49 |
Automatic Generation of Object Recognition Programs
- Ikeuchi, Kanade
- 1988
(Show Context)
Citation Context ...t recognition problem. The research emphasis to this end has historically been on the design of efficient matching algorithms from a manually designed feature set with hand-crafted shape rules (e.g., =-=[7]-=-). Hand-crafted shape rules can exploit the efficiency found in manually tuning features for a particular training image set. However, these rules have a severe limitation on the type of object classe... |

40 |
Indexing in Video Databases
- Hampapur, Jain, et al.
- 1995
(Show Context)
Citation Context ...ility of computers to rapidly and successfully retrieve information from image databases based on the objects contained in the images has a direct impact on the progress of digital library technology =-=[9]-=-. The complexity in the very nature of two-dimensional image data gives rise to a host of problems that alphanumeric information systems were never designed to handle [1]. A central task of these mult... |

32 |
Illumination Planning for Object Recognition in Structured Environments
- Murase, Nayar
- 1994
(Show Context)
Citation Context ...n this type of framework, a training phase finds salient features to use in the subsequent recognition phase of the system. These types of approaches can deal directly with complex, real-world images =-=[14] [20]-=- [21] because the system is general and adaptive. The efficient selection of good features, however, is an important issue to consider [2]. A well-known problem in pattern recognition is called "... |

19 | Learning Recognition and Segmentation Using the Cresceptron - Weng, Ahuja, et al. - 1993 |

17 | Genetic algorithms for object recognition in a complex scene
- Swets, Punch, et al.
- 1995
(Show Context)
Citation Context ...roblem in general. Techniques have been proposed to produce these types of images, using, for example, pixel-to-pixel search [20], hierarchical coarse-to-fine search [21], or genetic algorithm search =-=[18]-=-. This reliance on wellframed images is a limitation of the work; however, there are application domains where this limitation is not overly intrusive. In image databases, for example, the human opera... |

5 | A system for combining traditional alphanumeric queries with content-based queries by example in image databases. Multimedia Tools and Application
- Swets, Pathak, et al.
- 1996
(Show Context)
Citation Context ...e the capability indicated by the labels given to the training images. Each stored image can maintain pointers to a relational database to provide retrievals under various other desired organizations =-=[17]-=- such as gender, age group, etc. For this experiment, the database consisted of predominantly pairs of images to describe each class. Most classes in the database were represented by two images, and 1... |

3 |
Principal Component Analysis.Springer-Verlag
- Jolliffe
- 1986
(Show Context)
Citation Context ...m i=1 y i v i ; where the v i 's are the column vectors of V . Let the effectiveness of the approximation be defined as the mean-square error kX \GammasX(m)k 2 . Then we can use the proven result [6] =-=[10]-=- [12] that the best vectors v 1 ; v 2 ; \Delta \Delta \Delta vm to use are the unit eigenvectors associated with the m largest eigenvalues of the covariance matrix of X, \Sigma X = E [(X \Gamma MX )(X... |

3 |
Probability Theory.Princeton
- Lo`eve
- 1955
(Show Context)
Citation Context ... y i v i ; where the v i 's are the column vectors of V . Let the effectiveness of the approximation be defined as the mean-square error kX \GammasX(m)k 2 . Then we can use the proven result [6] [10] =-=[12]-=- that the best vectors v 1 ; v 2 ; \Delta \Delta \Delta vm to use are the unit eigenvectors associated with the m largest eigenvalues of the covariance matrix of X, \Sigma X = E [(X \Gamma MX )(X \Gam... |

2 | Efficient image retrieval using a network with complex neurons
- Swets, Weng
- 1995
(Show Context)
Citation Context ...g on a Sun SPARC 20. When this Most Discriminating Feature subspace computation is placed into a hierarchy and the resulting spaces decomposed into a hierarchical Voronoi tessellation as described in =-=[19]-=-, the average time required for a test probe fell to 9.1 seconds; at the same time, the recognition rate rose to 95% for an image from the correct class being retrieved as the top choice and 99% for t... |

1 |
Affiliation of Authors Daniel L. Swets is with the computer science department at Augustana College
- Wilks, Wiley, et al.
- 1963
(Show Context)
Citation Context ...ne the projection matrix W that maximizes the ratio detfS b g detfSwg : In other words, we want to maximize the between-class scatter while minimizing the within-class scatter. It has been proven [5] =-=[22]-=- that this ratio is maximized when the column vectors of projection matrix W are the eigenvectors of S \Gamma1 w S b associated with the largest eigenvalues. Then the scalar components in Z are featur... |