Information or data fusion is one of the solutions adopted for improving the performance of a pattern recognition system. Information can be gathered either from multiple data sources or through the use of multiple representations generated from a single data source. A single representation summarizes the information and provides a single cue on the data, and thus may not be able to fully reveal the inherent characteristics of the data. In visual recognition, image representations are generally categorized into global and local based types. A global representation captures features corresponds to some holistic characteristic in the image, and produces a coarse representation. Differently, a local representation reveals detail variations and traits inherent to the image. Psychological findings have shown that humans equally rely on both local and global visual information. Moreover, there is a large agreement in literature that the combination of different features, i.e. a multiview perspective, can have a positive effect on the performance of a pattern recognition system. In fact, different features can represent different and complementary characteristics of the data; in other words, each feature set represent a different view on the original dataset. Thus it is expected that a visual recognition system can benefit from different representations (both local and global) through the use of information fusion. Information can generally be consolidated at three different levels: (i) decision level; (ii) match score level; and (iii) feature level. In the literature match level and decision level fusion (i.e. combining the output of different classifier, each of them working on different feature sets) have been extensively studied, whereas feature level fusion is a relatively understudied problem because of the difficulties inherent to its correct implementation. Feature level fusion may incorporate redundant, noisy or trivial information and the concatenated feature vectors may lead to the problem of curse of dimensionality. In addition, the feature sets may not be compatible and relationship between different feature spaces may not be known. Moreover, this integration comes at a cost, which may incur in units of time, computational resources or even money. Nevertheless, it is thought that fusing features at this level would still retain a richer source of discriminative information. Motivated by the belief, this thesis investigates the use of feature level fusion and its correlation with feature selection and classification tasks for two recent pattern recognition problems. These include the classification of six types of HEp2 staining patterns and the automatic verification of kinship relations in a pair of face images. several image attributes are proposed, that are better capable of characterizing the different kind of images associated with the two said classification tasks. Feature level fusion of the different attributes is performed followed by a careful reduction of features, through the use of pertinent feature selection and classification algorithms, that decide the most representative and discriminative feature sets for the patterns to classify. Results indicate that the proposed techniques working on the combination of features of different natures, which are capable of describing the data under different perspectives, is an effective strategy in achieving higher accuracy.
Feature Fusion for Pattern Recognition / UL-ISLAM, Ihtesham. - (2015).
Feature Fusion for Pattern Recognition
UL-ISLAM, IHTESHAM
2015
Abstract
Information or data fusion is one of the solutions adopted for improving the performance of a pattern recognition system. Information can be gathered either from multiple data sources or through the use of multiple representations generated from a single data source. A single representation summarizes the information and provides a single cue on the data, and thus may not be able to fully reveal the inherent characteristics of the data. In visual recognition, image representations are generally categorized into global and local based types. A global representation captures features corresponds to some holistic characteristic in the image, and produces a coarse representation. Differently, a local representation reveals detail variations and traits inherent to the image. Psychological findings have shown that humans equally rely on both local and global visual information. Moreover, there is a large agreement in literature that the combination of different features, i.e. a multiview perspective, can have a positive effect on the performance of a pattern recognition system. In fact, different features can represent different and complementary characteristics of the data; in other words, each feature set represent a different view on the original dataset. Thus it is expected that a visual recognition system can benefit from different representations (both local and global) through the use of information fusion. Information can generally be consolidated at three different levels: (i) decision level; (ii) match score level; and (iii) feature level. In the literature match level and decision level fusion (i.e. combining the output of different classifier, each of them working on different feature sets) have been extensively studied, whereas feature level fusion is a relatively understudied problem because of the difficulties inherent to its correct implementation. Feature level fusion may incorporate redundant, noisy or trivial information and the concatenated feature vectors may lead to the problem of curse of dimensionality. In addition, the feature sets may not be compatible and relationship between different feature spaces may not be known. Moreover, this integration comes at a cost, which may incur in units of time, computational resources or even money. Nevertheless, it is thought that fusing features at this level would still retain a richer source of discriminative information. Motivated by the belief, this thesis investigates the use of feature level fusion and its correlation with feature selection and classification tasks for two recent pattern recognition problems. These include the classification of six types of HEp2 staining patterns and the automatic verification of kinship relations in a pair of face images. several image attributes are proposed, that are better capable of characterizing the different kind of images associated with the two said classification tasks. Feature level fusion of the different attributes is performed followed by a careful reduction of features, through the use of pertinent feature selection and classification algorithms, that decide the most representative and discriminative feature sets for the patterns to classify. Results indicate that the proposed techniques working on the combination of features of different natures, which are capable of describing the data under different perspectives, is an effective strategy in achieving higher accuracy.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/11583/2592755
Attenzione
Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo