Authors:
Tetsu Matsukawa
1
and
Takio Kurita
2
Affiliations:
1
University of Tsukuba, Japan
;
2
National Institute of Advanced Industrial Science and Technology, Japan
Keyword(s):
Scene classification, higher order local autocorrelation features, bag-of-features, posterior probability image
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Computer Vision, Visualization and Computer Graphics
;
Data Manipulation
;
Feature Extraction
;
Features Extraction
;
Health Engineering and Technology Applications
;
Human-Computer Interaction
;
Image and Video Analysis
;
Informatics in Control, Automation and Robotics
;
Methodologies and Methods
;
Neurocomputing
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition
;
Physiological Computing Systems
;
Sensor Networks
;
Signal Processing, Sensors, Systems Modeling and Control
;
Soft Computing
;
Statistical Approach
Abstract:
This paper presents scene classification methods using spatial relationship between local posterior probabilities of each category. Recently, the authors proposed the probability higher-order local autocorrelations (PHLAC) feature. This method uses autocorrelations of local posterior probabilities to capture spatial distributions of local posterior probabilities of a category. Although PHLAC achieves good recognition accuracies for scene classification, we can improve the performance further by using crosscorrelation between categories. We extend PHLAC features to crosscorrelations of posterior probabilities of other categories. Also, we introduce the subtraction operator for describing another spatial relationship of local posterior probabilities, and present vertical/horizontal mask patterns for the spatial layout of auto/crosscorrelations. Since the combination of category index is large, we compress the proposed features by two-dimensional principal component analysis. We confirm
ed the effectiveness of the proposed methods using Scene-15 dataset, and our method exhibited competitive performances to recent methods without using spatial grid informations and even using linear classifiers.
(More)