Related Experiment Video
Updated: May 29, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
Multilinear supervised neighborhood embedding of a local descriptor tensor for scene/object recognition
Xian-Hua Han1, Yen-Wei Chen, Xiang Ruan
1College of Information Science and Engineering, Ritsumeikan University, Kusatsu, Japan. hanxhua@fc.ritsumei.ac.jp
Summary
This study introduces a new image recognition method using local descriptor tensors and multilinear supervised neighborhood embedding (MSNE). This approach enhances subject and scene recognition by efficiently extracting robust features, outperforming existing models.
Area of Science:
- Computer Vision
- Machine Learning
- Pattern Recognition
Background:
- Traditional image recognition methods struggle with variations in illumination and efficient feature representation.
- Existing bag-of-feature models can be computationally intensive and less effective in combining local features.
Purpose of the Study:
- To propose a novel image representation and feature extraction framework for improved subject and scene recognition.
- To develop a method robust to illumination variations and more efficient than current models.
Main Methods:
- Introduced a novel feature extraction approach: histogram of orientation weighted with a normalized gradient (NHOG) for robust local region representation.
- Developed a local descriptor tensor framework to efficiently combine multiple local features, offering an alternative to bag-of-feature models.
- Applied a multilinear supervised neighborhood embedding (MSNE) algorithm to directly process the local descriptor tensor for discriminant feature extraction.
Main Results:
- The proposed NHOG feature extraction is robust to significant illumination variations.
- The local descriptor tensor framework provides efficient image representation, outperforming the bag-of-feature model.
- The MSNE algorithm effectively extracts compact and discriminant features while preserving neighborhood structures in tensor-feature space.
Conclusions:
- The proposed approach demonstrates superior performance in subject and scene recognition tasks.
- The method shows significant advantages over existing techniques on various benchmark datasets, including scene, face, and object recognition datasets.