Related Experiment Videos
A robust classifier combined with an auto-associative network for completing partly occluded images
Takashi Takahashi1, Takio Kurita
1Department of Applied Mathematics and Informatics, Ryukoku University, Ootsu, Shiga 520-2194, Japan. takataka@math.ryukoku.ac.jp
Summary
This study introduces a novel classifier robust to image occlusions. By integrating an auto-associative network, it reconstructs occluded image parts, enabling accurate object classification even with significant data loss.
Area of Science:
- Computer Vision
- Machine Learning
- Image Processing
Background:
- Image occlusions present a significant challenge in object recognition tasks.
- Existing classifiers often suffer performance degradation when faced with incomplete visual data.
Purpose of the Study:
- To develop a robust image classifier that remains effective despite occlusions.
- To integrate an auto-associative network for image reconstruction within a classification framework.
Main Methods:
- Proposed a method integrating an auto-associative network with a simple classifier.
- Utilized the auto-associative network to detect occluded regions and reconstruct them with recalled pixels.
- Employed an iterative reconstruction process for enhanced robustness.
Main Results:
- Demonstrated the classifier's ability to handle occluded input images effectively.
- Achieved robust classification performance without performance decrease even with up to 30% of face images occluded.
- Validated the approach through experiments on face image classification.
Conclusions:
- The proposed integrated network offers a robust solution for image classification under occlusion.
- The method successfully reconstructs occluded image regions, preserving classification accuracy.
- This approach significantly enhances the reliability of classifiers in real-world scenarios with occluded objects.