SS-HCNN: Semi-Supervised Hierarchical Convolutional Neural Network for Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning for image classification faces challenges with large-scale annotated data availability and uneven data category separability.
• Existing methods struggle to efficiently utilize unlabeled data and handle complex class distributions.

Purpose of the Study:

• To propose a novel Semi-Supervised Hierarchical Convolutional Neural Network (SS-HCNN) to address data annotation and separability limitations in image classification.
• To develop an unsupervised clustering technique for hierarchical data organization and feature learning.

Main Methods:

• Implemented a large-scale unsupervised maximum margin clustering technique to iteratively split images into hierarchical clusters.
• Learned cluster-level CNNs at parent nodes and category-level CNNs at leaf nodes using feature similarity for grouping.
• Introduced a novel cluster splitting criterion for automatic termination of image clustering in the hierarchy.

Main Results:

• SS-HCNN trained with partial labeled data achieved performance comparable to fully supervised CNNs using all labeled data.
• SS-HCNN trained with all labeled data significantly outperformed other fully supervised CNNs.
• The method effectively relieves constraints of uneven data separability and data annotation.

Conclusions:

• The proposed SS-HCNN effectively addresses key challenges in deep learning for image classification.
• This approach demonstrates superior performance and efficiency, especially in scenarios with limited labeled data.
• SS-HCNN offers a promising direction for advancing large-scale image classification.