A New Method of Deep Convolutional Neural Network Image Classification Based on Knowledge Transfer in Small Label...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning image classification faces challenges with large datasets and scarce annotated samples.
• Overfitting is a significant issue in neural network training when using limited data.

Purpose of the Study:

• To address the problem of deep learning image classification with limited annotated data.
• To develop a method that effectively utilizes a large number of unlabeled images alongside a small set of labeled ones.
• To mitigate overfitting in deep neural network training.

Main Methods:

• Constructed a support vector machine (SVM) expert labeling system using a bag-of-words model for feature extraction from limited labeled samples.
• Automatically annotated a large unlabeled image dataset using the developed SVM system.
• Created an augmented training set by combining original labeled and automatically labeled samples.
• Developed a deep convolutional neural network (CNN) model trained on the augmented dataset, incorporating knowledge transfer from SVMs.

Main Results:

• Successfully implemented knowledge transfer from SVMs to deep neural network classifiers.
• Solved the overfitting problem in neural network training with small sample sizes.
• Demonstrated the method's performance through experimental verification on the Caltech-256 public dataset.

Conclusions:

• The proposed method effectively enhances deep learning image classification accuracy with limited annotated data.
• The integration of SVM-based automated labeling and CNNs provides a robust solution for data augmentation and overfitting prevention.
• This approach offers a viable strategy for real-world image classification tasks where manual annotation is resource-intensive.