An Efficient Image Categorization Method With Insufficient Training Samples.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Convolutional Neural Networks (CNNs) achieve high performance with ample training data.
• Real-world applications often face challenges due to limited sample availability for training.
• Existing CNN methods may struggle to generalize effectively with insufficient data.

Purpose of the Study:

• To propose a novel method, Classifier method based on Variational Autoencoder (CFVAE), to enhance CNN performance under data scarcity.
• To improve the robustness and accuracy of image classification models when training samples are limited.

Main Methods:

• CFVAE combines a standard CNN as a prior classifier and a CNN-based Variational Autoencoder (VAE) as a posterior classifier.
• The prior classifier generates initial labels and latent variable distributions.
• The posterior classifier augments latent variables from regularized distributions, optimizing with a stochastic gradient variational Bayes method.

Main Results:

• The proposed CFVAE method demonstrates improved performance in image classification tasks with limited samples.
• Analysis using Hoeffding's inequality and Chernoff's bounding confirms the method's effectiveness.
• Experimental results show CFVAE surpasses state-of-the-art methods in accuracy.

Conclusions:

• CFVAE effectively addresses the challenge of limited samples in CNN-based image classification.
• The latent variable augmentation strategy enhances model generalization and predictive accuracy.
• The method offers a promising solution for practical applications requiring high performance with scarce data.