Deep one-class probability learning for end-to-end image classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• One-class learning is crucial for novelty, anomaly, and outlier detection.
• Existing methods often struggle with end-to-end training and indirect decision-making.
• Deep networks offer potential but require specialized training strategies for one-class problems.

Purpose of the Study:

• To develop a deep, end-to-end binary image classifier for one-class learning.
• To enable direct classification without relying on negative training samples.
• To improve the adaptation of positive sample distributions and classification accuracy.

Main Methods:

• Designed a convolutional neural network (CNN) for direct binary image classification.
• Established a probabilistic model driven by network-derived energy to learn positive sample distributions.
• Utilized a novel particle swarm optimization (PSO) algorithm for sampling and distribution estimation during optimization.

Main Results:

• The proposed method directly outputs classification results, eliminating the need for post-processing thresholding or estimation.
• Direct optimization of the deep network via the probabilistic model enhances positive distribution adaptation.
• Experimental results demonstrate the method's effectiveness and state-of-the-art performance.

Conclusions:

• The developed deep end-to-end classifier offers a direct and effective solution for one-class learning tasks.
• The integration of a probabilistic model and PSO-based sampling overcomes key challenges in distribution estimation.
• This approach advances anomaly and outlier detection capabilities in machine learning.