Deeply Encoding Stable Patterns From Contaminated Data for Scenery Image Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models achieve competitive performance in scene categorization but rely on accurate image labels.
• Real-world datasets often contain noisy labels due to external predictors, limiting model robustness.
• Existing models are too restrictive by assuming 100% correct image-level labels.

Purpose of the Study:

• To propose a new deep architecture for robust scene categorization despite noisy image labels.
• To develop a method that hierarchically derives stable templates using a generative model.
• To enhance the accuracy and reliability of AI systems in recognizing diverse sceneries.

Main Methods:

• Constructing a semantic space using subspace embedding of image-level labels.
• Utilizing a probabilistic generative model to learn stable templates from superpixel distributions.
• Developing a novel aggregation network to concatenate CNN features for deep representation.
• Integrating learned representations into an image kernel for multiclass SVM classification.

Main Results:

• The proposed method demonstrates effective scene categorization even with contaminated image labels.
• Stable templates learned by the generative model remain consistent under significant label noise (up to 36%).
• Experiments show the superior performance of the developed deep architecture in distinguishing scene categories.

Conclusions:

• The novel deep architecture effectively handles noisy image labels in scene categorization tasks.
• The hierarchical derivation of stable templates offers a robust approach to improve AI system reliability.
• This method advances AI's ability to recognize complex sceneries under realistic data conditions.