Streaming View Classification With Noisy Label.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Dynamic scenes in image processing generate data views in a streaming manner.
• Existing streaming view learning methods assume accurate labels, which is often not true in real-world applications.
• Noisy labels from initial views degrade classification performance in dynamic environments.

Purpose of the Study:

• To address the challenge of simultaneous view evolving and label ambiguity in streaming data.
• To propose a novel method for streaming view classification that can handle noisy labels.

Main Methods:

• Introduced Streaming View Classification with Noisy Label (SVCNL) method.
• Calibrated noisy labels based on emerging new views to reflect data dynamics.
• Utilized a label transition matrix and graph embedding for progressive noisy label correction.

Main Results:

• The proposed SVCNL method effectively calibrates noisy labels in streaming view learning.
• The approach accurately reflects dynamic data changes by adapting to new views.
• Extensive experiments and theoretical analyses demonstrate the method's effectiveness and generalization bounds.

Conclusions:

• SVCNL offers a robust solution for streaming view classification problems with noisy labels.
• The method successfully handles both evolving data views and label ambiguity.
• This work provides a significant advancement for real-world image processing tasks involving dynamic, streaming data.