Noise Immunity and Robustness Study of Image Recognition Using a Convolutional Neural Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Convolutional neural networks (CNNs) face challenges in robustness and noise immunity.
• Understanding the impact of data uncertainty on CNN performance is crucial.

Purpose of the Study:

• To propose a technique for estimating CNN robustness and improving stability.
• To analyze the influence of training and testing dataset uncertainty on recognition probability.

Main Methods:

• Estimated recognition accuracies across datasets with varying uncertainties.
• Analyzed the relationship between training dataset uncertainty and recognition accuracy.
• Employed statistical modeling to determine optimal uncertainty levels.

Main Results:

• Demonstrated the existence of an optimal level of uncertainty in training data for improved recognition accuracy.
• Showcased that adding a specific amount of noise can enhance CNN noise immunity.
• Provided a method to determine this optimal uncertainty using statistical modeling.

Conclusions:

• Optimal uncertainty in training data can significantly improve CNN recognition quality and noise immunity.
• Statistical modeling is effective for identifying the optimal data uncertainty.
• This approach offers a practical method for enhancing CNN performance in uncertain environments.