Suppressing catastrophic forgetting in deep-learning-based imaging through scattering media.

Area of Science:

• Computer Vision
• Machine Learning
• Optics

Background:

• Scattering media like fog obscure objects.
• Deep learning reconstructs objects from scattered images (speckle images).
• Continual learning in deep learning models leads to catastrophic forgetting.

Purpose of the Study:

• Investigate methods to suppress catastrophic forgetting in deep-learning-based scattering imaging.
• Evaluate the effectiveness of rehearsal and elastic weight consolidation (EWC) methods.

Main Methods:

• Defined Task 1 (KMNIST character classification) and Task 2 (MNIST digit classification).
• Sequentially trained a model on KMNIST then MNIST.
• Applied rehearsal (reusing past data) and EWC (preserving parameters) to mitigate forgetting.

Main Results:

• Sequential training caused Task 1 accuracy to drop from 85.2% to 65.2%.
• Rehearsal method maintained Task 1 accuracy at 85.7%.
• EWC method achieved Task 1 accuracy between 80.4% and 86.07%.

Conclusions:

• Catastrophic forgetting is a significant issue in continual learning for scattering imaging.
• Both rehearsal and EWC methods are effective in mitigating catastrophic forgetting.
• These methods enable robust performance in deep-learning-based scattering imaging models over time.