Overcoming Catastrophic Forgetting in Neural Network During Continuous Learning: A Selective Layer Freezing Approach for Crack Detection

This paper investigates the phenomenon of catastrophic forgetting in continuous learning, focusing on crack detection problems. We propose a specific configuration of the neural network involving selective layer freezing to mitigate this issue. Our study uses the ResNet18 neural network and evaluates it on two databases featuring distinct types of surface cracks. Continuous learning scenarios are simulated by introducing various crack types to the system over time. Results demonstrate that by configuring ResNet18 with specific frozen layers, catastrophic forgetting can be overcome, leading to reduced computational demands during continuous learning.