You have trained a custom object detection model to detect defects in manufactured parts. The model has been trained on a diverse set of images but is still underperforming in detecting subtle defects. What strategies would you employ to enhance the model's ability to detect these subtle defects, including potential modifications to the training data and model architecture?