As a machine learning engineer optimizing an image classification system for surveillance, which must handle high-resolution images under varied lighting conditions, which strategy best improves model accuracy while maintaining manageable computational load?