Neural Network Training for Semantic Segmentation Boundary Accuracy

The patent divides the training process into two distinct phases: initial training using conventional loss functions (dice coefficient loss or IOU) to establish baseline segmentation performance, and subsequent additional training using boundary-specific loss functions (boundary dice coefficient loss or boundary IOU) to specifically improve boundary line prediction accuracy. This segmentation of training stages allows each phase to optimize for its specific goal without interfering with the other.

The patent applies different loss functions to different regions of the image: conventional loss functions are applied to the entire image during initial training, while boundary-specific loss functions are applied specifically to boundary regions during additional training. This local differentiation ensures that boundary regions receive specialized attention and optimization, improving boundary prediction accuracy without compromising overall segmentation performance.

The patent performs preliminary training using conventional loss functions before conducting additional training with boundary-focused images. This preliminary action establishes a solid baseline model that already possesses general segmentation capabilities, so the subsequent additional training only needs to refine boundary-specific features rather than learn from scratch, thereby reducing the relative time cost of additional training.

The patent maintains continuous training action by seamlessly transitioning from initial training to additional training without interrupting the learning process. The neural network model continues to learn and improve throughout both stages, ensuring that the useful action of training is uninterrupted and cumulative, which maximizes efficiency and minimizes total training time.