CPS-Net: Segmentation Network Based on Feature Fusion and Siamese

Traditional deep convolutional neural networks are weak in handling multi-scale issues in tooth CT image segmentation. Based on the evolution of siamese network vision and image cutting algorithms, we make progress when teeth are occluded or deformed. Thus, a siamese network image segmentation algorithm combined with feature fusion is proposed. Due to network degradation and gradient dispersion, different backbone networks are selected for distillation experiments for overly complex networks. The siamese model constructs a feature pyramid, improving the overall model's ability. The new method is better at handling multi-scale features of tooth deformation.