Graph Cut-based automatic image segmentation driven by seeds from deep semantic segmentation models

抄録

This paper proposes a unified framework for Graph Cut-based automatic image segmentation driven by seeds from deep semantic segmentation models. Graph Cut is a high-precision image segmentation method, whose performance has traditionally depended on manually specified object and background seeds. Our group previously proposed an automatic seed generation method using U-Net [17]. Although more advanced architectures, such as Transformer-based and pretrained models, have since been developed, their effectiveness and generalization for automatic seed generation, as well as suitable architectural designs, remain insufficiently explored. In this study, we propose a unified framework for automatic seed generation across diverse deep segmentation architectures, including convolutional, transformer-based, and pretrained models. Using representative architectures such as U-Net, TransUNet, Swin-Unet, and Convolutional Transformer (FCT), we analyze how structural differences affect seed generation performance, providing practical guidance for model selection under extremely small-data conditions and demonstrating the applicability of the proposed methodology. Experiments were conducted under four scenarios: images of the same species with the same color, the same species with different colors, different flower species, and the PASCAL VOC dataset [26]. The results show that TransUNet combined with Graph Cut (TU+GC) achieved the highest mean IoU (up to 0.98 in the flower dataset) and demonstrated promising generalization even under limited data conditions. This study suggests that combining CNN and Transformer architectures, leveraging pre-trained models, and incorporating diverse background conditions are effective design elements for automatic seed image generation under extremely small data settings. It also demonstrates that integrating Graph Cut with deep learning is effective for weakly supervised segmentation.