A Novel Deep Seed Generation Model for Graph Cut Segmentation

Abstract

We propose a novel convolutional neural network specialized for generating foreground/background seeds for Graph Cut segmentation. While seed generation was automated in prior work using deep models (e.g., TransUNet), such models were originally designed for general-purpose semantic segmentation and are not optimized for Graph Cut seed generation. On the basis of insights gained from our previous study, we introduce a lightweight architecture that combines a convolutional neural network (CNN) encoder with a low-frequency vision Transformer (LF-ViT) at the bottleneck to produce reliable seeds. The model is trained using 400 images generated by data augmentation from a single image of a lantana flower and evaluated on ten test images. Experimental results show that the proposed method improves Graph Cut performance over TransUNet-generated seeds (average mIoU from 0.8946 to 0.9699).