基于DeepLab v3+的涂鸦式图像分割算法  

Graffiti-based Image Segmentation Algorithm Based on DeepLab v3+

作　　者：俞颖晖 洪茂雄 Yu Ying-hui;Hong Mao-xiong(School of Computer Science and Mathematics,Fujian University of Technology,Fuzhou 350118,Fujian Province,China)

机构地区：[1]福建理工大学计算机科学与数学学院,福建福州350118

出　　处：《科学与信息化》2025年第2期95-97,共3页Technology and Information

摘　　要：在现有的基于深度学习的交互式图像分割算法的研究中,主要以点击以及边界框的交互方式为主。本文在Deep GrabCut算法的基础上,选择DeepLab v3+作为模型的架构,并提出了“米”字形采样策略,经过大量的训练,最终生成的模型能够很好地适应涂鸦的交互方式。在分割精度上比原方法提升了5%以上,并有效地简化了用户交互要求,拓展了基于深度学习的交互式图像分割技术在涂鸦交互方式上的研究。In the current research on interactive image segmentation algorithms based on deep learning,the main interaction methods are clicks and bounding boxes.This paper,based on the Deep GrabCut algorithm,selects DeepLab v3+as the model architecture and proposes a“rice”character-shaped sampling strategy.After extensive training,the final model can well adapt to the graffiti interaction method.The segmentation accuracy is improved by more than 5%compared to the original method,and the user interaction requirements are effectively simplified,expanding the research on interactive image segmentation technology based on deep learning in the graffiti interaction method.

关 键 词：深度学习 交互式图像分割 Deep GrabCut DeepLab v3+ 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]