交互式持续学习的图像目标语义感知分割

Image segmentation is a key step from image processing to image analysis. The quality of segmentation has a direct impact on the subsequent technical tasks. Focusing on the semantic awareness of the user’s demands for the segmentation, this project proposes a semantic-aware image segmentation model based on interactive continuous learning and further researches on the corresponding theoretical method to obtain high-quality segmentation results meeting the demands of users. (1) To overcome the sensitivity to interactive information and further reduce the user’s interactive burden, we propose a user intent semantic-aware model, which can accurately obtain the user’s intent under limited interaction. (2) To overcome the limitations of the local relationship for the common problems such as low contrast, complex texture, uneven brightness and so on, we extend the segmentation strategy to the entire image and construct the image segmentation model based on local to the whole laws of perceptual grouping, which can achieve high-precision segmentation results. (3) To break the limitations of different image scenarios, types and wrong segmentation, the interactive continuous learning strategy is introduced and a parallel game theory-based optimization model is constructed, which can achieve universal application of the algorithm to meet the actual segmentation demands of users. The researches in this project will vastly promote the theoretical development for the interactive segmentation model, and can help solve real-world related image processing problems in the form of software, which are very meaningful and valuable in both theory and practice.

图像分割是从图像处理到图像分析的一个关键步骤，分割质量的好坏直接影响后续的图像技术任务。本项目旨在语义感知用户的分割需求，深入开展交互式持续学习的图像目标语义感知分割理论与方法研究，实现满足用户需求的高质量图像分割：（1）构建用户意图语义感知模型，克服算法对于交互信息的敏感性，减轻用户交互负担，实现有限交互下用户意图的准确获取；（2）将局部分割策略延伸到图像整体，构建基于局部到整体知觉组织原则的图像分割模型，克服局部关系对于低对比度、复杂纹理结构、亮度不均等分割领域共性问题的局限，实现高精度的图像目标分割；（3）引入交互式持续学习策略，构建并行博弈论优化模型，突破算法对于不同图像情景、类型以及错误分割的限制，实现普适且满足用户实际分割需求的算法应用。本项目对于推动交互式分割模型的理论和应用研究具有重要意义，能够以软件的形式辅助解决现实中相关的图像处理问题。

图像分割是从图像处理到图像分析的一个关键步骤，分割质量的好坏直接影响后续的图像技术任务。本项目旨在语义感知用户的分割需求，深入开展交互式持续学习的图像目标语义感知分割理论与方法研究，实现满足用户需求的高质量图像分割：（1）构建用户意图语义感知模型，克服算法对于交互信息的敏感性，减轻用户交互负担，实现有限交互下用户意图的准确获取；（2）将局部分割策略延伸到图像整体，构建基于局部到整体知觉组织原则的图像分割模型，克服局部关系对于低对比度、复杂纹理结构、亮度不均等分割领域共性问题的局限，实现高精度的图像目标分割；（3）引入交互式持续学习策略，构建持续分割修正模型，突破算法对于不同图像情景、类型以及错误分割的限制，实现普适且满足用户实际分割需求的算法应用。通过本项目研究所提出的理论方法来解决目前已有方法存在的问题，在目标先验感知、标签传递、持续学习修正等方面取得了一定突破，研究结果对于推动图论统计模型以及深度分割网络模型的应用研究具有一定的理论意义，为相关智能化图像分析工作提供可靠保障，同时也为真实环境中开放式大规模数据集的快速标注提供支撑。