摘要
针对抠图任务中人像毛发这类细微结构局部区域出现的误抠问题,其本质是此区域中混合信息所导致的图像像素点透明度遮罩回归预测不准确问题,提出一种端到端的层次特征聚合抠图网络模型。该模型通过一个共享编码器和两个独立解码器,利用通道和位置注意力机制,以层次特征聚合方式聚合低层次纹理线索和高层次语义信息,能够在没有额外输入的情况下从单个人像的精细边界和自适应语义中感知前景透明度遮罩。基于此,结合交叉熵损失、未知区域的透明度遮罩预测损失和结构性损失,以引导层次特征聚合抠图网络模型完善前景整体结构,恢复毛发纹理细节。为验证所设计模型的有效性,在自建的MCP-1k和公开的P3M-500-NP数据集上进行验证分析,实验结果表明所提模型在MSE和SAD指标上的误差分别为0.0076,25.59与0.0072,25.52,与其他典型深度抠图模型相比在恢复人像细微毛发和完善人像语义结构方面有较大提升,解决人像毛发区域的误抠问题。
Addressing the issue of erroneous extraction of fine structures such as human hair in image matting tasks,the problem essentially stemmed from inaccurate prediction of pixel alpha mattes due to mixed information within these regions.To address this problem,a novel end-to-end hierarchical feature aggregation matting network model is proposed.This model incorporates a shared encoder and two independent decoders,leveraging channel and positional attention mechanisms to aggregate low-level texture clues and high-level semantic information in a hierarchical manner.It enables perceiving foreground transparency masks from fine boundaries of individual portraits and adaptive semantics without additional inputs.To guide the hierarchical feature aggregation matting network model in refining the overall foreground structure and restoring hair texture details,cross-entropy loss,alpha matte prediction loss for unknown regions,and structural losses are integrated.To validate the effectiveness of the proposed model,experiments were conducted on the self-constructed MCP-1k dataset and the publicly available P3M-500-NP dataset.Experimental results demonstrated that the proposed model achieved errors of 0.0076 MSE and 25.59 SAD on MCP-1k dataset,and 0.0072 MSE and 25.52 SAD on P3M-500-NP dataset,respectively.Compared with other typical deep matting models,it showed significant improvements in restoring fine human hair and enhancing semantic structure in portraits,effectively addressing the issue of erroneous extraction in human hair regions.
作者
汪小梅
谭棉
罗太维
冯夫健
Wang Xiaomei;Tan Mian;Luo Taiwei;Feng Fujian(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Key Laboratory of Pattern Recognition and Intelligent System,Guizhou Minzu University,Guiyang 550025,China)
出处
《电子测量技术》
北大核心
2024年第14期170-177,共8页
Electronic Measurement Technology
基金
贵州省科技计划项目(黔科合基础-ZK[2022]一般195,黔科合基础-ZK[2023]一般143,黔科合平台人才-ZCKJ[2021]007)
贵州省教育厅自然科学研究项目(黔教技[2023]012号,黔教技[2022]015号,黔教技[2023]061号)
贵州省模式识别与智能系统重点实验室开放课题(GZMUKL[2022]KF01,GZMUKL[2022]KF05)项目资助。
关键词
图像抠图
透明度遮罩
特征融合
层次特征聚合
自动人像抠图
image matting
alpha matte
feature fusion
hierarchical feature aggregation
automatic portrait matting