多类别标签弱监督语义分割热力图生成算法  

Heat Map Generation Algorithm for Weakly-supervised Semantic Segmentation of Multi-category Labels

作　　者：迟津生 杨大伟 毛琳 CHI Jin-sheng;YANG Da-wei;MAO Lin(School of Electromechanical Engineering,Dalian Minzu University,Dalian Liaoning 116605,China)

机构地区：[1]大连民族大学机电工程学院,辽宁大连116605

出　　处：《大连民族大学学报》2023年第1期40-46,共7页Journal of Dalian Minzu University

基　　金：国家自然科学基金项目(61673084);辽宁省自然科学基金项目(20170540192,20180550866,2020-MZLH-24)。

摘　　要：针对弱监督语义分割过程中,类激活映射只关注影响分类结果的像素点,而忽略了图像内弱监督学习过程中出现的不确定像素点,导致类激活映射无法生成高精度热力图的问题,提出一种多类别标签弱监督语义分割热力图生成算法。通过将图像特征与种子区域生成的多类别标签相结合,构造一种注意力机制,使得算法可以关注到每类中的像素点,解决了弱监督语义分割过程中出现的不确定像素点分类问题,并生成高精度热力图,进一步提高弱监督语义分割算法的精度。在PascalVOC2012数据集进行仿真测试,与该算法同类激活映射相比,平均交并比提升14.57%。在无人车自主驾驶等领域具有较好的应用前景。In the process of weakly-supervised semantic segmentation,the class activation map(CAM)only focuses on the pixels that affect the classification results,ignoring uncertain pixels that appear in the weakly-supervised learning process in the image,resulting in that CAM cannot generate high-precision heat maps.To solve this problem,this paper proposes a multi-category label weakly-supervised semantic segmentation heat map generation algorithm.By combining image features with multi-category labels generated in seed regions,the algorithm constructs an attention mechanism to pay attention to which class every pixel is in,solving the problem of uncertain pixel classification in the process of weakly-supervised semantic segmentation.And the algorithm generates high-precision heat maps to further improve the accuracy of weakly-supervised semantic segmentation.Simulation tests were carried out on the PascalVOC2012 dataset,and the simulation results of the algorithm increased by 14.57%compared with CAM.It has a good application prospect in the field of autonomous driving of unmanned vehicles.

关 键 词：弱监督语义分割 热力图 种子区域 边界探索 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]