【作者】 赵军；

【导师】 张秀玲； 徐春阳；

【作者基本信息】 燕山大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 语义分割是计算机视觉任务中的一个重要分支,可以把它看作是底层像素的分类问题,归因于深度学习理论的发展以及计算机并行运算能力的提高,当下的语义分割算法的精度得到了极大的提升。但是当下的语义分割算法采用全监督的训练模式,也就是采用像素级别的图像标签进行训练网络,然而这种标签的标注需要花费大量的人力和财力,阻碍了语义分割的应用场景。为了降低对像素标签的依赖性,一些学者提出了弱监督语义分割算法,这种算法摒弃了传统的像素级的图像标签,采用更加容易获得和标注的图像标签,比如图像级标签,边界框级标签,涂鸦级标签等,其中图像级标签是最容易标注的,因此许多弱监督语义分割采用这种标签。由于图像级标签仅仅表示图像中目标物体是否存在,丢失了对于语义分割至关重要的语义信息,类激活映射(Class Activation Maps,CAM)的出现为弱监督语义分割提供了必不可少的目标位置先验信息。但是类激活映射仅仅确定小的显著性区域,这对于语义分割是不足够的。针对这个问题,本文进行了如下工作:(1)设计了多特征融合模块(Multi-Feature Fusion Module,MFFM),充分利用了各个阶段获得的特征映射,弥补了丢失的浅层的语义信息,与此同时结合利用级联和并联的空洞卷积操作来扩大特征映射的感受野。为了获得周围区域与它相邻以及相距较远区域的关系,进一步引入了全局推理单元来计算周围像素之间的关系。(2)设计了注意力混合池化模块(Attention Hybrid Pooling Module,AHPM)来代替传统的全局平均池化(Global Average Pooling,GAP)操作,解决了背景像素的误分类问题,利用了广泛使用的空间注意力机制和通道注意力机制为特征映射重新安排权重。(3)结合MFFM和AHPM设计出了全局推理网络(Global Reasoning Network,GRN),在PASCAL VOC 2012数据集训练该网络获得了包括更大目标范围的类激活映射,而不仅仅局限在小的显著性目标区域。为了获得更加良好的分割性能,本文设计了新的孪生网络(Siamese Global Reasoning Network,SGRN),实验结果验证了孪生网络的优越性。更多还原

【Abstract】 Semantic segmentation is an important branch of computer vision tasks,which can be considered as the underlying pixel classification problem.Thanks to the development of deep learning theory and the improvement of computers’ parallel computing power,the accuracy of current semantic segmentation algorithms has been greatly improved.However,contemporary semantic segmentation algorithms use a fully supervised training model,which means that pixel-level image labels are used to train the network,and it is obvious that such labeling requires a lot of human and financial resources,which hinders the application scenarios of semantic segmentation.In order to reduce the dependence on pixel labels,some scholars have proposed weakly supervised semantic segmentation algorithms,which eschew traditional pixel-level image labels in favor of more easily obtained and labeled image labels,such as image-level labels,bounding box-level labels,and scribblelevel labels,among which image-level labels are the easiest to label,and thus many weakly supervised semantic segmentation algorithms use such labels.Since image-level labels only indicate the presence or absence of the target object in the image,losing semantic information that is crucial for semantic segmentation,the emergence of class activation maps(CAM)provides crucial a priori information about the target location for weakly supervised semantic segmentation.But the class activation maps only determine small discriminative regions,which is not sufficient for semantic segmentation.In response to this problem,the following work is carried out in this paper.(1)A multi-feature fusion module(MFFM)is designed to make full use of the feature maps obtained at each stage to compensate for the loss of semantic information at a shallow level,and combine the use of cascaded and parallel dilated convolution operations to expand the perceptual field of the feature maps.To obtain the relationship between surrounding regions and adjacent and distant regions,a global reasoning unit is further introduced to calculate the relationship between surrounding pixels.(2)The attention hybrid pooling module(AHPM)is designed to replace the traditional global average pooling(GAP)operation to solve the misclassification problem of background pixels,and the widely used spatial attention mechanism and channel attention mechanism are utilized to rearrange weights for feature maps.(3)A global reasoning network is designed by combining MFFM and AHPM,and training the network on the PASCAL VOC 2012 datasets obtains class activation maps that include a larger range of targets,rather than being limited to only small salient target regions.In order to obtain better segmentation performance,the paper designs siamese network,and the experimental results verify the superiority of the siamese network.更多还原