【作者】 韩震；

【导师】 朱敏；

【作者基本信息】 华东师范大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 图像语义分割旨在实现像素级的标签分类,是人工智能的热门研究方向之一。语义分割作为一项基础的视觉感知任务,其可以被广泛的应用于多种场景下,如无人驾驶、视频监控以及智能医疗等。由于卷积神经网络拥有较高的可迁移性和鲁棒性,基于卷积神经网络的语义分割方法被广泛使用。注意力机制通过模拟人类视觉机制的原理可以使模型有选择性地关注图像中的重要区域,以此提升分割效果。因此本文主要对基于注意力机制的图像语义分割方法进行研究,通过改进和优化注意力机制算法改善模型的上下文信息提取能力,从而提升语义分割的精度。本文主要工作包括以下两个方面:(1)针对当前语义分割方法中存在的上下文信息以及边界信息提取不足的问题,本文提出了基于语义边界增强和位置注意力的图像语义分割网络(SBEPNet)。为了增强目标的语义边界信息和上下文信息的提取,SBEPNet提出了语义边界增强注意力模块,以此来丰富语义边界信息并获取语义边界周围长距离依赖的上下文信息。与此同时,SBEPNet融合了位置注意力模块,通过捕获每个位置像素之间的相关性,增强重要特征的响应程度从而帮助改善语义分割效果。(2)针对空间注意力模块存在的计算量较大、占用内存高以及获取特征信息尺度单一等问题,提出了基于多尺度注意力和局部注意力的图像语义分割网络(MSALNet),该方法可以帮助模型提取多尺度全局上下文信息和局部上下文信息。MSALNet设计了一种多尺度注意力模块,该模块构建了一套新的注意力建模流程,通过使用多个尺度的特征信息进行建模,在没有增加计算量的前提下,捕获了多尺度下的全局上下文信息。此外,MSALNet还引入了局部注意力模块,帮助改善局部特征的学习能力。为了验证本文提出的图像语义分割方法的有效性,我们将本文提出的方法与当前先进的方法在公开数据集上进行了实验对比。本文提出的SBEPNet在Cityscapes和PASCAL VOC 2012测试集上分别达到了82.2%m IoU和84.6%m IoU,超越了其它所有对比方法。此外,本文提出的MASLNet在保证分割效果的同时,大幅度地降低了模型的参数量和计算量,实现了模型的分割精度和性能之间的平衡。更多还原

【Abstract】 Image semantic segmentation aims to achieve pixel-level label classification,and is one of the popular research topics of artificial intelligence.semantic segmentation is a basic visual perception task,it can be widely used in a variety of scenarios,such as autonomous cars,video surveillance,and smart medical.Due to the high transferability and robustness of convolutional neural networks,semantic segmentation methods based on convolutional neural networks have become the mainstream.The attention mechanism can make the model selectively focus on important areas in the image by simulating the principle of human visual mechanism,so as to improve the segmentation accuracy.Therefore,this paper mainly studies the image semantic segmentation method based on the attention mechanism,and improves the context information extraction ability of the model by improving and optimizing the attention mechanism algorithm,thereby improving the effect of semantic segmentation.The contributions of this paper include two aspects:(1)Aiming at the problem of insufficient extraction of context information and boundary information in current semantic segmentation methods,this paper proposes a semantic boundary enhancement and position attention network(SBEPNet).In order to enhance the extraction of semantic boundary information and contextual information of the target,SBEPNet proposes a semantic boundary enhanced attention module to enrich the semantic boundary information and obtain long-distance dependent context information around the semantic boundary.At the same time,SBEPNet incorporates a position attention module to help improve semantic segmentation by capturing the correlation between each location pixel and enhancing the responsiveness of important features.(2)Aiming at the problems of the spatial attention module,such as large amount of computation,high memory consumption,and single scale of feature information,multi scale attention and local attention network(MSALNet)is proposed.This method can help model extraction.Multi-scale global context information and local context information.MSALNet designs a multi-scale attention module,which builds a new attention modeling process,and captures multi-scale by using feature information of multiple scales for modeling without increasing the amount of computation.In addition,MSALNet also introduces a local attention module to help improve the learning ability of local features.To prove the effectiveness of the methods proposed in this paper,we compare them with current state-of-the-art methods on public datasets.The SBEPNet achieves 82.2%m IoU and 84.6% m IoU on the Cityscapes and PASCAL VOC 2012 test sets,respectively,achieving the highest segmentation accuracy compared to other methods.In addition,the MASLNet greatly reduces the amount of parameters and computation of the model while ensuring the segmentation effect,and achieves a balance between the segmentation accuracy and performance of the model.更多还原