【作者】 赵婷；

【导师】 邬向前；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 随着科技的发展,网络化、智能化开始渗透到各个行业并逐渐影响着我们的日常生活。很多行业已经通过技术的进步解放了人们的双手,机器代替了传统繁杂而重复的工作,很多要花费大量人力,甚至使用人力无法进行的工作在机器的帮助下可以很好的完成。其中,计算机视觉技术作为计算机重要的感知手段,在很多行业发挥着独特且重要的作用。计算机视觉模拟了人类视觉感知的方式,通过对图像、视频等多媒体数据的处理,进而获取到人们需要的场景信息,形象的说,计算机视觉就像是机器的视觉系统,使得机器能够感知环境,为人类观察或传送给下一阶段设备进行决策提取帮助。计算机视觉模拟人类视觉系统,使机器可以像人类一样观察和理解图片,一个关键任务就是显著性检测。人类能够很快的在丰富的视觉场景中找到少量的重要信息,对应到计算机视觉上,就是显著性检测这个任务。目前大多数显著性检测的研究成果是基于全卷积深度神经网络的。全卷积神经网络通过叠加大量的卷积层和池化层获取丰富的语义特征,但是其中存在着一个隐患——空间结构缺失。池化层能够扩大特征提取的感受野、获取语义上下文信息,但它也造成了图片空间结构信息的损失。在一些逐像素的检测任务中,要求输出图与原图大小一致,因此需要保留图像的空间结构信息。本文针对目前显著性检测的问题出发,从增强显著性检测的空间结构信息角度设计了不同的方法。一、基于底层特征增强的图像显著性检测方法,该方法将边缘图和颜色块等信息加入到网络的学习中,通过多任务学习的方法加强底层特征的空间结构信息,然后通过短连接的方式将具有强空间信息的底层特征融入具有高语义特征的高层特征中,得到能够保持边缘信息的显著性检测结果。二、基于金字塔特征注意力模型的图像显著性检测方法,该方法针对网络同层次特征的特性,使用通道注意力和空间注意力机制来选取合适的语义特征图,同时过滤掉低层次特征的背景噪声,使网络学习到的特征能够关注到正确的显著性区域特别是显著性区域与背景的边界上,从而生成边界清晰,前景背景平衡的显著性检测结果。三、基于分类模型和导向滤波优化的弱监督图像显著性检测方法,该方法利用已训练好的分类模型基本网络,经过阈值限制,得到显著性检测的伪标签,然后利用伪标签进行显著性检测网络模型的训练。由于利用伪标签得到的显著性图是不准确的,最后,我们使用导向滤波对网络生成的结果进行后处理,得到最终的显著性图。更多还原

【Abstract】 With the development of modern science and technology,internet,digitalization and intelligence are applied to various industries and gradually permeate into people to live aspect,influencing the life style and values.Through the progress of technology,everything seems to be getting faster,more efficient,more compact and hands-free in many industries.Machines have replaced the traditional complex and repetitive work which requires a lot of manpower can be well completed with the help of machines in saome industries.As an important machine perception,computer vision plays a vital role in many tasks.Computer vision simulates the human visual perception,the scene sematic information can be obtained by processing images,video and multimedia data.Computer vision is like a machine’s visual system,which enables the machine to perceive the environment,and providing convenience for human observation and decision-making.One of the key tasks of how computer vision simulates human visual system and enables machines to observe and understand pictures is saliency detection.Human can find a small amount of important information in rich visual scenes quickly,image saliency detection plays the same role in computer vision.At present,most of the methods of saliency detection are based on full convolution depth neural network.Full convolution networks extract contest semantic features by stacking numbers of convolution layers and pooling layers.Pooling layer can enlarge the receptive field and aggregate the semantic context,and it also causes lack of spatial structure information.In some pixel-by-pixel detection tasks,such as image saliency detection,the output image required to be the same size as the original image,the spatial structure information of the image is important to be preserved.For solving the problem in saliency detection,this paper proposes some methods to enhance spatial structure information of saliency detection.The first one is image saliency detection with low-level features enhancement,the method introduces edge map and color block information into network learning.The spatial structure information of bottom features is enhanced by multi-task learning method.Then,the bottom features with strong spatial information are integrated into the high-level features by short connection,and saliency detection that can maintain edge information.The second method,pyramid feature attention network for saliency detection considers about different characteristics of different level features.The channel attention and spatial attention mechanism are used to select the appropriate semantic feature map and filter the background noise of the low level feature,so that the features learnt by the network can focus on the correct saliency region and the boundary between saliency region and background.The last one is weakly supervised image saliency detection via classification model and guided filter optimization.A classificationed model of classification is used to obtain the pseudo-label of saliency detection by threshold restriction.Then the pseudo-label is used to train a saliency detection model.Because the saliency map obtained by using pseudo-labels is inaccurate,we use guided filter to post-process the saliency maps and get the final result.更多还原