【作者】 李翠萍；

【导师】 陈振学；

【作者基本信息】 山东大学 ， 控制科学与工程， 2019， 硕士

【摘要】 随着移动电子设备的不断升级与应用,使用图像来记录或表达信息已成为一种常态。我们要想快速地在海量图像中提取出有价值的信息,那么需要模拟人类视觉系统在机器视觉系统进行计算机视觉热点问题的研究。图像显著性目标检测对图像中最引人注意且最能表征图像内容的部分进行检测。在图像显著性目标检测任务中,传统的方法一般利用纹理、颜色等低层级视觉信息自下向上地进行数据驱动式检测。对于含有单一目标或高对比度的自然场景图像,可以从多个角度去挖掘其显著性信息,如先验知识、误差重构等。然而,对于那些具有挑战性的自然场景图像,如复杂的背景、低对比度等,传统的方法通常会检测失败。基于深度卷积神经网络的算法利用高层级语义信息结合上下文充分挖掘潜在的细节,相较于传统的方法已取得了更优越的显著性检测性能。本文对于图像显著性检测任务存在的主要问题提出了相应的解决方法。本文的主要贡献如下:为充分挖掘图像多种显著性信息,并使其能够达到优势互补效果,本文提出了一种有效的模型,即融合先验信息和重构信息的显著性目标检测模型。重构过程包括密度重构策略与稀疏重构策略。密度重构其优势在于能够更准确地定位存在于图像边缘的显著性物体。而稀疏重构更具鲁棒性,能够更有效地抑制复杂背景。先验过程包含背景先验策略与中心先验策略,通过先验信息可更均匀地突出图像中的显著性目标。最后,把重构过程与先验过程生成的显著特征做非线性融合操作。实验结果充分说明了该模型的高效性能与优越性能。针对图像中存在多个显著性目标或者检测到的显著性目标存在边界模糊问题,本文提出了一种基于多层级连续特征细化的深度显著性目标检测模型。该模型包括三个阶段:多层级连续特征提取、分层边界细化和显著性特征融合。首先,在多个层级上连续提取和编码高级语义特征,该过程充分挖掘了全局空间信息和不同层级的细节信息。然后,通过反卷积操作对多层级特征做边界细化处理。分层边界细化后,把不同层级的显著特征做融合操作得到结果显著图。在具有挑战性的多个基准数据集上使用综合评价指标进行性能测试,实验结果表明该方法具有优越的显著性检测性能。对于低对比度或者小目标等问题,本文提出一种新颖模型,即通道层级特征响应模型。该模型包含三个部分:通道式粗特征提取,层级通道特征细化和层级特征图融合。该方法基于挤压激励残差网络,依据卷积特征通道之间的相关性进行建模。首先,输入图像通过通道式粗特征提取过程生成空间信息丢失较多的粗糙特征图。然后,从高层级到低层级逐步细化通道特征,充分挖掘潜在的通道相关性细节信息。接着,对多层级特征做融合操作得到结果显著图。在含有复杂场景的多个基准数据集上与其它先进算法进行比较,实验结果证明该算法具有较高的计算效率和卓越的显著性检测性能。更多还原

【Abstract】 With the widespread utility and escalation of mobile electronic devices,the use of images to record or express information has become a norm.In order to quickly extract valuable information from massive images,we need to simulate the human visual system to study computer vision hotspots in machine vision systems.Image saliency target detection is to detect the most attractive and most representative part of the image.In the image saliency target detection task,the conventional method generally performs bottom-up and data-driven detection using low-level visual information such as texture,color,etc.For natural scene images with a single target or high contrast,the saliency information can be explored from multiple angles,such as prior knowledge,error reconstruction,etc.However,for the challenging natural scene images,such as complex backgrounds,low contrast,etc.,traditional methods often fail to detect.Deep convolutional neural-network based algorithms take advantage of high-level semantic information combined with context to fully exploit potential details,and have achieved superior saliency detection performance compared to traditional methods.This paper proposes the corresponding methods for the main problems existed in image saliency target detection task.The main contributions of this paper are as follows:In order to fully exploit the various saliency information of the image and make it achieve the complementary effect,this paper proposes a saliency target detection model that combines the reconstruction and priori.Reconstruction includes dense reconstruction and sparse reconstruction.The advantage of dense reconstruction is to make more accurate detection when the salient targets appear at the edges of the image.Sparse reconstruction is more robust and can suppress the complex backgrounds more effectively.The priori includes the background priori and central priori,and the priori knowledge can highlight the salient targets more uniformly.Finally,the saliency maps obtained by the reconstruction and the priori are nonlinearly fused.The experimental results prove the efficiency and superiority of the proposed work in the saliency target detection task.For the problem of multiple saliency targets existed in the image or that of the boundary ambiguity of the detected saliency targets,this paper proposes a deep saliency target detection model based on multi-level continuous features and hierarchical refinement.The model consists of three structures:multi-level continuous feature extraction,hierarchical boundary refinement,and saliency feature fusion.At first,high-level semantic features are continuously extracted and encoded at multiple levels.The process fully exploits global spatial information and details of different levels.Then,the deconvolution operation is used to perform boundary refinement on the multi-level features.Finally,the saliency feature maps obtained are merged to generate the final saliency map.The performance evaluation is carried out using comprehensive metrics on several challenging datasets.The experimental results demonstrate that the method has superior performance.For the problems of low contrast or small targets,this paper proposes a deep saliency target detection model based on channel-wise feature response.The model consists of three parts:channel-wise coarse feature extraction,hierarchical channel-wise feature refinement,and hierarchical feature map fusion.The method is based on the squeeze and excitation residual network,which is modeled based on the correlation between the convolution feature channels.Firstly,channel-wise coarse feature extraction is performed on the input image to generate the global coarse feature maps with more information loss.Then,the channel features are gradually refined from the high level to the low level,and the potential channel correlation details are fully exploited.Finally,the multi-level feature maps are fused to generate the final saliency map.Compared with other state-of-the-art algorithms on multiple saliency datasets with complex scenes,the experimental results demonstrate that the proposed algorithm has high computational efficiency and excellent saliency detection performance.更多还原