【作者】 王健；

【导师】 王世刚；

【作者基本信息】 吉林大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 深度学习的兴起引发了人工智能领域的研究热潮,深度卷积神经网络是计算机视觉领域中非常可靠的方法。随着图像数据显著增加和智能设备日益普及,快速而准确地理解图像的内容并自动识别和分割图像中的目标对象是很有必要的。目前,国内外已有许多关于图像分割算法的优秀研究成果。当将其应用于实际操作中时,发现仍然存在许多问题。例如,在诸如部分物体重叠和遮挡、背景杂乱、光强度等干扰下,目标物体难以获得满意的分割精度。本文对现有语义分割算法进行了深入的研究和分析,基于现有的算法上分析改进,并设计了很多自适应管道和模块,本文的具体工作如下:首先,为了更好的提取不同尺度下目标图像特征,本文改变了原有残差结构设计了并行卷积模块,提出了一种基于改进残差网络的特征提取网络,通过扩展网络的感受野来获取更多语义信息,为下文嵌入平衡的混合级联语义分割算法减少运算参数做了充分的准备。其次,本文提出了一种嵌入平衡改进残差网络的混合级联语义分割算法,解决了多个目标重叠时分割边缘线过度模糊的问题。在获取特征时,采用可变形卷积加改进式并联残差的网络结构以获取更多空间信息;在进行特征融合时,发现成功的语义分割融合的关键是能否充分利用边界框检测结果和掩码(Mask)预测分割之间的关联信息,因此为了进一步处理边界框信息,改善信息流,本文使用级联管道将Mask RCNN和Cascade RCNN网络结构交并起来,在每个阶段,通过掩模预测和边界框回归并行融合来获取它们之间的相互信息关系,从而进一步改善RCNN中的信息流。此外,网络模型中还加入了Io U平衡采样、平衡特征金字塔和平衡L1损失函数,分别用于降低样本采集、提取特征和目标检测三方面的不平衡。最后,本文提出了一种可得到高分辨率图像语义分割结果的优化网络,本质上介绍了本文研究的第二种图像语义分割算法,解决了传统全卷积模型中多次变形和采样导致图片分辨率过低、对象无法识别等问题。引入分离-转换-合并思想来优化网络,用增添交叉单元的HRNet网络代替嵌入平衡改进残差网络的混合级联结构中的特征提取网络。此外,在提取特征模块提出了一种用于处理多尺度图像对象的混合扩展单元,该模块包括一个可以执行多尺度融合的替代通信策略,增强了图片模糊边界信息间信息共享,大大提高了模型分割的准确性。更多还原

【Abstract】 The rise of deep learning has caused a research boom in the field of artificial intelligence,and deep convolutional neural networks are a very reliable method in the field of computer vision.With the significant increase in image data and the increasing popularity of smart devices,it is necessary to quickly and accurately understand the content of the image and automatically identify and segment the target object in the image.At present,there are many excellent research results on image segmentation algorithms at home and abroad.When it is applied to the actual operation process,it is found that there are still many problems.For example,it is difficult to obtain satisfactory segmentation accuracy for the target object under interference such as partial object overlap and occlusion,background clutter,and light intensity.This paper conducts in-depth research and analysis on existing semantic segmentation algorithms,analyzes and improves on existing algorithms,and designs many adaptive pipelines and modules.The specific work of this paper is as follows:First of all,in order to better extract target image features at different scales,this paper changes the original residual structure and designs a parallel convolution module,and proposes a feature extraction network based on an improved residual network to expand the receptive field of the network.Ways to obtain more semantic information,and make sufficient preparations for the following embedded balanced hybrid cascaded semantic segmentation algorithm to reduce operational parameters.Secondly,this paper proposes a hybrid cascaded semantic segmentation algorithm with embedded balance improved residual network,which solves the problem of excessive blurring of segmentation edge lines when multiple targets overlap.When acquiring features,a network structure of deformable convolution and improved parallel residuals is used to obtain more spatial information;in feature fusion,it is found that the key to successful semantic segmentation fusion is to make full use of bounding box detection results and masks(Mask)predicts the associated information between the segmentation.Therefore,in order to further process the bounding box information and improve the information flow,this paper uses a cascaded pipeline to merge the Mask RCNN and Cascade RCNN network structures,and at each stage,through the mask prediction and boundary Box regression parallel fusion to obtain the mutual information relationship between them,so as to further improve the information flow in RCNN.In addition,Io U balanced sampling,balanced feature pyramid,and balanced L1 loss function are also added to the network model,which are used to reduce the imbalances in sample collection,feature extraction,and target detection,respectively.Finally,this paper proposes an optimized network that can obtain high-resolution image semantic segmentation results.In essence,it introduces the second image semantic segmentation algorithm studied in this paper,which solves the multiple deformation and sampling processes in the traditional full convolution model.This leads to problems such as low image resolution and unrecognizable objects.The idea of separation-transformation-merging is introduced to optimize the network,and the feature extraction network in the hybrid cascade structure of the residual network is improved by replacing the embedded balance with the HRNet network with the added cross unit.In addition,in the feature extraction module,a hybrid extension unit for processing multi-scale image objects is proposed.This module includes an alternative communication strategy that can perform multi-scale fusion,enhances the information sharing between the fuzzy boundary information of the picture,and greatly improves the model.Accuracy of segmentation.更多还原