基于深度学习的图像分割研究

发布时间：2018-04-04 17:35

本文选题：图像分割　切入点：深度学习　出处：《吉林大学》2017年硕士论文

【摘要】：图像分割是数字图像处理领域的一个主要研究方向,在许多计算机视觉应用的研究中作为第一步,起到十分关键的作用。由于图像分割的重要性和困难性,图像分割技术自20世纪70年代以来一直备受研究人员的重视。虽然人们针对各种问题提出了许多方法,但到目前为止仍然没有一个适用于大多数图像的分割算法,也没有制定出标准来选择适用的分割算法。传统的图像分割方法大部分先对图像进行特征提取,然后映射到某种模型中,不仅过程复杂,效果不鲁棒,且往往不能给出语义信息。因此,提高图像分割算法的效率并且提升分割效果具有重要的意义。得益于“大数据”和高性能计算设备的发展,深度学习成为当前人工智能领域的一大研究热点。深度学习本质上是神经网络的拓展,目前在语音识别、人脸识别、物体识别、目标检测等诸多人工智能领域都取得了突破性进展。在计算机视觉领域,深度学习应用最广泛的一种模型就是卷积神经网络。其通过局部连接和权值共享两个特性,大大减少了模型参数的数量,从而使模型可以做到很深的层次。深度学习可以从大量标注图像数据直接学习和提取相关特征,省略过去十分繁杂的建模过程,为许多领域的研究提供了便利。本文就是针对基于深度学习的图像分割方法进行深入研究。基于深度学习的图像语义分割的主要思路是,使用像素级标注的图像,利用上采样、反卷积等特殊层,将普通卷积网络提取到的特征再还原回原图像尺寸,从而实现一种端到端的学习。本文在借鉴已有网络的基础上,提出了一种叫Fast-Seg Net的网络新架构,新架构使用了最新的残差网络结构,结合了稀疏卷积、分解卷积等操作,配合批归一化,将网络深度做到了28层,在提高了精度的同时,依然保持了运行的速度。通过迁移学习的训练方法,加快了网络的收敛速度和精度。另外,本文提出了一套对图像进行像素级标注的工具和流程,并编写了图像像素级标注检查工具,采集了吉林大学校园里的街景,合格标注了1566张照片,并在其上面训练了Fast-Seg Net,取得了不错的效果,平均联合交叉度量m Io U达到了70.2%。
[Abstract]:Image segmentation is one of the main research directions in the field of digital image processing. As the first step in many computer vision applications, image segmentation plays a very important role.Because of the importance and difficulty of image segmentation, image segmentation technology has been paid much attention by researchers since 1970s.Although many methods have been proposed to solve various problems, up to now, there is still not a segmentation algorithm suitable for most images, nor has a standard been developed to select suitable segmentation algorithms.Most of the traditional image segmentation methods firstly extract the features of the image and then map to a certain model. The process is not only complex the effect is not robust but also the semantic information is often not given.Therefore, it is of great significance to improve the efficiency of image segmentation algorithm and improve the segmentation effect.With the development of big data and high performance computing equipment, deep learning has become a hot topic in the field of artificial intelligence.Depth learning is essentially an extension of neural network. At present, many artificial intelligence fields such as speech recognition, face recognition, object recognition, target detection and so on have made breakthrough progress.In the field of computer vision, one of the most widely used models of deep learning is convolution neural network.By means of local connection and weight sharing, the number of model parameters is greatly reduced, thus the model can achieve a very deep level.Depth learning can directly learn and extract relevant features from a large number of annotated image data, omitting the complicated modeling process of the past, and providing convenience for the research in many fields.In this paper, the method of image segmentation based on depth learning is studied in depth.The main idea of image semantic segmentation based on depth learning is to use the image tagged at pixel level, using special layers such as upper sampling, deconvolution and so on, to restore the features extracted from the general convolution network back to the original image size.In order to achieve an end-to-end learning.Based on the existing network, a new network architecture called Fast-Seg Net is proposed in this paper. The new architecture uses the latest residual network structure, combines the operations of sparse convolution, decomposition convolution and batch normalization.The network depth is 28 layers, which improves the accuracy and still keeps the speed of operation.The convergence speed and accuracy of the network are accelerated by the training method of transfer learning.In addition, this paper puts forward a set of tools and flow chart for image pixel level labeling, and compiles the image pixel level label inspection tool, collects the street view of Jilin University campus, and marks 1566 photos qualified.The Fast-Seg net is trained on it, and good results are obtained, and the average joint cross metric Mio U reaches 70.2%.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】