CASIA OpenIR  > 毕业生  > 硕士学位论文
Thesis Advisor王亮
Degree Grantor中国科学院研究生院
Place of Conferral北京
Keyword深度神经网络 不规则卷积核 旋转线形卷积核 计算机视觉
Other Abstract
       提出了一种形状不规则的卷积核的学习算法。我们通过赋予卷积核以形状属性,得到了不规则卷积神经网络 ICNN。与使用规则卷积核的传统卷积神经网络不同,我们的算法在理论上可以学习任意的卷积核形状,以更好地适应不同的输入特征。也就是说,形状是和权重一样的可学习变量。借助于基本的反向传播算法,两者可以同时在端对端的训练中得到学习。我们使用对形状最敏感的图像分割任务检验本算法的有效性。
       提出了一种可旋转的线形卷积核的学习算法 RotateConv。我们设计使用最简单的几何形状,线段,来作为新的卷积核形状。但我们还赋予了卷积核旋转的能力,以用来帮助网络学习不同角度的特征。同样借助于基本的反向传播算法,卷积核的权重和角度可以在训练中同时学习。在是否使用权重共享的基础上,我们一共提出了参数量分别为2和4的两个版本的卷积核,而这相较于基本9参数的3x3卷积核来说已经压缩了非常多。
    In the background of the era of big data, deep learning has developed rapidly and made a great impact on all sectors of society. The Convolutional Neural Network (CNN) is a branch of deep learning, which is widely used in computer vision. The algorithms improved by CNN like face recognition, object detection and image segmentation have been playing a role in many fields like security, autopilot and medical treatment.
    For CNN, neural network is composed by layers, and convolutional layers' responses are determined by kernels. The approaches of deformable convolutional kernels proposed in this paper are devoted to improve the micro structure of Convolutional Neural Networks through convolutional kernels.
    Considering two aspects as expanding modeling capacity and reducing redundant parameters, we all propose two different deformations as follows:
    This paper proposes a new method to learn irregular convolutional kernels. In this paper, we equip convolutional kernels with shape attributes to generate the deep Irregular Convolutional Neural Networks (ICNN). Compared with traditional CNN applying regular convolutional kernels like 3x3, our approach trains irregular kernel shapes to better fit the geometric variations of input features. In other words, shapes are learnable parameters in addition to weights. The kernel shapes and weights are learned simultaneously during end-to-end training with the standard back-propagation algorithm. Experiments for semantic segmentation are implemented to validate the effectiveness of our proposed ICNN.
    We propose a new method, RotateConv, to learn asymmetric and rotatable convolutional kernels. The new shape is extremely simple as a line segment one, and we equip it with the rotatable ability which aims to learn diverse features with respect to different angles. The kernel weights and angles are learned simultaneously during end-to-end training via the standard back-propagation algorithm. There are two variants of RotateConv that only have 2 and 4 parameters respectively depending on whether using weight sharing, which are much compressed than the normal 3x3 kernel with 9 parameters.
    Experiments show that deformation changes the sampling method for convolutional kernels, helps extract more valuable information from input and gets better results for neural networks.
Document Type学位论文
First Author AffilicationInstitute of Automation, Chinese Academy of Sciences
Recommended Citation
GB/T 7714
马佳彬. 形变卷积核的研究与视觉应用[D]. 北京. 中国科学院研究生院,2018.
Files in This Item:
File Name/Size DocType Version Access License
马佳彬硕士毕业论文-中科院自动化研究所-(15089KB)学位论文 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[马佳彬]'s Articles
Baidu academic
Similar articles in Baidu academic
[马佳彬]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[马佳彬]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.