CASIA OpenIR  > 毕业生  > 硕士学位论文
面向显微图像分割的深度学习方法研究
高楷涵
2024-05-15
Pages88
Subtype硕士
Abstract

随着高通量电子显微镜成像技术的发展,使用最先进的扫描电子显微镜可以在数日内获取大量显微图像,其规模可达数万张。利用显微图像分割技术,来自不同领域的研究人员可以对微观结构开展更高分辨率和更大的体积的研究。近年来,深度学习的进步催生了新的图像处理算法,与传统图像分割方法相比,基于深度神经网络的方法可以更加有效地针对显微图像数据进行识别与分割。尽管深度学习方法已经在显微图像分割领域取得了广泛的应用,但如何充分利用有限的标注数据,实现对显微图像在不同场景下的高效、准确分割,仍是一个亟待解决的难题。

本文以深度表示方法为切入点,聚焦显微图像分割精度低和有标注数据匮乏的问题,改进了显微图像语义分割深度模型,并引入半监督分割方法以利用大量无标注数据,旨在达到更高的图像分割准确度,为显微图像分割领域提供泛化性更强的模型。本文的主要贡献和创新点如下:

第一,本文提出了一种面向显微图像分割的金字塔多尺度通道注意力模型(PmcaNet)。首先,针对多尺度信息缺失造成的显微图像分割不完整问题,使用多尺度通道注意力金字塔模块捕获不同维度的全局上下文依赖关系。其次,设计了轻量级自适应通道注意力模块,通过计算不同尺度的通道注意力,从而捕获长短距离交互信息。最后,高维通道注意力向量由多尺度注意力融合模块产生,该模块可以自适应地融合来自不同尺度的注意力向量。为了验证模型的效果,在通用显微图像分割数据集和自制的高温合金材料显微图像数据集上进行了实验。实验结果表明,与其他先进模型相比,PmcaNet可以在显微图像语义分割任务中取得有竞争力的性能表现,在多个数据集上均取得了性能提升。

第二,本文提出了一种基于多重扰动的半监督显微图像分割方法(EMmatch)。该方法针对显微图像语义分割领域像素级标注数据获取困难的问题,尝试采用特征空间扰动的方法,利用无标注数据提升模型泛化性。在数据扰动层面,该方法采用了三重数据扰动策略,并针对显微图像分割中多尺度信息缺失的问题加入了尺度变化操作。在特征扰动层面,利用非线性激活函数对主编码器的输出特征进行扰动,以提升模型处理非线性边界的能力。该方法在显微图像数据集上与其他半监督方法进行了公平对比。实验结果表明,EMmatch可以在使用少量有标注数据和大量无标注数据的情况下达到令人满意的分割效果,与其他先进方法相比均有一定的性能提升。

Other Abstract

With the development of high-throughput electron microscopy imaging techniques, a large number of microscopy images can be acquired in a matter of days using advanced scanning electron microscopes that can be tens of thousands of images in size. Utilizing microscopy image segmentation techniques, researchers from different fields can carry out studies of microstructures at higher resolutions and larger volumes. In recent years, advances in deep learning have given rise to new image processing algorithms, and deep neural network-based methods can be more effective in recognizing and segmenting microscopy image data than traditional image segmentation methods. Although deep learning methods have been widely used in the field of microscopy image segmentation, how to make full use of the limited labeled data to achieve efficient and accurate segmentation of microscopy images in different scenarios is still a difficult problem to be solved.

In this paper, taking the deep representation method as the entry point, focusing on the problems of low segmentation accuracy of microscopy images and the lack of labeled data, we improve the depth model of semantic segmentation of microscopy images, and introduce the semi-supervised segmentation method to take advantage of the large amount of unlabeled data, aiming to achieve a higher image segmentation accuracy and provide a more generalized model for the field of microscopy image segmentation. The main contributions and innovations of this paper are as follows:

First, this paper proposes a pyramidal multi-scale channel attention model (PmcaNet) for microscopy image segmentation. Firstly, to address the problem of incomplete microscopy image segmentation caused by the lack of multiscale information, a multiscale channel attention pyramid module is used to capture the global contextual dependencies in different dimensions. Secondly, a lightweight adaptive channel attention module is designed to capture the long and short distance interaction information by computing the channel attention at different scales. Finally, high-dimensional channel attention vectors are generated by a multiscale attention fusion module that adaptively fuses attention vectors from different scales. In order to validate the effectiveness of the model, experiments were conducted on a generalized micro-image segmentation dataset and a self-made micro-image dataset of high-temperature alloy materials. The experimental results show that compared with other state-of-the-art models, PmcaNet can achieve competitive performance in the microscopy image semantic segmentation task, and performance improvement is achieved on several datasets.

Second, this paper proposes a semi-supervised microscopy image segmentation method (EMmatch) based on multiple perturbations. Aiming at the problem of difficulty in acquiring pixel-level labeled data in the field of semantic segmentation of microscopy images, the method tries to adopt feature space perturbation to improve the model generalization using unlabeled data. At the data perturbation level, the method adopts a triple data perturbation strategy and incorporates a scale change operation to address the lack of multiscale information in microscopy image segmentation. At the feature perturbation level, a nonlinear activation function is utilized to perturb the output features of the master encoder in order to enhance the model's ability to handle nonlinear boundaries. The method is fairly compared with other semi-supervised methods on a microscopy image dataset. The experimental results show that EMmatch can achieve satisfactory segmentation results using a small amount of labeled data and a large amount of unlabeled data, all of which show some performance improvement over other state-of-the-art methods.

Keyword深度学习 显微图像 半监督学习 语义分割
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56590
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
高楷涵. 面向显微图像分割的深度学习方法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
final-高楷涵.pdf(8046KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[高楷涵]'s Articles
Baidu academic
Similar articles in Baidu academic
[高楷涵]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[高楷涵]'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.