LLC编码BoW与Softmax回归：微观图像分类新策略

135 浏览量更新于2024-08-26 收藏 195KB PDF 举报

本文探讨了一种结合局部线性编码（LLC）与Bag-of-Words (BoW) 特征以及Softmax回归的微观图像分类方法。在传统的计算机视觉领域中，BoW模型常用于将图像分解为可描述的视觉单词集合，这些单词代表了图像中的局部特征。然而，传统的方法如矢量量化（VQ）在处理微观图像时可能存在较大的量化误差，无法充分捕捉到图像的细微结构。 LLC是一种改进的局部特征编码技术，它通过限制特征之间的空间邻域关系，提高了特征编码的精度。相比于VQ，LLC能够在保持局部结构的同时减少量化误差，从而生成更为稀疏且信息丰富的图像表示。这种稀疏性使得基于线性分类器的计算复杂度得以降低，对于大规模数据集和实时性能要求较高的应用场景具有显著优势。在该研究中，作者提出了一种新型的微观看图分类流程：首先，利用LLC对图像的局部特征进行编码，形成一个包含丰富局部结构信息的BoW特征向量。然后，引入Softmax回归作为分类器，这是一个多类概率估计算法，能够处理多个类别间的非线性关系，并输出每个类别的概率分布，从而实现图像的精确分类。 Softmax回归通过最大化所有类别的预测概率之和，同时确保每个类别的概率和为1，确保了分类决策的合理性。这种方法能够有效地解决传统BoW框架下可能存在的过拟合问题，提高模型的泛化能力。总结来说，本文的主要贡献在于提出了一种利用LLC编码的BoW特征和Softmax回归相结合的微观图像分类策略，它既保持了BoW的全局上下文信息，又通过LLC的局部精确性降低了计算复杂度，提升了分类性能。这对于实际的微观图像分析，例如生物医学图像识别、细胞分析等领域有着重要的应用价值。

LLC Encoded BoW Features And Softmax

Regression for Microscopic Image Classiﬁcation

Dongyun Lin

∗

, Zhiping Lin

∗

, Lei Sun

†

, Kar-Ann Toh

‡

and Jiuwen Cao

∗

School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore

†

School of Information and Electronics, Beijing Institute of Technology, Beijing, 100081, PR China

‡

School of Electrical and Electronic Engineering, Yonsei University, Seoul 120-749, Korea

Key Lab for IOT and Information Fusion Technology of Zhejiang, Hangzhou Dianzi University, 310018, China

Abstract—This paper proposes a method based on the bag-of-

words (BoW) and the softmax regression for microscopic image

classiﬁcation. Essentially, the locality-constrained linear coding

(LLC) is adopted for local feature encoding. Compared with the

traditionally adopted vector quantization (VQ) in the BoW frame-

work, the LLC encodes local structures of microscopic images

with lower quantization errors and generates a sparse image

representation. This enables the use of linear classiﬁers with

low computational complexity. A softmax regression classiﬁer is

then adopted to address the multi-categorical classiﬁcation task

where the conﬁdence of categorical prediction is quantiﬁed by

posterior probabilities. Compared with other linear classiﬁers

(such as the linear SVM) which only assign labels to images,

such probabilistic outputs provide extra quantitative information

to analyze misclassiﬁed images. Our experiments on the 2D-

Hela and the PAP smear data sets show signiﬁcant performance

improvement of the proposed method comparing with competing

methods using different features and classiﬁers under the BoW

framework.

Keywords—bag-of-words (BoW); locality-constrained linear

coding (LLC); softmax regression classiﬁer; microscopic image

I. INTRODUCTION

Recent advances in cell biology and microscope automation

produce a large-scale of biomedical images. The microscopic

image is one of the most important type of biomedical images

which plays a crucial role in the study of subcellular biological

structures. For instance, the subcellular localization of proteins

is obtained through microscopic images of cells stained with

monoclonal antibodies with respect to speciﬁc endogenous

proteins [1]. In addition, multiple diseases [2][3] are studied

through biomedical screening and imaging. One important

medical application is that human cells can be stained using

the Papanicolaou method and the microscopic images of these

cells are examined for possibility of cancers [4]. Due to

the high complexity and diversity of such images, the task

of microscopic image classiﬁcation is traditionally addressed

by visual inspection based on human knowledge. However,

such human inspection is not only time consuming, but also

subjective and inaccurate. Hence, there is a growing number of

research works on developing automated methods to classify

microscopic images.

A general pipeline for microscopic image classiﬁcation

consists of two stages, i.e., feature extraction and classiﬁer

training. In the feature extraction stage, features such as

Zernike moments [5], Gabor [6], curvelet [7], local binary

patterns (LBP) [8] are extracted to form feature representations

of training images. In the classiﬁer training stage, classiﬁers

are trained based on these features. Popular classiﬁers adopted

in microscopic image classiﬁcation include support vector

machines [5][9] and artiﬁcial neural networks [7]. The bag-

of-words (BoW) [10] is a popular model for forming a

discriminative representation of local features from an image.

Motivated by the successful application of BoW in na-

ture object/scene recognition tasks [11][12][13], we propose

a classiﬁcation method for microscopic image classiﬁcation

under the BoW framework. In the feature extraction stage,

the dense SIFT features are extracted and encoded using the

locality-constrained linear coding (LLC) [12] to form effective

sparse representation of training images. In the classiﬁer train-

ing stage, a softmax regression classiﬁer is trained on these

features for multi-categorical classiﬁcation.

The main contributions of this paper are three-folds: (i) The

encoding of local structures of microscopic images by LLC

produces lower quantization errors than the traditional vector

quantization (VQ) technique. The ﬁnal image representation

through LLC can be classiﬁed using linear classiﬁers with

low computational complexity. (ii) The design of a softmax

regression classiﬁer shows better generalization performance

compared with the competing methods for multi-categorical

classiﬁcation problems. (iii) The softmax regression can gen-

erate probabilistic outputs to quantify classiﬁcation conﬁdence

which are useful for analysis of misclassiﬁed images. Our

experiments on two benchmark data sets of microscopic im-

ages, namely the 2D-Hela [1] and the PAP smear [4] data

sets, show signiﬁcantly improved classiﬁcation accuracy of

the proposed method compared with the recently published

microscopic image classiﬁcation methods under the BoW

framework. Speciﬁcally, the proposed method achieves 89.37%

and 89.96% classiﬁcation accuracies for the 2D-Hela and the

PAP smear data sets, respectively. The performance gains for

2D-Hela and PAP are 3.16% and 2.33% compared with the

competing methods.

The remaining parts of the paper are organized as follows.

Section II presents the details of the proposed method. In

Section III, the proposed method is compared with the state-of-

art competing methods under the BoW framework on 2D-Hela

[1] and PAP smear [4] data sets. Finally, Section IV concludes

the entire paper.

下载后可阅读完整内容，剩余3页未读，立即下载

weixin_38657848

粉丝: 5
资源: 906

LLC编码BoW与Softmax回归：微观图像分类新策略

OpenCV基于SVM与BoW类型的图片分类的实现

kmeans提取特征训练softmax分类器

图像分类中BOW模型

BOW图像分类程序

BoW|Pyramid BoW+SVM进行图像分类

考虑用于图像分类的特征包（BoF）框架的空间布局信息

Bag-of-Words:使用 BoW 进行图像分类

VS2013+opencv2.4.10提取SURF_BOW特征使用svm多分类识别手势

基于图正则化局部特征编码算法的图像分类方法.docx

计算图像的视觉词表示包：计算图像的 BOV/BOF/BOW 表示。 码本和图像特征是用户输入。-matlab开发

最新资源

计算图像的视觉词表示包：计算图像的 BOV/BOF/BOW 表示。码本和图像特征是用户输入。-matlab开发