融合卷积神经网络与主题模型的图像标注 下载: 843次
Image Annotation Based on Convolutional Neural Network and Topic Model
江南大学物联网工程学院, 无锡 江苏 214122
图 & 表
图 1. LDA的图模型
Fig. 1. Graphical model of LDA
下载图片 查看原文
图 2. 基于迁移学习的卷积神经网络结构
Fig. 2. Structure of CNN based on transfer learning
下载图片 查看原文
图 3. 融合CNN和主题模型的图像标注框架
Fig. 3. Framework of image annotation that combines CNN and topic model
下载图片 查看原文
表 1符号及其意义
Table1. Symbols and their meaning
Symbol | Meaning of symbol | Symbol | Meaning of symbol |
---|
M | Size of training set | N | Number of vocabulary | K | Number of topics | w | Vocabulary | z | Potential topic | θ | Proportion of topic | α | Parameter of model | β | Parameter of model | γ | Variational parameter α | φ | Variational parameter β | Pdir | Dirichlet distribution | Mult(·) | Polynomial distribution |
|
查看原文
表 2CNN各层参数设置
Table2. Parameters of different layers of CNN
Type ofnetwork layer | Kf | F | S | P | Df |
---|
conv1 | 96 | 11×11 | 4 | 0 | 55×55×96 | Max-Pooling1 | - | 3×3 | 2 | 0 | 27×27×96 | conv2 | 256 | 5×5 | 1 | 2 | 27×27×256 | Max-Pooling2 | - | 3×3 | 2 | 0 | 13×13×256 | conv3 | 384 | 3×3 | 1 | 1 | 13×13×384 | conv4 | 384 | 3×3 | 1 | 1 | 13×13×384 | conv5 | 256 | 3×3 | 1 | 1 | 13×13×256 | Max-Pooling5 | - | 3×3 | 2 | 0 | 6×6×256 |
|
查看原文
表 3模型在Corel5K上的标注结果
Table3. Annotation results of different models on Corel5K
Model | Visual feature | AP | AR | F1 |
---|
PLSA-WORDS | TVS | 0.121 | 0.221 | 0.191 | fc7 | 0.217 | 0.275 | 0.269 | HGDM | TVS | 0.293 | 0.321 | 0.263 | fc7 | 0.305 | 0.364 | 0.297 | Proposed model | fc7 | 0.380 | 0.490 | 0.420 |
|
查看原文
表 4通用数据集上所有图像标注方法的标注结果
Table4. Annotation results of all image annotation models on common datasets
Model | Corel5K | IAPR TC-12 |
---|
AR | AP | F1 | AR | AP | F1 |
---|
MBRM | 0.25 | 0.24 | 0.25 | 0.23 | 0.24 | 0.24 | JEC | 0.32 | 0.27 | 0.29 | 0.29 | 0.28 | 0.29 | TagProp-ML | 0.37 | 0.31 | 0.34 | 0.25 | 0.48 | 0.33 | 2PKNN | 0.40 | 0.39 | 0.40 | 0.32 | 0.49 | 0.39 | CNN-R | 0.41 | 0.32 | 0.37 | 0.31 | 0.49 | 0.37 | CNN-MSE | 0.35 | 0.41 | 0.38 | 0.35 | 0.40 | 0.37 | Proposed model | 0.49 | 0.38 | 0.43 | 0.40 | 0.44 | 0.42 |
|
查看原文
张蕾, 蔡明. 融合卷积神经网络与主题模型的图像标注[J]. 激光与光电子学进展, 2019, 56(20): 201004. Lei Zhang, Ming Cai. Image Annotation Based on Convolutional Neural Network and Topic Model[J]. Laser & Optoelectronics Progress, 2019, 56(20): 201004.