CN111767964A - Improved DenseNet-based multi-channel feature re-labeling image classification method - Google Patents
Improved DenseNet-based multi-channel feature re-labeling image classification method Download PDFInfo
- Publication number
- CN111767964A CN111767964A CN202010649381.1A CN202010649381A CN111767964A CN 111767964 A CN111767964 A CN 111767964A CN 202010649381 A CN202010649381 A CN 202010649381A CN 111767964 A CN111767964 A CN 111767964A
- Authority
- CN
- China
- Prior art keywords
- training
- data set
- network
- densenet
- training data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 22
- 238000002372 labelling Methods 0.000 title claims abstract description 11
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000013145 classification model Methods 0.000 claims abstract description 16
- 230000000694 effects Effects 0.000 claims abstract description 5
- 238000011156 evaluation Methods 0.000 claims abstract description 4
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000000605 extraction Methods 0.000 claims description 4
- 230000001174 ascending effect Effects 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000001514 detection method Methods 0.000 abstract description 4
- 239000010410 layer Substances 0.000 description 16
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000011160 research Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000004904 shortening Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a multichannel feature re-labeling image classification method based on improved DenseNet, which comprises the following steps: step S1, collecting a training data set, and preprocessing the training data set to obtain a preprocessed training data set; step S2: based on the multi-channel characteristic heavy-mark intensive connection network, building a classification model; step S3, training a classification model according to a training data set, introducing training effect evaluation indexes Acc and F1-Score, evaluating the model training process in real time, storing the training model and data in real time, and drawing according to a final iteration result to obtain a trained classification model; step S4: and inputting the data set to be classified into the trained classification model to obtain a classification result. The invention effectively improves the image recognition classification detection precision.
Description
Technical Field
The invention relates to the technical field of image recognition and classification, in particular to a multichannel feature re-labeling image classification method based on improved DenseNet.
Background
In recent years, Convolutional Neural Networks (CNN) have made a series of contributions to computer vision tasks, and have made great progress in the fields of image recognition, target detection, medical images, semantic segmentation, and the like. Research on convolutional neural networks is also ongoing, and a large set of efficient and good models achieve significant performance gains in various computer vision tasks, such as AlexNet, VGGNet, google lenet, ResNet, and densneet, in the corner.
At present, the convolutional neural network is developing towards deeper depth and wider width due to the promotion of hardware devices such as GPU. In the research in the depth direction of the CNN network, the VGG19 network proves that the depth of the network is a key part for improving the performance of the architecture. And the high way and ResNet structure realizes the network depth to more than 100 layers by adopting the data bypass and the skip connection idea. Henceforth, the DenseNet network proposed a completely new dense connection model, i.e. the flow of information between all layers connects each layer to another layer in a feed-forward manner, so that each layer in the network accepts as input the characteristics of all layers before it. Certainly, some researchers also explore the influence of the network channel width on the CNN network performance, and GoogleNet enhances the feature extraction capability by using a multi-scale convolution kernel on a single-layer convolution layer, so as to widen the network channel. The Wide residual network follows the principle that depth is not the only important parameter, and by shortening the depth and increasing the functional characteristics of each layer, a wider neural network is realized.
However, researches show that the deeper and wider the network structure is, the better optimization of the performance cannot be obtained, and the problems of poor convergence, severe overfitting and gradient disappearance are easy to occur. Meanwhile, when the network structure reaches a certain depth and width, the performance of the network structure tends to average precision. These problems, which have been a hindrance to further advance of image recognition classification, have been initiated by many researchers for new structural studies.
Disclosure of Invention
In view of this, the present invention provides an improved DenseNet-based multi-channel feature re-labeling image classification method, which effectively improves the image recognition classification detection accuracy.
In order to achieve the purpose, the invention adopts the following technical scheme:
a multi-channel feature re-labeling image classification method based on improved DenseNet comprises the following steps:
step S1, collecting a training data set, and preprocessing the training data set to obtain a preprocessed training data set;
step S2: based on the multi-channel characteristic heavy-mark intensive connection network, building a classification model;
step S3, training a classification model according to a training data set, introducing training effect evaluation indexes Acc and F1-Score, evaluating the model training process in real time, storing the training model and data in real time, and drawing according to a final iteration result to obtain a trained classification model;
step S4: and inputting the data set to be classified into the trained classification model to obtain a classification result.
Further, the training data set preprocessing specifically includes: and performing stretching rotation transformation and data expansion on the data sample by using the picture generator to obtain a sample set training larger than the original data set.
Further, the step S2 is specifically:
step S21: reconstructing an ascending and descending dimensional Bottleneck structure based on a DenseNet-BC network 1 x 1conv +3 x 3conv to form a Reconstructed Bottleneck structure suitable for the same dimension 3 x 3conv + 3conv of a low-depth and narrow-width network;
step S22: the method comprises the following steps of forming a dense connection block DenseBlock by using a multichannel parallel connection DenseBlock while forming the dense connection block structure of a multipath;
step S23: and introducing an SE structure with characteristic recalibration capability after the multipath dense connection block, and marking important characteristics.
Further, the step S21 is specifically: the method is characterized in that a Bottleneck infinitesimal structure in the Densenet is improved, 1 × 1conv in the original structure is replaced by 3 × 3conv, and two convolution layers with the size of 3 × 3 are stacked to replace one convolution layer with the size of 5 × 5; and the 4k-k lifting dimension feature extraction mode in the original micro-element structure is directly changed into a 4k-4k same-dimension structure according to the structure with shallow depth and narrow width, so that the network channel is further widened.
Further, the step S22 is specifically: parallel dense blocks with similar depths are added to DenseBlocks with improved reinforced Bottleneck structures to increase the width of the network, namely, the network is connected in parallel by multiple channels;
since the number n of multipaths and the depth d of the network do not exhibit a single inverse relationship, the values of both in the different training data sets need to be set through a small range to determine the appropriate multipath value;
for a single densebclk:
let x0Is input, H1Is x0The output is x1，H2Is x0And x1I.e. L-layer DenseNet network has
And directly connecting, connecting all the previous layers as input, and expressing the following expression:
xl＝Hl([x0,x1,…,xl-1])
then, the DenseBlcok for multipath is expressed as:
xj＝xl1+xl2+…+xln＝Hl1([x0,x1,…,xl-1])+Hl2([x0,x1,…,xl-1])+…+Hln([x0,x1,…,xl-1])，
wherein xjIs a characteristic of the output of multiple parallel channels, HlnThe characteristics output by the nth individual DenseBlcok are expressed.
Compared with the prior art, the invention has the following beneficial effects:
the invention effectively improves the image recognition classification detection precision.
Drawings
FIG. 1 is an overall block diagram of an embodiment of the present invention;
FIG. 2 is a flow chart of the overall structure improvement in one embodiment of the present invention;
fig. 3 is a network training flow chart based on multi-channel feature re-labeling of the DenseNet network in an embodiment of the present invention.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
Referring to fig. 1, the present invention provides a multi-channel feature re-labeling image classification method based on improved DenseNet, which includes the following steps:
s1, collecting public training data sets such as CIFAR-10/100, SVHN, MNIST and the like, and performing transformation such as stretching rotation and the like and data expansion on the data samples by using a picture generator to obtain a sample set training larger than the original data set, improving the network image recognition capability to obtain a model fitting effect and obtaining a preprocessed training data set;
step S2: based on the multi-channel characteristic heavy-mark intensive connection network, building a classification model;
step S3, training a classification model according to a training data set, introducing training effect evaluation indexes Acc and F1-Score, evaluating the model training process in real time, storing the training model and data in real time, and drawing according to a final iteration result to obtain a trained classification model;
step S4: and inputting the data set to be classified into the trained classification model to obtain a classification result.
In this embodiment, the step S2 specifically includes:
step S21: reconstructing an ascending and descending dimensional Bottleneck structure based on a DenseNet-BC network 1 x 1conv +3 x 3conv to form a Reconstructed Bottleneck structure suitable for the same dimension 3 x 3conv + 3conv of a low-depth and narrow-width network;
step S22: the method comprises the following steps of forming a dense connection block DenseBlock by using a multichannel parallel connection DenseBlock while forming the dense connection block DenseBlock by a constrained Bottleneck structure, and enhancing the characteristic multiplexing capability;
step S23: and an SE structure with the characteristic recalibration capability is introduced after the multi-path dense connecting blocks, important characteristics are marked, redundant characteristics are reduced, and the classification speed is improved.
In this embodiment, step S21 specifically includes: the method is characterized in that a Bottleneck infinitesimal structure in the Densenet is improved, 1 × 1conv in the original structure is replaced by 3 × 3conv, and two convolution layers with the size of 3 × 3 are stacked to replace one convolution layer with the size of 5 × 5; and the 4k-k lifting dimension feature extraction mode in the original micro-element structure is directly changed into a 4k-4k same-dimension structure according to the structure with shallow depth and narrow width, so that the network channel is further widened.
In this embodiment, the step S22 specifically includes: parallel dense blocks with similar depths are added to DenseBlocks with the improved ReconstructedButtleneck structure to increase the width of the network, namely the network is connected in parallel with multiple channels;
since the number n of multipaths and the depth d of the network do not exhibit a single inverse relationship, the values of both in the different training data sets need to be set through a small range to determine the appropriate multipath value;
for a single densebclk:
let x0Is input, H1Is x0The output is x1，H2Is x0And x1I.e. L-layer DenseNet network has
And directly connecting, connecting all the previous layers as input, and expressing the following expression:
xl＝Hl([x0,x1,…,xl-1])
then, the DenseBlcok for multipath is expressed as:
xj＝xl1+xl2+…+xln＝Hl1([x0,x1,…,xl-1])+Hl2([x0,x1,…,xl-1])+…+Hln([x0,x1,…,xl-1])，
wherein xjIs a characteristic of the output of multiple parallel channels, HlnThe characteristics output by the nth individual DenseBlcok are expressed.
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.
Claims (5)
1. A multi-channel feature re-labeling image classification method based on improved DenseNet is characterized by comprising the following steps:
step S1, collecting a training data set, and preprocessing the training data set to obtain a preprocessed training data set;
step S2: based on the multi-channel characteristic heavy-mark intensive connection network, building a classification model;
step S3, training a classification model according to a training data set, introducing training effect evaluation indexes Acc and F1-Score, evaluating the model training process in real time, storing the training model and data in real time, and drawing according to a final iteration result to obtain a trained classification model;
step S4: and inputting the data set to be classified into the trained classification model to obtain a classification result.
2. The improved DenseNet-based multi-channel feature re-labeling image classification method according to claim 1, characterized in that the training data set is preprocessed specifically as follows: and performing stretching rotation transformation and data expansion on the data sample by using the picture generator to obtain a sample set training larger than the original data set.
3. The improved DenseNet-based multi-channel feature relabeling image classification method according to claim 1, wherein the step S2 specifically comprises:
step S21: reconstructing an ascending and descending dimensional Bottleneck structure based on a DenseNet-BC network 1 x 1conv +3 x 3conv to form a Reconstructed Bottleneck structure suitable for the same dimension 3 x 3conv + 3conv of a low-depth and narrow-width network;
step S22: the method comprises the following steps of forming a dense connection block DenseBlock by using a multichannel parallel connection DenseBlock while forming the dense connection block structure of a multipath;
step S23: and introducing an SE structure with characteristic recalibration capability after the multipath dense connection block, and marking important characteristics.
4. The improved DenseNet-based multi-channel feature relabeling image classification method according to claim 3, wherein the step S21 is specifically: the method is characterized in that a Bottleneck infinitesimal structure in the Densenet is improved, 1 × 1conv in the original structure is replaced by 3 × 3conv, and two convolution layers with the size of 3 × 3 are stacked to replace one convolution layer with the size of 5 × 5; and the 4k-k lifting dimension feature extraction mode in the original micro-element structure is directly changed into a 4k-4k same-dimension structure according to the structure with shallow depth and narrow width, so that the network channel is further widened.
5. The improved DenseNet-based multi-channel feature re-labeling image classification method according to claim 1, characterized in that: the step S22 specifically includes: parallel dense blocks with similar depths are added to DenseBlocks with improved reinforced Bottleneck structures to increase the width of the network, namely, the network is connected in parallel by multiple channels;
since the number n of multipaths and the depth d of the network do not exhibit a single inverse relationship, the values of both in the different training data sets need to be set through a small range to determine the appropriate multipath value;
for a single densebclk:
let x0Is input, H1Is x0The output is x1，H2Is x0And x1I.e. L-layer DenseNet network has
And directly connecting, connecting all the previous layers as input, and expressing the following expression:
xl＝Hl([x0,x1,…,xl-1])
then, the DenseBlcok for multipath is expressed as:
xj＝xl1+xl2+…+xln＝Hl1([x0,x1,…,xl-1])+Hl2([x0,x1,…,xl-1])+…+Hln([x0,x1,…,xl-1])，
wherein xjIs a characteristic of the output of multiple parallel channels, HlnThe characteristics output by the nth individual DenseBlcok are expressed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010649381.1A CN111767964A (en) | 2020-07-08 | 2020-07-08 | Improved DenseNet-based multi-channel feature re-labeling image classification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010649381.1A CN111767964A (en) | 2020-07-08 | 2020-07-08 | Improved DenseNet-based multi-channel feature re-labeling image classification method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111767964A true CN111767964A (en) | 2020-10-13 |
Family
ID=72725166
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010649381.1A Pending CN111767964A (en) | 2020-07-08 | 2020-07-08 | Improved DenseNet-based multi-channel feature re-labeling image classification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111767964A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20170091963A (en) * | 2016-02-02 | 2017-08-10 | 삼성전자주식회사 | Gesture classification apparatus and method using electromyogram signals |
| CN110276148A (en) * | 2019-06-27 | 2019-09-24 | 上海交通大学 | The feature extraction of micro-structure dimensionality reduction and reconstruct implementation method based on self-encoding encoder |
| CN110458842A (en) * | 2019-06-26 | 2019-11-15 | 南京邮电大学 | Brain tumor dividing method based on the three-dimensional intensive connection network of binary channels |
| CN110472668A (en) * | 2019-07-22 | 2019-11-19 | 华北电力大学（保定） | A kind of image classification method |
| CN110570431A (en) * | 2019-09-18 | 2019-12-13 | 东北大学 | Medical image segmentation method based on improved convolutional neural network |
| US10635951B1 (en) * | 2018-10-24 | 2020-04-28 | Alibaba Group Holding Limited | Fast computation of a convolutional neural network |
-
2020
- 2020-07-08 CN CN202010649381.1A patent/CN111767964A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20170091963A (en) * | 2016-02-02 | 2017-08-10 | 삼성전자주식회사 | Gesture classification apparatus and method using electromyogram signals |
| US10635951B1 (en) * | 2018-10-24 | 2020-04-28 | Alibaba Group Holding Limited | Fast computation of a convolutional neural network |
| CN110458842A (en) * | 2019-06-26 | 2019-11-15 | 南京邮电大学 | Brain tumor dividing method based on the three-dimensional intensive connection network of binary channels |
| CN110276148A (en) * | 2019-06-27 | 2019-09-24 | 上海交通大学 | The feature extraction of micro-structure dimensionality reduction and reconstruct implementation method based on self-encoding encoder |
| CN110472668A (en) * | 2019-07-22 | 2019-11-19 | 华北电力大学（保定） | A kind of image classification method |
| CN110570431A (en) * | 2019-09-18 | 2019-12-13 | 东北大学 | Medical image segmentation method based on improved convolutional neural network |
Non-Patent Citations (3)
| Title |
|---|
| 张彤: "基于深度学习的图像超分辨率重构研究", 《中国优秀硕士学位论文全文数据库（信息科技I辑）》 * |
| 张锡英 等: "融合STN和DenseNet的深度学习网络及其应用", 《计算机工程与应用》 * |
| 郭玉荣 等: "端到端双通道特征重标定DenseNet图像分类", 《中国图象图形学报》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112308158B (en) | Multi-source field self-adaptive model and method based on partial feature alignment | |
| CN111325751B (en) | CT image segmentation system based on attention convolution neural network | |
| CN108509978B (en) | Multi-class target detection method and model based on CNN (CNN) multi-level feature fusion | |
| CN111695467B (en) | Spatial spectrum full convolution hyperspectral image classification method based on super-pixel sample expansion | |
| CN115049936B (en) | High-resolution remote sensing image-oriented boundary enhanced semantic segmentation method | |
| CN113344188A (en) | Lightweight neural network model based on channel attention module | |
| CN114821249B (en) | Vehicle weight recognition method based on grouping aggregation attention and local relation | |
| CN108596044B (en) | Pedestrian detection method based on deep convolutional neural network | |
| CN116740439A (en) | Crowd counting method based on trans-scale pyramid convertors | |
| CN114693577B (en) | Infrared polarized image fusion method based on Transformer | |
| Al-Amaren et al. | RHN: A residual holistic neural network for edge detection | |
| CN115100165A (en) | Colorectal cancer T staging method and system based on tumor region CT image | |
| CN114972753A (en) | Lightweight semantic segmentation method and system based on context information aggregation and assisted learning | |
| CN113239949A (en) | Data reconstruction method based on 1D packet convolutional neural network | |
| CN117058386A (en) | Asphalt road crack detection method based on improved deep Labv3+ network | |
| CN111767964A (en) | Improved DenseNet-based multi-channel feature re-labeling image classification method | |
| CN115937693A (en) | Road identification method and system based on remote sensing image | |
| CN113192076B (en) | MRI brain tumor image segmentation method combining classification prediction and multi-scale feature extraction | |
| CN113691863B (en) | Lightweight method for extracting video key frames | |
| CN113052810B (en) | Small medical image focus segmentation method suitable for mobile application | |
| CN113436198A (en) | Remote sensing image semantic segmentation method for collaborative image super-resolution reconstruction | |
| CN114565511A (en) | Lightweight image registration method, system and device based on global homography estimation | |
| Cao et al. | Deep feature interactive aggregation network for single image deraining | |
| CN116698410B (en) | Rolling bearing multi-sensor data monitoring method based on convolutional neural network | |
| CN115457611B (en) | Vein recognition method based on characteristic decoupling network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20201013 |
|
| RJ01 | Rejection of invention patent application after publication |