CN112001431A - Efficient image classification method based on comb convolution - Google Patents

Efficient image classification method based on comb convolution Download PDF

Info

Publication number
CN112001431A
CN112001431A CN202010803534.3A CN202010803534A CN112001431A CN 112001431 A CN112001431 A CN 112001431A CN 202010803534 A CN202010803534 A CN 202010803534A CN 112001431 A CN112001431 A CN 112001431A
Authority
CN
China
Prior art keywords
convolution
comb
layer
mapping
layers
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.)
Granted
Application number
CN202010803534.3A
Other languages
Chinese (zh)
Other versions
CN112001431B (en
Inventor
周圆
李丹丹
霍树伟
曹涛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tianjin University
Original Assignee
Tianjin University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tianjin University filed Critical Tianjin University
Priority to CN202010803534.3A priority Critical patent/CN112001431B/en
Publication of CN112001431A publication Critical patent/CN112001431A/en
Application granted granted Critical
Publication of CN112001431B publication Critical patent/CN112001431B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a high-efficiency image classification method based on comb convolution, which comprises the following steps: preparing a data set and preprocessing the data; training a comb convolution-based deep convolutional neural network, comprising the following aspects: (1) comb convolutional layer internal operation: the linear mapping with low calculation cost is inserted into the standard convolution mapping in a parallel and staggered manner, so that one part of feature points on the output feature map are obtained by the standard convolution mapping, and the other part of feature points are obtained by the linear mapping; (2) interleaving mapping operation between comb convolution layers; (3) the whole network structure of the high-efficiency network model based on comb convolution; (4) training; and testing the classification effect of the model.

Description

Efficient image classification method based on comb convolution
Technical Field
The invention relates to the field of image processing and computer vision, in particular to a comb convolution-based efficient image classification method.
Background
Image classification belongs to the basic problem of various related research works in the field of computer vision, and is the basic composition of computer vision tasks. Image classification refers to the classification of an input image into a plurality of predefined classes, i.e. the input image is labeled with corresponding labels. In recent years, deep learning performs image classification by multi-layer information processing of a hierarchical structure, and compared with a conventional image classification method of learning shallow layers or obtaining image bottom-layer features, deep learning completely learns the hierarchical structural features of images from training data by using a set network structure, and can extract abstract features closer to high-level semantics of the images, so that the expression of image classification is far superior to that of the conventional method.
The image classification model based on the deep convolutional neural network has strong capability of hierarchically learning data characteristics. At present, an image classification algorithm based on a deep convolutional neural network model has become a main research algorithm in the field. However, as the performance of convolutional neural network models continues to approach the accuracy limit of computer vision tasks, the computational resources consumed also increase dramatically. For example, AlexNet[1]Requires 1.4X 1010FLOPs to process a single 224X 224 size image, and ResNet-152[2]Requires 2.26X 1011FLOPs. Therefore, in order to meet the requirements of the industry on the aspects of calculation overhead, real-time performance and the like, designing an efficient image classification method becomes a problem to be solved at present.
Recent studies of acceleration of convolution operations have focused on reducing redundancy of connections between feature map channels. Previous work, such as ResNeXt, has been done by building a structured sparse relationship of feature maps between channel dimensions[3]、Xception[4]、Shufflenet[5]、MobileNets[6]、Deep roots[7]、CondenseNet[8]And IGCNets[9]And the model efficiency is improved consistently. However, the above work ignores that there is a large amount of connection redundancy in the spatial dimension of the feature maps of adjacent layers, and further improvement of the model efficiency is limited.
Reference documents:
[1]Krizhevsky A,Sutskever I,Hinton G.ImageNet Classification with Deep Convolutional Neural Networks[C]//NIPS.Curran Associates Inc.2012.
[2]He K,Zhang X,Ren S,et al.Deep Residual Learning for Image Recognition[J].2015.
[3]Xie S,Girshick R,Dollár,Piotr,et al.Aggregated Residual Transformations for Deep Neural Networks[J].2016.
[4]Chollet F.Xception:Deep Learning with Depthwise Separable Convolutions[J].2016.
[5]Zhang X,Zhou X,Lin M,et al.ShuffleNet:An Extremely Efficient Convolutional Neural Network for Mobile Devices[J].2017.
[6]Howard A G,Zhu M,Chen B,et al.MobileNets:Efficient Convolutional Neural Networks for Mobile Vision Applications[J].2017.
[7]Ioannou Y,Robertson D,Cipolla R,et al.Deep Roots:Improving CNN Efficiency with Hierarchical Filter Groups[J].2016.
[8]Huang G,Liu S,Laurens V D M,et al.CondenseNet:An Efficient DenseNet using Learned Group Convolutions[J].2017.
[9]Zhang T,Qi G J,Xiao B,et al.Interleaved Group Convolutions for Deep Neural Networks[J].2017.
disclosure of Invention
In order to solve the above problems, the present invention provides an efficient image classification method based on comb convolution, which can significantly reduce the computation overhead generated by a network model during image classification by inserting linear mapping into standard convolution mapping in parallel and in an interlaced manner, so that the network model can classify images more efficiently and obtain a considerable classification effect. Meanwhile, the method provided by the invention can be used for end-to-end training, and is simple and efficient. Compared with the prior art, the method can solve the problem of intra-layer calculation redundancy, is compatible with an image classification method for reducing the inter-layer calculation redundancy, and has better expansibility. The technical scheme is as follows:
a high-efficiency image classification method based on comb convolution comprises the following steps:
firstly, preparing a data set and preprocessing the data;
secondly, training the comb convolution-based deep convolutional neural network, which comprises the following aspects:
(1) comb convolutional layer internal operation: the parallel and staggered insertion of the linear mapping with low calculation cost into the standard convolution mapping enables the feature points of one part on the output feature graph to be obtained by the standard convolution mapping and the feature points of the other part to be obtained by the linear mapping, and the method comprises the following steps:
the l-1 layer characteristic diagram xl-1While passing through standard convolution mapping
Figure BDA0002628273130000021
And linear mapping mu (-) to obtain the l-th layer feature map xlThe characteristic diagram of the jth channel of the ith layer is shown as
Figure BDA0002628273130000022
Which is related to the l-1 layer characteristic diagram xl-1The mapping relationship of (2) is defined as follows:
Figure BDA0002628273130000023
wherein the content of the first and second substances,
Figure BDA0002628273130000024
representing a profile corresponding to the jth channel of the ith layer
Figure BDA0002628273130000025
The convolution kernel of (a) is performed,
Figure BDA0002628273130000026
representation feature diagram xl-1And convolution kernel
Figure BDA0002628273130000027
Performing standard convolution mapping operation, wherein 1 represents an all-1 matrix with the same resolution as the characteristic diagram of the jth channel of the ith layer;
Figure BDA0002628273130000028
representing a profile corresponding to the jth channel of the ith layer
Figure BDA0002628273130000029
Binary chessboard maskA code matrix defined as follows:
Figure BDA00026282731300000210
wherein, (p, q) represents the characteristic diagram of the jth channel of the ith layer
Figure BDA00026282731300000211
Coordinates and signs of each point on
Figure BDA00026282731300000212
Representing a set of non-negative integers; at the same time, mu (x)l-1) A linear mapping operation is represented, defined as follows:
Figure BDA0002628273130000031
wherein, Cl-1Representing the total number of channels of the l-1 level feature map;
when convolution operation is performed
Figure BDA0002628273130000032
Denoted as a standard depth separable convolution mapping operation and linear mapping as μ (x)l-1)＝xl-1Then, the comb convolutional layer is converted into a comb convolutional layer fused with the depth separable convolution;
negating the binary chessboard mask matrix, repeating the calculation process of the formula (1) to obtain the characteristic diagram of the j +1 th channel of the l-th layer
Figure BDA0002628273130000033
And layer l-1 feature map xl-1The mapping relationship of (2); by analogy, continuously negating the binary chessboard mask matrix to finally obtain the characteristic diagram x of the l-th layerlAnd layer l-1 feature map xl-1The overall mapping relationship of (a);
(2) comb convolutional layer inter-layer interleaving mapping operation: the comb convolution operations are stacked while continuously inverting the binary mask tensor of the previous layer, i.e.
Maskl+1＝1-Maskl (4)
Wherein, MasklAnd Maskl+1Mask tensors of l-th and l + 1-th layers, 1 being MasklAll 1 tensors of the same size dimension; repeating the operation process of (1) comb-type convolution in the middle layer;
(3) the comb convolution-based high-efficiency network model overall network structure is as follows: based on the Xception network, 28 convolution layers are included in total, namely 2 layers of comb convolution layers and 26 layers of comb convolution layers which are fused with depth separable convolution, and an input image is input into a plurality of mutually stacked modules after the characteristics of the 2 layers of comb convolution layers are extracted; each module comprises 2 cascaded comb convolution layers fused with the depth separable convolution and 12 multiplied by 2 maximum pooling layers, 12 modules are stacked in total, and each module adopts residual connection operation in ResNet; finally, after 2 layers of comb-shaped convolution layers which are combined with depth separable convolution are processed, global average pooling is carried out, and then classification results are output through a full-connection layer; all the used convolution kernels have the size of 3 multiplied by 3, the sliding step length of the convolution kernels is 1, and a ReLU activation function is adopted;
(4) and inputting the image after data preprocessing into a constructed efficient network model based on comb convolution, setting parameters of the neural network model, updating the parameter weight of the network once after forward conduction and backward conduction, and iterating to obtain the trained efficient image classification network.
And thirdly, testing the classification effect of the model.
Drawings
FIG. 1 is a schematic diagram of the algorithm
FIG. 2 TensorFlow based implementation flow chart
FIG. 3 Overall network architecture diagram
FIG. 4 algorithm execution results
Detailed Description
In order to make the technical scheme of the invention clearer, the invention is further explained with reference to the attached drawings. The invention is realized by the following steps:
first, a data set is prepared and data is preprocessed.
The proposed method uses the CIFAR-100 classification dataset to train the image classification model. The CIFAR-100 dataset consists of 60000 32 × 32 color images, divided into 100 small categories according to image content, including: airplan, automobile, bird, cat, deer, dog, frog, horse, ship, hip, and truck, 10 subclasses under the 10 major classes, and the intersection between the classes is empty. Each class has 600 images per class, 500 training images and 100 test images per class. The experimental implementation used a standard data enhancement scheme, i.e., zero-padding the image by 4 pixels per edge, randomly cropping to generate an image of 32 × 32 size, and horizontally flipping with a probability of 0.5. During the training process, 5% of the training images were randomly drawn as the validation set.
And secondly, training the comb convolution-based deep convolutional neural network.
The proposed comb convolution principle is mainly divided into two major parts, see fig. 1, and its principle details are as follows:
1. comb convolutional layer internal operation:
linear mapping with low calculation cost is parallelly and alternately inserted into the standard convolution mapping by the operation in the comb convolution layer, so that half of feature points on an output feature map are obtained by the standard convolution mapping, and the other half of feature points are obtained by the linear mapping, thereby greatly reducing the operation expense of the convolution operation. The specific design is as follows:
(1) the l-1 layer characteristic diagram xl-1While passing through standard convolution mapping
Figure BDA0002628273130000041
And linear mapping mu (-) to obtain the l-th layer feature map xl. Wherein, the characteristic diagram of the jth channel of the ith layer is shown as
Figure BDA0002628273130000042
Which is related to the l-1 layer characteristic diagram xl-1The mapping relationship of (2) is defined as follows:
Figure BDA0002628273130000043
wherein the content of the first and second substances,
Figure BDA0002628273130000044
representing a profile corresponding to the jth channel of the ith layer
Figure BDA0002628273130000045
The convolution kernel of (a) is performed,
Figure BDA0002628273130000046
representation feature diagram xl-1And convolution kernel
Figure BDA0002628273130000047
A standard convolution mapping operation is performed with 1 representing the full 1 matrix at the same resolution as the profile for the jth channel of the ith layer.
Figure BDA0002628273130000048
Representing a profile corresponding to the jth channel of the ith layer
Figure BDA0002628273130000049
The binary chessboard mask matrix of (2) is defined as follows:
Figure BDA00026282731300000410
wherein, (p, q) represents the characteristic diagram of the jth channel of the ith layer
Figure BDA00026282731300000411
Coordinates and signs of each point on
Figure BDA00026282731300000412
Representing a set of non-negative integers. At the same time, mu (x)l-1) A linear mapping operation is represented, defined as follows:
Figure BDA00026282731300000413
wherein, Cl-1The total number of channels of the l-1 level profile is shown.
In particular, when convolution operations
Figure BDA00026282731300000414
Denoted as a standard depth separable convolution mapping operation and linear mapping as μ (x)l-1)＝xl-1The comb convolutional layer is then converted into a comb convolutional layer fused with the depth separable convolution.
(2) Negating the binary chessboard mask matrix, repeating the calculation process of the formula (1) to obtain the characteristic diagram of the j +1 th channel of the l-th layer
Figure BDA00026282731300000415
And layer l-1 feature map xl-1The mapping relationship of (2). By analogy, continuously negating the binary chessboard mask matrix to finally obtain the characteristic diagram x of the l-th layerlAnd layer l-1 feature map xl-1The overall mapping relationship of (1).
2. Comb convolutional layer inter-layer interleaving mapping operation:
stacking the comb convolution operation obtained in the above process, and continuously inverting the binary mask tensor of the previous layer, that is
Maskl+1＝1-Maskl (4)
Wherein, MasklAnd Maskl+1Mask tensors of l-th and l + 1-th layers, 1 being MasklAll 1 tensors with the same size dimension. And repeating the operation process of the comb convolution in the layer 1 to finally obtain the comb convolution-based efficient network model. An equivalent TensorFlow framework based efficient comb convolution implementation flow is shown in FIG. 2.
The overall network structure of the experiment is shown in figure 3. The overall network structure is inspired by the Xception network and comprises 28 convolutional layers in total, namely 2 comb convolutional layers and 26 comb convolutional layers fused with depth separable convolutions. After extracting features from the input image by the 2-layer comb convolution layer, the input image is input into a plurality of mutually stacked modules. Each module comprises 2 cascaded comb convolution layers fused with the depth separable convolution and 12 multiplied by 2 maximum pooling layers, 12 modules are stacked in total, and each module adopts residual connection operation in ResNet. And finally, after 2 layers of comb-shaped convolution layers with separable convolution and depth fusion are carried out, global average pooling is carried out, and then a classification result is output through a full-connection layer. The convolution kernels used are all 3 x 3 in size, the convolution kernel sliding step is 1, and the ReLU activation function is used.
In specific operation, an image after data preprocessing is input into a built neural network model, the initial learning rate of the neural network model is set to be 0.1, a Momentum optimization method with the parameter of 0.9 is adopted, and the weight attenuation rate is 10-4And after the training process reaches 50% and 75%, the learning rate is reduced by 10 times, the He initialization method is adopted for parameter initialization, and the size of the training batch is 100. And updating the parameter weight of the network once after forward conduction and backward conduction, and obtaining the trained efficient image classification network after 300 iterations.
And thirdly, testing the classification effect of the model.
During testing, the test set image is input into the network model, and please refer to fig. 4 for the test result. The invention compares the model classification results before and after the efficient comb convolution. From the experimental effect, the method can effectively reduce the operation complexity of the model and obtain the equivalent classification effect. Meanwhile, the method provided by the invention can be used for end-to-end training, and is simple and efficient. In addition, the method can solve the problem of intra-layer calculation redundancy, is compatible with an image classification method for reducing the inter-layer calculation redundancy, and has better expansibility.

Claims (3)

1. A high-efficiency image classification method based on comb convolution comprises the following steps:
firstly, preparing a data set and preprocessing the data;
and secondly, training the comb convolution-based deep convolutional neural network. The method comprises the following aspects:
(1) comb convolutional layer internal operation: the parallel and staggered insertion of the linear mapping with low calculation cost into the standard convolution mapping enables the feature points of one part on the output feature graph to be obtained by the standard convolution mapping and the feature points of the other part to be obtained by the linear mapping, and the method comprises the following steps:
the l-1 layer characteristic diagram xl-1While passing through standard convolution mapping
Figure FDA0002628273120000011
And linear mapping mu (-) to obtain the l-th layer feature map xlThe characteristic diagram of the jth channel of the ith layer is shown as
Figure FDA0002628273120000012
Which is related to the l-1 layer characteristic diagram xl-1The mapping relationship of (2) is defined as follows:
Figure FDA0002628273120000013
wherein the content of the first and second substances,
Figure FDA0002628273120000014
representing a profile corresponding to the jth channel of the ith layer
Figure FDA0002628273120000015
The convolution kernel of (a) is performed,
Figure FDA0002628273120000016
representation feature diagram xl-1And convolution kernel
Figure FDA0002628273120000017
Performing standard convolution mapping operation, wherein 1 represents an all-1 matrix with the same resolution as the characteristic diagram of the jth channel of the ith layer;
Figure FDA0002628273120000018
representing a profile corresponding to the jth channel of the ith layer
Figure FDA0002628273120000019
The binary chessboard mask matrix of (2) is defined as follows:
Figure FDA00026282731200000110
wherein, (p, q) represents the characteristic diagram of the jth channel of the ith layer
Figure FDA00026282731200000111
Coordinates and signs of each point on
Figure FDA00026282731200000112
Representing a set of non-negative integers; at the same time, mu (x)l-1) A linear mapping operation is represented, defined as follows:
Figure FDA00026282731200000113
wherein, Cl-1Representing the total number of channels of the l-1 level feature map;
when convolution operation is performed
Figure FDA00026282731200000114
Denoted as a standard depth separable convolution mapping operation and linear mapping as μ (x)l-1)＝xl -1Then, the comb convolutional layer is converted into a comb convolutional layer fused with the depth separable convolution;
negating the binary chessboard mask matrix, repeating the calculation process of the formula (1) to obtain the characteristic diagram of the j +1 th channel of the l-th layer
Figure FDA00026282731200000115
And layer l-1 feature map xl-1The mapping relationship of (2); by analogy, continuously negating the binary chessboard mask matrix to finally obtain the characteristic diagram x of the l-th layerlAnd layer l-1 feature map xl-1The overall mapping relationship of (a);
(2) comb convolutional layer inter-layer interleaving mapping operation: the comb convolution operations are stacked while continuously inverting the binary mask tensor of the previous layer, i.e.
Maskl+1＝1-Maskl (4)
Wherein, MasklAnd Maskl+1Mask tensors of l-th and l + 1-th layers, 1 being MasklAll 1 tensors of the same size dimension; repeating the operation process of (1) comb-type convolution in the middle layer;
(3) establishing a comb convolution-based efficient network model;
(4) training a high-efficiency image classification network;
and thirdly, testing the classification effect of the model.
2. The method of claim 1, wherein the efficient network model based on comb convolution is established by: based on the Xception network, 28 convolution layers are included in total, namely 2 layers of comb convolution layers and 26 layers of comb convolution layers which are fused with depth separable convolution, and an input image is input into a plurality of mutually stacked modules after the characteristics of the 2 layers of comb convolution layers are extracted; each module comprises 2 cascaded comb convolution layers fused with the depth separable convolution and 12 multiplied by 2 maximum pooling layers, 12 modules are stacked in total, and each module adopts residual connection operation in ResNet; finally, after 2 layers of comb-shaped convolution layers which are combined with depth separable convolution are processed, global average pooling is carried out, and then classification results are output through a full-connection layer; the convolution kernels used are all 3 x 3 in size, the convolution kernel sliding step is 1, and the ReLU activation function is used.
3. The method of claim 1, wherein the method of training an efficient image classification network is: and inputting the image after data preprocessing into a constructed efficient network model based on comb convolution, setting parameters of the neural network model, updating the parameter weight of the network once after forward conduction and backward conduction, and iterating to obtain the trained efficient image classification network.
CN202010803534.3A 2020-08-11 2020-08-11 Efficient image classification method based on comb convolution Active CN112001431B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010803534.3A CN112001431B (en) 2020-08-11 2020-08-11 Efficient image classification method based on comb convolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010803534.3A CN112001431B (en) 2020-08-11 2020-08-11 Efficient image classification method based on comb convolution

Publications (2)

Publication Number Publication Date
CN112001431A true CN112001431A (en) 2020-11-27
CN112001431B CN112001431B (en) 2022-06-28

Family

ID=73464070

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010803534.3A Active CN112001431B (en) 2020-08-11 2020-08-11 Efficient image classification method based on comb convolution

Country Status (1)

Country Link
CN (1) CN112001431B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115170917A (en) * 2022-06-20 2022-10-11 美的集团(上海)有限公司 Image processing method, electronic device, and storage medium

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103489454A (en) * 2013-09-22 2014-01-01 浙江大学 Voice endpoint detection method based on waveform morphological characteristic clustering
CN110211064A (en) * 2019-05-21 2019-09-06 华南理工大学 A kind of mixing degeneration text image restoration methods based on margin guide
CN110427990A (en) * 2019-07-22 2019-11-08 浙江理工大学 A kind of art pattern classification method based on convolutional neural networks
CN110781912A (en) * 2019-09-10 2020-02-11 东南大学 Image classification method based on channel expansion inverse convolution neural network
CN110799991A (en) * 2017-06-28 2020-02-14 奇跃公司 Method and system for performing simultaneous localization and mapping using a convolutional image transform
CN110995382A (en) * 2019-11-29 2020-04-10 中国科学院微电子研究所 Interference avoidance communication model based on meta-learning and training method thereof
CN111340907A (en) * 2020-03-03 2020-06-26 曲阜师范大学 Text-to-image generation method of self-adaptive attribute and instance mask embedded graph

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103489454A (en) * 2013-09-22 2014-01-01 浙江大学 Voice endpoint detection method based on waveform morphological characteristic clustering
CN110799991A (en) * 2017-06-28 2020-02-14 奇跃公司 Method and system for performing simultaneous localization and mapping using a convolutional image transform
CN110211064A (en) * 2019-05-21 2019-09-06 华南理工大学 A kind of mixing degeneration text image restoration methods based on margin guide
CN110427990A (en) * 2019-07-22 2019-11-08 浙江理工大学 A kind of art pattern classification method based on convolutional neural networks
CN110781912A (en) * 2019-09-10 2020-02-11 东南大学 Image classification method based on channel expansion inverse convolution neural network
CN110995382A (en) * 2019-11-29 2020-04-10 中国科学院微电子研究所 Interference avoidance communication model based on meta-learning and training method thereof
CN111340907A (en) * 2020-03-03 2020-06-26 曲阜师范大学 Text-to-image generation method of self-adaptive attribute and instance mask embedded graph

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
安坤: "基于神经网络的梳状谱信号峰均比优化技术", 《中国电子科学研究院学报》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115170917A (en) * 2022-06-20 2022-10-11 美的集团(上海)有限公司 Image processing method, electronic device, and storage medium
CN115170917B (en) * 2022-06-20 2023-11-07 美的集团(上海)有限公司 Image processing method, electronic device and storage medium

Also Published As

Publication number Publication date
CN112001431B (en) 2022-06-28

Similar Documents

Publication Publication Date Title
Zhang et al. Latentgnn: Learning efficient non-local relations for visual recognition
CN111275107A (en) Multi-label scene image classification method and device based on transfer learning
CN109063719B (en) Image classification method combining structure similarity and class information
Khasoggi et al. Efficient mobilenet architecture as image recognition on mobile and embedded devices
CN110046249A (en) Training method, classification method, system, equipment and the storage medium of capsule network
CN110222718B (en) Image processing method and device
CN111259904B (en) Semantic image segmentation method and system based on deep learning and clustering
CN111488917A (en) Garbage image fine-grained classification method based on incremental learning
Wang et al. Pruning blocks for CNN compression and acceleration via online ensemble distillation
Yue et al. Face recognition based on histogram equalization and convolution neural network
CN111461175A (en) Label recommendation model construction method and device of self-attention and cooperative attention mechanism
Liu et al. A multiscale image denoising algorithm based on dilated residual convolution network
CN114882278A (en) Tire pattern classification method and device based on attention mechanism and transfer learning
Ale et al. Lightweight deep learning model for facial expression recognition
Luan et al. Sunflower seed sorting based on convolutional neural network
CN112001431B (en) Efficient image classification method based on comb convolution
Ge et al. Super-resolution reconstruction of biometric features recognition based on manifold learning and deep residual network
CN110728352A (en) Large-scale image classification method based on deep convolutional neural network
CN114492581A (en) Method for classifying small sample pictures based on transfer learning and attention mechanism element learning application
Abdulsattar et al. Facial expression recognition using transfer learning and fine-tuning strategies: A comparative study
CN113239949A (en) Data reconstruction method based on 1D packet convolutional neural network
CN116109868A (en) Image classification model construction and small sample image classification method based on lightweight neural network
Sang et al. Image recognition based on multiscale pooling deep convolution neural networks
Nie et al. Multi-scale CNNs for 3D model retrieval
Xia et al. Efficient synthesis of compact deep neural networks

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
GR01 Patent grant
GR01 Patent grant