CN111832621A - Image classification method and system based on dense multipath convolutional network - Google Patents
Image classification method and system based on dense multipath convolutional network Download PDFInfo
- Publication number
- CN111832621A CN111832621A CN202010531567.7A CN202010531567A CN111832621A CN 111832621 A CN111832621 A CN 111832621A CN 202010531567 A CN202010531567 A CN 202010531567A CN 111832621 A CN111832621 A CN 111832621A
- Authority
- CN
- China
- Prior art keywords
- dense
- module
- network
- dense multipath
- multipath
- 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 31
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000013145 classification model Methods 0.000 claims abstract description 12
- 230000001131 transforming effect Effects 0.000 claims abstract description 5
- 230000004931 aggregating effect Effects 0.000 claims abstract description 4
- 230000002776 aggregation Effects 0.000 claims description 16
- 238000004220 aggregation Methods 0.000 claims description 16
- 230000009466 transformation Effects 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 description 12
- 238000013527 convolutional neural network Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 101100153586 Caenorhabditis elegans top-1 gene Proteins 0.000 description 2
- 101100370075 Mus musculus Top1 gene Proteins 0.000 description 2
- 230000003321 amplification Effects 0.000 description 2
- 238000003199 nucleic acid amplification method Methods 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 238000000638 solvent extraction Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000012795 verification Methods 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
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention provides a picture classification method and a picture classification system based on a dense multipath convolutional network, which comprise the following steps: constructing a basic unit formed by grouping, transforming and aggregating, constructing a new dense module based on the basic unit, and replacing the original dense module in the densely connected convolution DenseNet network framework with the new dense module to obtain a dense multipath convolution network; using the image data with the marked categories as training data, updating the weight in the dense multipath convolution network through gradient back propagation, and training the dense multipath convolution network to obtain an image classification model; and inputting the picture data to be classified into the classification model to obtain a classification result of the picture data to be classified. The dense multipath convolution network provided by the invention improves and optimizes the basic module of the DenseNet network to obtain the characteristic of stronger expressive force.
Description
Technical Field
The method belongs to the field of computer vision, and particularly relates to a problem of image classification in computer vision.
Background
Convolutional Neural Networks (CNNs) make a major breakthrough in the field of picture classification. At present, many image classification methods are further improved on the basis of CNN, and classification performance is improved by methods of increasing network width or depth, novel network module structures and the like. The resenext network and the DenseNet network are convolutional neural networks with excellent classification effects proposed in the last two years.
The resenext network is a variant of the residual network Resnet, and for convenience, is hereinafter referred to as the residual network. The residual error network (resenxt) designs a homogeneous, multi-path Transformation (Transformation) module, which proposes a dimension different from depth and width: cardinality (Cardinality), a new network resenext is constructed. The network achieves the 2 nd achievement on the classification task of identifying challenge games on a large scale in the year 2017 by ImageNet, and experiments show that the increasing of the base number of the network is more effective than the increasing of the depth and the width of the network.
The dense network densnet designs a densely connected module in which each layer is connected to all other layers based on the observation that "short circuit connections between layers (skipconnections) help to make the network deeper, more accurate, and more efficient in training". In a conventional convolutional network, there are only L direct connections in the L layer (the connection of the current layer to the next layer counts once), but there are L (L +1)/2 direct connections in the DenseNet. For each layer, all layers in the module in front of it are their inputs, and that layer is the input of all layers in the module behind. The connection mode can reduce gradient disappearance, strengthen feature propagation, enhance feature reuse and greatly reduce the number of parameters. DenseNet achieved the best performance at that time on the CIFAR-10, CIFAR-100, SVHN and ImageNet datasets.
DenseNet has many advantages, and the basic unit is from ResNet, and the representation capability of the basic unit is weaker than that of ResNeXt division, transformation and aggregation.
Disclosure of Invention
The invention improves and optimizes DenseNet. The dense multipath convolution network provided by the invention improves and optimizes the basic module of the DenseNet network to obtain the characteristic of stronger expressive force.
The network comprises the following features:
(1) the dense multipath convolutional network is the same as the dense convolutional network, and the original image is taken as the input of the network.
(2) The dense multipath convolutional network follows the partitioning and transformation steps of the resenext basic Block, thus proposing a new basic Block _ XT as shown in fig. 1. A new dense module DenseBlock _ XT is designed as shown in fig. 2 based on the proposed Block _ XT module. In the DenseNet framework, the DenseNet old dense module is replaced with the new dense module DenseBlock _ XT.
(3) The dense multipath convolutional network has 3 superparameters, which are: the base number, depth and width, and the 3 hyper-parameters are adjusted, so that the performance of the network can be adjusted to achieve good classification effect.
(4) During the training process, the dense multipath convolutional network uses the cross entropy of the model predicted value and the true value as a loss function. And simultaneously, replacing the ReLU used by the original DenseNet with an activation function ELU with better performance.
Aiming at the defects of the prior art, the invention provides a picture classification method based on a dense multipath convolutional network, which comprises the following steps:
step 1, constructing a basic unit formed by grouping, transforming and aggregating, constructing a new dense module based on the basic unit, and replacing an original dense module in a densely connected convolution network frame with the new dense module to obtain a dense multipath convolution network;
step 2, using the image data with the marked categories as training data, updating the weight in the dense multipath convolution network through gradient back propagation, and training the dense multipath convolution network to obtain an image classification model;
and 3, inputting the picture data to be classified into the classification model to obtain a classification result of the picture data to be classified.
The image classification method based on the dense multipath convolutional network is characterized in that the aggregation type is a serial aggregation structure.
The picture classification method based on the dense multipath convolutional network, wherein the step 1 comprises the following steps: and adding a normalization module, a scale adjustment layer and an activation function layer in front of the basic module to obtain the new dense module.
The picture classification method based on the dense multipath convolutional network comprises the following specific steps of: training data is input into the dense multipath convolutional network to obtain a model predicted value, loss is obtained by comparing the model predicted value with the marked category, gradient is calculated, and weight in the dense multipath convolutional network is updated through layer-by-layer reverse propagation.
The picture classification method based on the dense multipath convolutional network is characterized in that the cross entropy of the model prediction value and the marked category is used as the loss function.
The invention also provides a picture classification system based on the dense multipath convolutional network, which comprises the following steps:
the method comprises the following steps that 1, a basic unit formed by grouping, transformation and aggregation is constructed, a new dense module is constructed based on the basic unit, and an original dense module in a densely connected convolution network frame is replaced by the new dense module to obtain a dense multipath convolution network;
the module 2 uses the image data with the marked categories as training data, updates the weight in the dense multipath convolution network through gradient back propagation, trains the dense multipath convolution network and obtains an image classification model;
and the module 3 inputs the image data to be classified into the classification model to obtain a classification result of the image data to be classified.
The image classification system based on the dense multipath convolutional network is characterized in that the aggregation type is a serial aggregation structure.
The picture classification system based on the dense multipath convolutional network, wherein the module 1 comprises: and adding a normalization module, a scale adjustment layer and an activation function layer in front of the basic module to obtain the new dense module.
The picture classification system based on the dense multipath convolutional network comprises the following steps of: training data is input into the dense multipath convolutional network to obtain a model predicted value, loss is obtained by comparing the model predicted value with the marked category, gradient is calculated, and weight in the dense multipath convolutional network is updated through layer-by-layer reverse propagation.
The picture classification system based on the dense multipath convolutional network is characterized in that the cross entropy of the model prediction value and the marked category is used as the loss function.
According to the scheme, the invention has the advantages that: the dense multipath convolution network (DenseXT) provided by the invention has strong representation capability, enables information of each layer to be fully transmitted, and can obtain better picture characteristics than ResNeXt and DenseNet, thereby obtaining better picture classification effect.
Drawings
Fig. 1 is a structural diagram of a newly designed basic module Block _ XT;
fig. 2 newly designs the structure of the dense module DenseBlock _ XT.
Detailed Description
The invention is based on the idea of ResNeXt basic module multi-path transformation, improves and optimizes the basic module of a DenseNet network (Densely Connected Convolutional network), and designs a new Convolutional neural network which is named as dense multi-path Convolutional network (DenseXT).
The dense multipath convolution network provided by the invention is mainly used for acquiring image features with stronger expressive force so as to improve the performance of a plurality of tasks such as image classification and the like. The dense multipath convolution network combines the respective advantages of DenseNet and ResNeXt, overcomes the disadvantages of DenseNet and ResNeXt, and designs a new basic module Block _ XT and a dense module DenseBlock _ XT, so that the dense multipath convolution network has strong representation capability and enables information of each layer to be fully transmitted. Experimental results show that the DenseXT network designed by the invention can obtain higher classification precision than DenseNet by using fewer parameters and smaller models, and shows that DenseXT has stronger expression capability and better classification performance than DenseNet.
In order to make the aforementioned features and effects of the present invention more comprehensible, embodiments accompanied with figures are described in detail below.
The mainstream image classification method at present is mainly based on a convolutional neural network and its deformation. The methods adopt the idea of transfer learning, utilize a convolutional neural network model pre-trained on a large-scale image classification data set, and retrain or fine tune on a specific data set.
In order to improve the picture classification precision, the basic module of the DenseNet network is improved and optimized, and a new convolutional neural network, namely a dense multipath convolutional network, is designed. The use of a dense multipath convolutional network significantly improves the picture classification accuracy.
The dense multipath convolution network is mainly based on the dividing and transforming steps of the ResNeXt basic module, and preliminarily solves the problem of weak expression of the DenseNet basic module on the premise of keeping information capable of being spread in each layer. A new DenseXT basic module Block _ XT is proposed as shown in fig. 1, the structure of the left and right sides of fig. 1 being equivalent, the right side being a simplified version of the left side. The left 128 channels, divided into 32 groups of 4 channels, were output through 4 filters of 3 × 3. The simplified representation on the left is on the right, with group 32 representing a division into 32 groups, 128 channels for the inputs and 128 channels for the outputs. A new dense module DenseBlock _ XT is designed based on the proposed Block _ XT module, the structure of which is shown in FIG. 2, Batch normalization of Batch Nortch, Scale: Scale, ELU: an activation function. Using the overall framework of DenseNet and the new dense module DenseBlock _ XT, a dense multipath convolutional network is obtained. The structure is shown in the following table, in the table, parameters K, C can be adjusted according to needs, and the numbers of the Dense Block and the Transition Layer can also be adjusted:
TABLE 1 DenseXT architecture
Dense multipath convolutional networks capture global content information using a deeper convolutional neural network structure. During the forward propagation of the neural network, each position of the back layer will be fully connected to the adjacent region in its front layer, thus being affected by all values in that region. Therefore, as the number of layers of the neural network increases, the receptive field also increases continuously. The convolutional neural network structure used by the dense multipath convolutional network comprises a plurality of convolutional kernels and a plurality of down-sampling operations with the step size of 2, so that a larger receptive field can be obtained, and global content information can be captured better.
In the training process of the dense multipath convolutional network, as in DenseNet, the cross entropy of the model predicted value and the true value is used as a loss function. And simultaneously, replacing the ReLU used by the original DenseNet with an activation function ELU with better performance.
A dense multipath convolutional network provides a new basic module Block _ XT and a new dense module DenseBlock _ XT based on the idea of ResNeXt basic module multipath transformation.
The ResNeXt basic module has three steps of dividing, transforming and aggregating, and in order to better adapt to a DenseNet framework, the Block _ XT basic module newly proposed by the invention is modified as follows: first, the superposition aggregation step in the ResNeXt basic module is eliminated, and the aggregation purpose is realized by the series connection step in the DenseNet structure, so that the last convolution layer of 1x1 in the ResNeXt basic module is eliminated. Secondly, a Group Convolution (Group Convolution) operation in the deep learning Caffe framework can limit the connection of each Convolution kernel in a subset of the input feature map, input and output channels are automatically divided into a plurality of groups, and the output of the ith Group is only related to the input channel of the ith Group, namely the Group Convolution operation can automatically complete the division step. To simplify the model structure and reduce the parameters of the model, the Block _ XT basic Block replaces the partitioning operation in the resenext basic Block with the above-described set of convolution operations.
And designing a new dense module DenseBlock _ XT based on the newly proposed basic unit Block _ XT. And replacing the old dense module DenseBlock in the framework with a new dense module DenseBlock _ XT by using an integral framework of DenseNet to obtain the dense multipath convolutional network.
And during training, optimizing by using a random gradient descent method by taking the cross entropy of the model prediction result and the true value as a loss function. And selecting proper optimization parameters in experiments according to the characteristics of the data set. In the training process, the training data can also adopt a multi-scale data amplification method so as to improve the robustness of the model and reduce the degree of overfitting.
In the neural network training stage, image data is propagated backwards layer by layer to obtain a prediction result, the prediction result is compared with a label, loss is calculated, gradient backward propagation is calculated layer by layer according to the loss, the network weight is updated, and the process is repeated until model training is finished. The testing stage is that the image data is propagated backward from the input layer to the output layer until the prediction result is output from the output layer. The specific process of processing the picture data is to follow the DenseXT structure in Table 1
The following experimental results show that: compared with ResNeXt and DenseNet, the dense multipath convolution network provided by the invention has a remarkable improvement on the classification effect.
The experimental results are as follows:
to verify the effectiveness of the method of the present invention, we performed experimental verification of dense multipath convolutional networks on the now popular CIFAR-10 dataset.
The CIFAR-10 dataset is a subset of 8000 ten thousand tiny image datasets collected by Alex Krizhevsky, Vinod Nair and Geoffrey Hinton, comprising a total of 60000 color pictures in 10 categories, 6000 pictures in each category. Of these, 50000 were used for training and 10000 were used for testing. The data set was divided into a total of 5 training subsets (training Batch) and one Test subset (Test Batch), 10000 pieces per Batch. The test batch consisted of 1000 pictures randomly selected from each class for a total of 10000. The training batch consists of the remaining pictures, the number of each type of picture in each batch is not equal, and each type of picture in all data is just 5000 images.
Top-1 accuracy was used as an evaluation criterion. The Top-1 accuracy rate means that 1 label predicted by the algorithm is just the real label of the image, and the classification is correct, so that the error rate is calculated on the whole data set.
The experimental data set only adopts a Gaussian normalization method of subtracting a mean value and dividing the mean value by a standard deviation, no data amplification is carried out, the DenseNet result is directly run out by the author of the invention by adopting a code of a DenseNet presenter, and results of other comparison methods are all cited from corresponding papers.
The experimental results are shown in table 2, where Baseline is DenseNet, K represents the growth rate, and C represents Cardinality (i.e., Cardinality).
TABLE 2 results of classification experiments of DenseXT on CIFAR-10
As can be seen from Table 2, when the number of parameters is 26% of DenseNet, the DenseXT classification error rate is 0.54% higher than that of DenseNet; when the number of parameters is 45.4% of DenseNet, the DenseXT classification error rate is equivalent to DenseNet and is slightly lower than 0.16%; when the number of parameters is 69.6% of DenseNet, the DenseXT classification error rate is already lower than 0.75% of DenseNet. The same architecture of the DenseXT and the DenseNet network, only the basic module is different, which shows that the designed DenseXT network basic module has higher representation capability than that of the DenseNet network.
In terms of identification accuracy, DenseXT can exceed ResNet of 110 layers, ResNet of random depth of 110 layers, and Pre-activation ResNet of 1001 layers already at 22 layers, and the results illustrate the superiority of the DenseXT structure designed herein.
Experimental results show that the DenseXT network designed by the method can obtain higher precision than DenseNet by using fewer parameters and smaller models, and shows that DenseXT has stronger expression capability and better classification performance than DenseNet.
The following is a system example corresponding to the above method example, and the present implementation system can be implemented in cooperation with the above embodiments. The related technical details mentioned in the above embodiments are still valid in the present implementation system, and are not described herein again for the sake of reducing repetition. Accordingly, the related-art details mentioned in the present embodiment system can also be applied to the above-described embodiments.
The invention also provides a picture classification system based on the dense multipath convolutional network, which comprises the following steps:
the method comprises the following steps that 1, a basic unit formed by grouping, transformation and aggregation is constructed, a new dense module is constructed based on the basic unit, and an original dense module in a densely connected convolution network frame is replaced by the new dense module to obtain a dense multipath convolution network;
the module 2 uses the image data with the marked categories as training data, updates the weight in the dense multipath convolution network through gradient back propagation, trains the dense multipath convolution network and obtains an image classification model;
and the module 3 inputs the image data to be classified into the classification model to obtain a classification result of the image data to be classified.
The image classification system based on the dense multipath convolutional network is characterized in that the aggregation type is a serial aggregation structure.
The picture classification system based on the dense multipath convolutional network, wherein the module 1 comprises: and adding a normalization module, a scale adjustment layer and an activation function layer in front of the basic module to obtain the new dense module.
The picture classification system based on the dense multipath convolutional network comprises the following steps of: training data is input into the dense multipath convolutional network to obtain a model predicted value, loss is obtained by comparing the model predicted value with the marked category, gradient is calculated, and weight in the dense multipath convolutional network is updated through layer-by-layer reverse propagation.
The picture classification system based on the dense multipath convolutional network is characterized in that the cross entropy of the model prediction value and the marked category is used as the loss function.
Claims (10)
1. A picture classification method based on a dense multipath convolutional network is characterized by comprising the following steps:
step 1, constructing a basic unit formed by grouping, transforming and aggregating, constructing a new dense module based on the basic unit, and replacing an original dense module in a densely connected convolution network frame with the new dense module to obtain a dense multipath convolution network;
step 2, using the image data with the marked categories as training data, updating the weight in the dense multipath convolution network through gradient back propagation, and training the dense multipath convolution network to obtain an image classification model;
and 3, inputting the picture data to be classified into the classification model to obtain a classification result of the picture data to be classified.
2. The method of claim 1, wherein the aggregation type is a tandem aggregation structure.
3. The method of claim 1, wherein the step 1 comprises: and adding a normalization module, a scale adjustment layer and an activation function layer in front of the basic module to obtain the new dense module.
4. The method of claim 1, wherein the training of the dense multipath convolutional network comprises: training data is input into the dense multipath convolutional network to obtain a model predicted value, loss is obtained by comparing the model predicted value with the marked category, gradient is calculated, and weight in the dense multipath convolutional network is updated through layer-by-layer reverse propagation.
5. The method of claim 1 or 4, wherein cross entropy between the model prediction value and the labeled class is used as the loss function.
6. A picture classification system based on a dense multipath convolutional network, comprising:
the method comprises the following steps that 1, a basic unit formed by grouping, transformation and aggregation is constructed, a new dense module is constructed based on the basic unit, and an original dense module in a densely connected convolution network frame is replaced by the new dense module to obtain a dense multipath convolution network;
the module 2 uses the image data with the marked categories as training data, updates the weight in the dense multipath convolution network through gradient back propagation, trains the dense multipath convolution network and obtains an image classification model;
and the module 3 inputs the image data to be classified into the classification model to obtain a classification result of the image data to be classified.
7. The system of claim 6, wherein the aggregation type is a concatenated aggregation structure.
8. The system of claim 6, wherein the module 1 comprises: and adding a normalization module, a scale adjustment layer and an activation function layer in front of the basic module to obtain the new dense module.
9. The system of claim 6, wherein the training of the dense multipath convolutional network is specifically: training data is input into the dense multipath convolutional network to obtain a model predicted value, loss is obtained by comparing the model predicted value with the marked category, gradient is calculated, and weight in the dense multipath convolutional network is updated through layer-by-layer reverse propagation.
10. The system according to claim 6 or 1, wherein the cross entropy between the model prediction value and the labeled class is used as the loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010531567.7A CN111832621A (en) | 2020-06-11 | 2020-06-11 | Image classification method and system based on dense multipath convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010531567.7A CN111832621A (en) | 2020-06-11 | 2020-06-11 | Image classification method and system based on dense multipath convolutional network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111832621A true CN111832621A (en) | 2020-10-27 |
Family
ID=72897794
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010531567.7A Pending CN111832621A (en) | 2020-06-11 | 2020-06-11 | Image classification method and system based on dense multipath convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111832621A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109584248A (en) * | 2018-11-20 | 2019-04-05 | 西安电子科技大学 | Infrared surface object instance dividing method based on Fusion Features and dense connection network |
| CN110189308A (en) * | 2019-05-17 | 2019-08-30 | 山东财经大学 | A kind of lesion detection approach and device based on BM3D and the fusion of dense convolutional network |
| CN111209972A (en) * | 2020-01-09 | 2020-05-29 | 中国科学院计算技术研究所 | Image classification method and system based on hybrid connectivity deep convolution neural network |
| CN111242862A (en) * | 2020-01-09 | 2020-06-05 | 西安理工大学 | Multi-scale fusion parallel dense residual convolution neural network image denoising method |
-
2020
- 2020-06-11 CN CN202010531567.7A patent/CN111832621A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109584248A (en) * | 2018-11-20 | 2019-04-05 | 西安电子科技大学 | Infrared surface object instance dividing method based on Fusion Features and dense connection network |
| CN110189308A (en) * | 2019-05-17 | 2019-08-30 | 山东财经大学 | A kind of lesion detection approach and device based on BM3D and the fusion of dense convolutional network |
| CN111209972A (en) * | 2020-01-09 | 2020-05-29 | 中国科学院计算技术研究所 | Image classification method and system based on hybrid connectivity deep convolution neural network |
| CN111242862A (en) * | 2020-01-09 | 2020-06-05 | 西安理工大学 | Multi-scale fusion parallel dense residual convolution neural network image denoising method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110929603B (en) | Weather image recognition method based on lightweight convolutional neural network | |
| CN108764317B (en) | Residual convolutional neural network image classification method based on multipath feature weighting | |
| CN111242288B (en) | Multi-scale parallel deep neural network model construction method for lesion image segmentation | |
| CN108614997B (en) | Remote sensing image identification method based on improved AlexNet | |
| CN107480726A (en) | A kind of Scene Semantics dividing method based on full convolution and shot and long term mnemon | |
| CN110175628A (en) | A kind of compression algorithm based on automatic search with the neural networks pruning of knowledge distillation | |
| CN109948029A (en) | Based on the adaptive depth hashing image searching method of neural network | |
| CN109740679B (en) | Target identification method based on convolutional neural network and naive Bayes | |
| CN112070768B (en) | Anchor-Free based real-time instance segmentation method | |
| CN110135460B (en) | Image information enhancement method based on VLAD convolution module | |
| CN107564007B (en) | Scene segmentation correction method and system fusing global information | |
| Zhang et al. | Channel-wise and feature-points reweights densenet for image classification | |
| CN111695513A (en) | Facial expression recognition method based on depth residual error network | |
| CN105989375A (en) | Classifier, classification device and classification method for classifying handwritten character images | |
| CN115035418A (en) | Remote sensing image semantic segmentation method and system based on improved deep LabV3+ network | |
| CN114821058A (en) | Image semantic segmentation method and device, electronic equipment and storage medium | |
| CN114219824A (en) | Visible light-infrared target tracking method and system based on deep network | |
| CN110598746A (en) | Adaptive scene classification method based on ODE solver | |
| Du et al. | Efficient network construction through structural plasticity | |
| CN110782001A (en) | Improved method for using shared convolution kernel based on group convolution neural network | |
| CN111144497B (en) | Image significance prediction method under multitasking depth network based on aesthetic analysis | |
| CN111832621A (en) | Image classification method and system based on dense multipath convolutional network | |
| da Silva Oliveira et al. | Comparing transfer learning approaches applied to distracted driver detection | |
| CN110992320A (en) | Medical image segmentation network based on double interleaving | |
| CN113947706B (en) | Image diversity enhancement method and system based on generation countermeasure 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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20201027 |