CN104866900B - A kind of deconvolution neural network training method - Google Patents
A kind of deconvolution neural network training method Download PDFInfo
- Publication number
- CN104866900B CN104866900B CN201510046974.8A CN201510046974A CN104866900B CN 104866900 B CN104866900 B CN 104866900B CN 201510046974 A CN201510046974 A CN 201510046974A CN 104866900 B CN104866900 B CN 104866900B
- Authority
- CN
- China
- Prior art keywords
- training
- layer
- filter
- mrow
- images
- 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.)
- Active
Links
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
本发明公开了一种反卷积神经网络训练方法，其能够有效地提取图像特征，有益于分类正确率的提高，提高反卷积神经网络的训练收敛效率及收敛精度，降低反卷积神经网络在实际应用中的训练成本，同时可被应用于其他基于卷积运算的优化问题求解中。这种反卷积神经网络训练方法，包括训练阶段和重建阶段，训练阶段包括步骤：(1)对训练图像进行预处理；(2)对训练图像进行批设置；(3)设置训练图像的网络训练参数；(4)开始第一层训练；重建阶段包括步骤：(5)对待重构图像进行预处理；(6)设置待重构图像的网络训练参数；(7)按批输入待重构图像直到完成所有批图像的重构。
The invention discloses a deconvolution neural network training method, which can effectively extract image features, is beneficial to the improvement of the classification accuracy rate, improves the training convergence efficiency and convergence accuracy of the deconvolution neural network, and reduces the deconvolution neural network. The training cost in practical applications can also be applied to solve other optimization problems based on convolution operations. This deconvolutional neural network training method includes a training phase and a reconstruction phase, and the training phase includes steps: (1) preprocessing the training images; (2) batch setting the training images; (3) setting the network of the training images Training parameters; (4) start the first layer of training; the reconstruction stage includes steps: (5) preprocessing the image to be reconstructed; (6) setting the network training parameters of the image to be reconstructed; (7) inputting the image to be reconstructed by batch images until the reconstruction of all batch images is completed.
Description
Technical Field
The invention belongs to the technical field of artificial intelligence, and particularly relates to a deconvolution neural network training method.
Background
Deep learning is one of the research hotspots in the field of artificial intelligence at present. Based on the thought of neural network, deep learning aims to simulate the layered Perception mode of human brain, and an abstract high-level representation capable of representing attribute categories or characteristics is combined from low-level characteristics by constructing a Multilayer perceptron (MLP), so that the model becomes one of effective models for extracting complex image characteristics at present.
The classical Deep learning model mainly includes Deep Belief Networks (DBNs), multi-layer sparse automatic coding models (AE), and Convolutional Neural Networks (CNNs). In general, these models extract features from an input image through an encoder, greedily transform the image from bottom to top to a feature space layer by layer, and in contrast, reconstruct features of the feature space from top to bottom through a network using a decoder. Among them, DBNs and AE are unsupervised network models that can hierarchically learn rich image features from bottom to top and provide certain gains for high-level applications, such as pattern recognition. However, since the encoder usually uses some simple non-linear functions, it may severely limit the extraction of potential features, so that the model learns suboptimal features, thereby restricting the improvement of the practical accuracy of the application in the recognition field. In addition, their training complexity and time overhead are enormous.
The CNNs model is a supervised learning model and is known as the first algorithm for really and effectively training the multilayer network model. The model constructs a multi-layer network using a trainable set of convolution filters and corresponding pooled downsampling algorithms. Like DBNs and AE, the model trains the network from bottom to top, and can only complete the function of feature extraction and related high-level applications, such as pattern recognition, and cannot complete the function of image reconstruction and corresponding low-level applications.
Zeiler et al first proposed a deconvolution neural network model (DN) in 2010. The model is an unsupervised deep learning model. The model expands the convolutional sparse coding to multiple layers and uses a network structure similar to CNNs, so that the model can not only cope with high-level application for extracting features, but also realize low-level application for image reconstruction, and the defect that the translation invariance of a natural image is ignored in the traditional sparse coding algorithm is overcome. In particular, the convolutional sparse coding model can be considered as a special case of a single-layer deconvolution neural network model. Through the learning set representation of the multilayer structure, the model can effectively extract the features of the images in different scales, and obtains a plurality of satisfactory effects in high-level application, such as identification and classification. According to the Chenyangti method, the DN image denoising method has excellent performance in the application of DN image denoising, and the potential of the DN image denoising method in low-level application is shown.
However, solving the DN model involves a large number of convolution operations and a series of optimization problems. Zeiler et al decomposes the entire optimization problem into two sub-problems: the fixed eigen map solves the filter sub-problem of the filter and the fixed filter solves the eigen map sub-problem of the eigen map. And when actually solving, carrying out alternate optimization solution on the two sub-problems until the whole model converges. Both the two subproblems are solved by using a Conjugate Gradient (CG) method, and the computational complexity is high and the computation is slow. The Zeiler introduces the FISTA algorithm to solve the subproblems of the characteristic diagram, and adds the pooling downsampling algorithm between layers to improve the overall operation speed of the model, but the solution of the objective function is still high in complexity and slow in convergence due to the large use of convolution and a multi-layer complex structure. Besides the optimization training algorithm, the mainstream DN optimization method also includes using GPU operation and using a multithread programming method.
Because the convolution sparse coding model is a special case of a single-layer deconvolution neural network model, some optimization solving algorithms based on the convolution sparse coding model can be regarded as special cases of optimization solving of the single-layer deconvolution neural network model. Bristow et al improve the convergence efficiency of the model by transforming the convolution operation into the fourier domain and using fast fourier operations to increase the speed of the convolution operation. Rigamonti et al propose convolutional sparse coding models with separable filter penalties that improve the speed of convolutional operations by training separable filters to reduce the complexity of the convolutional operations. However, extending such methods to a multi-layer deconvolution neural network model still requires much effort and improvement.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the training method of the deconvolution neural network can effectively extract image features, is beneficial to improving the classification accuracy, improves the training convergence efficiency and the convergence accuracy of the deconvolution neural network, reduces the training cost of the deconvolution neural network in practical application, and can be applied to solving other optimization problems based on convolution operation.
The technical solution of the invention is as follows: the training method of the deconvolution neural network comprises a training stage and a reconstruction stage, wherein the training stage comprises the following steps:
(1) preprocessing a training image: selecting a training set image, processing the training set image into a gray image, and unifying length and width pixels;
(2) carrying out batch setting on training images: according to the application of the trained network, carrying out batch division on the training images;
(3) setting network training parameters of a training image, wherein the network training parameters of the training image comprise the number of network layers, the size of each layer of filter, the number of characteristic graphs of each layer, FISTA reconstruction step number and reconstruction step length, total cycle epoch times and characteristic graph sparse control parameters;
(4) starting a first layer of training: initializing a first-layer characteristic graph and a first-layer filter by using random numbers, and inputting a batch of training images; solving the sub-problem of the feature map, solving the feature map by using a FISTA method, then solving the sub-problem of the filter, wherein the current residual error is only used for updating one filter every time of calculation, the gradient step length for updating the filter comes from the current recalculated reconstructed residual error every time, and repeatedly repeating the two steps of solving the sub-problem of the feature map and solving the sub-problem of the filter until the reconstructed residual error is converged; inputting a second batch of training images, initializing feature maps of the second batch of images by using random numbers and using filters obtained by the previous batch of training, repeating the steps until training of all batches of training images is completed to be used as an epoch, and then inputting the first batch of images to start a second epoch training until the total cycle epoch times of the step (3) are completed;
the reconstruction phase comprises the following steps:
(5) preprocessing an image to be reconstructed: processing an image to be reconstructed into a gray image, and unifying the length and width pixels;
(6) setting network training parameters of an image to be reconstructed, wherein the network training parameters of the image to be reconstructed comprise a FISTA reconstruction step length, reconstruction times and characteristic diagram sparsity control parameters;
(7) inputting images to be reconstructed in batches until the reconstruction of all the batches of images is completed: and sequentially inputting images to be reconstructed according to batches, and calculating the characteristic diagram subproblems of each layer of filter obtained by training the training images before the network is used until the convergence to obtain the reconstructed images and each layer of characteristic diagram.
According to the method, by improving the sub-problem of the filter solution, the current residual error is only used for updating one filter every time of calculation, and the gradient step length used for updating the filter every time is from the reconstructed residual error calculated again at present, so that the image characteristics can be effectively extracted, the improvement of the classification accuracy is facilitated, the training convergence efficiency and the convergence accuracy of the deconvolution neural network are improved, the training cost of the deconvolution neural network in practical application is reduced, and the method can be applied to the solution of other optimization problems based on convolution operation.
Drawings
Fig. 1 shows a logical structure of a deconvolution neural network model.
FIG. 2 is a comparison of objective data of SFU algorithm and traditional CG algorithm in training experiment; in fig. 2, (a) (b) shows the training results of the first layer, and (c) (d) shows the training results of the second layer.
FIG. 3 is a comparison of objective data of SFU algorithm and traditional CG algorithm in reconstruction experiments; in fig. 3, (a) (b) shows the training results of the first layer, and (c) (d) shows the training results of the second layer.
FIG. 4 is a flow chart of a method of deconvolution neural network training in accordance with the present invention.
Detailed Description
As shown in fig. 4, the training method of the deconvolution neural network includes a training phase and a reconstruction phase, and the training phase includes the following steps:
(1) preprocessing a training image: selecting a training set image, processing the training set image into a gray image, and unifying length and width pixels;
(2) carrying out batch setting on training images: according to the application of the trained network, carrying out batch division on the training images;
(3) setting network training parameters of a training image, wherein the network training parameters of the training image comprise the number of network layers, the size of each layer of filter, the number of characteristic graphs of each layer, FISTA reconstruction step number and reconstruction step length, total cycle epoch times and characteristic graph sparse control parameters;
(4) starting a first layer of training: initializing a first-layer characteristic graph and a first-layer filter by using random numbers, and inputting a batch of training images; solving the sub-problem of the feature map, solving the feature map by using a FISTA method, then solving the sub-problem of the filter, wherein the current residual error is only used for updating one filter every time of calculation, the gradient step length for updating the filter comes from the current recalculated reconstructed residual error every time, and repeatedly repeating the two steps of solving the sub-problem of the feature map and solving the sub-problem of the filter until the reconstructed residual error is converged; inputting a second batch of training images, initializing feature maps of the second batch of images by using random numbers and using filters obtained by the previous batch of training, repeating the steps until training of all batches of training images is completed to be used as an epoch, and then inputting the first batch of images to start a second epoch training until the total cycle epoch times of the step (3) are completed;
the reconstruction phase comprises the following steps:
(5) preprocessing an image to be reconstructed: processing an image to be reconstructed into a gray image, and unifying the length and width pixels;
(6) setting network training parameters of an image to be reconstructed, wherein the network training parameters of the image to be reconstructed comprise a FISTA reconstruction step length, reconstruction times and characteristic diagram sparsity control parameters;
(7) inputting images to be reconstructed in batches until the reconstruction of all the batches of images is completed: and sequentially inputting images to be reconstructed according to batches, and calculating the characteristic diagram subproblems of each layer of filter obtained by training the training images before the network is used until the convergence to obtain the reconstructed images and each layer of characteristic diagram.
According to the method, by improving the sub-problem of the filter solution, the current residual error is only used for updating one filter every time of calculation, and the gradient step length used for updating the filter every time is from the reconstructed residual error calculated again at present, so that the image characteristics can be effectively extracted, the improvement of the classification accuracy is facilitated, the training convergence efficiency and the convergence accuracy of the deconvolution neural network are improved, the training cost of the deconvolution neural network in practical application is reduced, and the method can be applied to the solution of other optimization problems based on convolution operation.
Preferably, in the step (2), the images of the same type with high similarity are used as the same batch of training.
Preferably, the filter subproblem is solved in the step (4) by the formula (1):
wherein u ═ fk，Wherein u ═ fkIs the filter that is currently to be updated, is the value obtained in the last iteration of the current filter to be updated, and L > 0 isThe upper constant parameter of the Labrinitz gradient, g (f)k) Is the total cost function expressed by taking the current filter to be updated as a variable.
Preferably, the following steps are further included between the steps (4) and (5):
(a) starting a second layer of training: initializing a second layer filter and a feature map by using a random number by using a first layer filter obtained by first layer training; inputting the first batch of training images into a network, and performing optimization calculation on the filter and the characteristic diagram until convergence; inputting a next batch of training images, initializing the next batch of feature maps by using a random number, training until convergence, and repeating the steps until the training of all batches of training images is completed;
(b) starting nth layer training, n is an integer greater than 2: using the filter obtained by the training of the first n-1 layers, and initializing the filter of the nth layer and the characteristic diagram by using a random number; inputting the first batch of training images into a network, and performing optimization calculation on the filter and the characteristic diagram until convergence; inputting a next batch of training images, initializing the next batch of feature maps by using random numbers, training until convergence, and repeating the steps until all batches of training images are trained.
Preferably, in the steps (a) and (b), the filter and the feature map are optimally calculated by formula (2):
wherein,is the filter that is currently to be updated at layer l,is the value of the filter obtained in the last iteration, LlIs the upper bound constant parameter, g, of the Labrunitz gradient of the first layerlIs the first layerIs a univariate total cost function.
The present invention is specifically illustrated below:
in order to improve the training convergence efficiency and the convergence precision of the deconvolution neural network and reduce the training cost of the deconvolution neural network in practical application, the invention provides a deconvolution neural network training algorithm based on a sequential update filter strategy. The algorithm can be applied to solving other optimization problems based on convolution operation, such as a convolution sparse coding model.
The classical deconvolution neural network model objective function can be recursively expressed as:
where, y is the input image,is a reconstructed image restored through a network using the feature map of the l-th layer,is the ith characteristic diagram of the l layer, Klis the number of characteristic graphs owned by the l-th layer, β is a reconstruction weight parameter, | | · |. luminance1Is thatThe norm of the number of the first-order-of-arrival,is the l-th layer feature map set, using the single vector z hereinafterlIs represented by RlIs a set of the l-th layer feature mapsReconstruction operator to restore to a reconstructed image, FlIs an operation matrix containing the l-th layer convolution and summation operation, which is defined as follows:
whereinIs a filter of the first layer of the filter,is a feature map of the first layer, representing a two-dimensional convolution operator,is the filter with the l layer number (i, j), which is used to reconstruct the feature map of the previous layerOne of the filters of (1). For the trained deconvolution neural network, all the filters in the network are common to all the input images, and each input image has a unique group of feature maps from the highest layer after being extracted by the network.
the logical structure of the deconvolution neural network model is shown in fig. 1, where ⊕ represents the two-dimensional convolution operator.
Traditional deconvolution neural network training methods are typically based on gradient descent algorithms. Both of these methods use the same reconstructed residual to calculate the optimal step size for all filters and update all filters in a batch, thereby introducing more error.
The main work of the invention is based on the above model, and the traditional batch updating filter algorithm is improved to greedy update filters one by one sequentially, that is, the current residual error is only used for updating one filter every time of calculation, and the gradient step length used for updating the filter every time is from the reconstructed residual error which is recalculated currently. In order to make the updating strategy calculable, the invention applies Taylor expansion formula to separate the filters to be updated one by one, thereby implementing the updating calculation.
1. Method for solving characteristic diagram subproblems
The solution of the deconvolution neural network model comprises an alternating solution optimization problem of two sub-problems, a filter sub-problem and a feature map sub-problem. The present invention improves only the solution method of the filter subproblem, while the solution of the eigen map subproblem continues to use the classical solution scheme, i.e. the FISTA method. To facilitate an understanding of the computational method of the present invention and to enable a reproduction of the experiments of the present invention, the application of the FISTA method to the computational steps for solving the signature subproblems is briefly described herein.
The feature map sub-problem is intended to fix the current filter set and optimally solve the feature map. When the problem is solved layer by layer, the problem is degraded into an exemplar convolution sparse coding problem. First, a reconstructed image is calculated using the above model objective functionThe feature map set is then calculated using the following equation:
where, y is the input image,is an operator for passing the residual from the first layer to the l-th layer,lis the l-th layer residual, αlIs the ISTA gradient step.
2. Solving method of filter subproblem
The content of the filter subproblem is a fixed characteristic diagram, and a filter set is optimized and solved. For convenience of explanation and understanding, the invention firstly describes a training solution method of a single-layer deconvolution neural network, and then extends to a multi-layer network solution method.
1.1 Single layer deconvolution neural network model solution
As shown by the objective function above, willIs decomposed intoThe single-layer objective function can thus be reduced to the following form:
whereinAnd isAnd then can be rewritten into the following form:
filter set for first layerIn particular, contain K1-set of 1 elementIs fixed and known, only the k-th filterIs an unknown quantity. When the feature map is fixed, only the update is performedGamma is a combination of1Is a constant, a functionIs a continuous convex function and thus can be approximated by taylor expansions and labbuchtz residuals. A function g is arranged inThe points are subjected to Taylor expansion, which yields the following equation:
the value of the expansion approximates the function g in each iterationTrue value of the point, L > 0The upper bound constant parameter of the Lei-Bluenz gradient. This formula is simplified to give the following formula:
using u-fk，Replacing the argument, rewrite the formula as:
where v is a known quantity and u can be calculated directly from v.
1.2 multilayer deconvolution neural network model solution
According to the multi-layer deconvolution neural network model objective function shown above, the objective function to solve the l-th layer network can be expressed as:
wherein, KlIs the number of characteristic maps of the l-th layer,is the jth characteristic diagram of the ith layer,is the l-th layer, indexed (i, j) filter, which is used to reconstruct the feature map of the previous layerone element of the filter set, βlIs a regularization parameter.
Expanding the solution of the first layer to the l layer, and sequentially filtering the l layer of filterSequential updates are separated from the overall objective function, whereby the above equation can be rewritten as:
wherein When the characteristic diagram is fixedUpdatingWhen Z islIs a constant. The formula is iterated back to the multilayer objective function by recursion to obtain a multilayer total objective function:
wherein Method of analogy to the first layer, for the equationTaylor expansion is carried out, and simplification can obtain:
whereinEquation (2) can be solved for multiple iterative optimizations using the same method as for the first layer.
In order to verify the effectiveness of the algorithm provided by the text, the text respectively performs effect verification through two types of experiments. Experiment one is implemented by training and reconstruction. The experiment used a face94 face dataset based on which a two-layer deconvolution neural network model DN, hereinafter designated DN-CG and DN-SFU, was trained using the conventional conjugate gradient CG and the text SFU, respectively. The accuracy and convergence performance of the two algorithms in the training stage of the first layer and the second layer network and the reconstruction stage are respectively compared through the reconstruction Mean Square Error (MSE) and the reconstruction peak signal-to-noise ratio (PSNR). The structure of the deconvolution neural network is set to be 9 feature maps in the first layer with a filter size of 11 × 11, and 4 feature maps in the second layer with a filter size of 21 × 21. The training image is preprocessed into a grayscale 200 x 200 image, 10 images are arbitrarily taken as a training set, and the other 10 different images are arbitrarily taken as a test set.
For clarity, the following notations are used in Table 1.
TABLE 1 Experimental notation
| Symbol | Model (model) | Algorithm |
| DN-SFU | DN | SFU |
| DN-CG | DN | CG |
Wherein DN-SFU is used for training the deconvolution neural network by using the text algorithm, and DN-CG is used for training the deconvolution neural network by using the traditional conjugate gradient method.
In fig. 2, (a) (b) shows the training results of the first layer, and (c) (d) shows the training results of the second layer. The ordinate of (a) and (c) is mean square error MSE of the reconstructed image, the abscissa is the total iteration times, and the change condition of the reconstructed mean square error along with the iteration times in the training process is shown. (b) And (d) the ordinate is the mean square error MSE of the reconstructed image, the abscissa is the training time, and the condition that the reconstructed mean square error changes along with the time in the training process is shown.
The objective experimental data prove that when the DN-SFU algorithm trains the deconvolution neural network, the convergence speed of the first layer network is obviously improved, and meanwhile, the convergence precision is improved. The DN-SFU algorithm does not obviously improve the training convergence speed of the second layer, but improves the convergence precision. Because the training of the second-layer network starts from the convergence point of the first-layer network, the algorithm can still obtain larger training time gain when the two-layer network is comprehensively trained.
The results of fig. 3 are the reconstruction experiments of the test images using the filter sets trained using the DN-CG and DN-SFU algorithms, respectively, where (a) (b) is the first layer training result and (c) (d) is the second layer training result. The ordinate of (a) and (c) is the peak signal-to-noise ratio (PSNR) of the reconstructed image and the original image, the abscissa is the iteration frequency of the reconstructed ISTA, and the change condition of the reconstructed PSNR along with the iteration frequency in the reconstruction process is shown. (b) And (d) the ordinate is the PSNR of the reconstructed image, the abscissa is the reconstruction time, and the condition that the reconstructed PSNR changes along with the time in the reconstruction process is shown. The reconstruction experiment adopts test set data which is not coincident with a training set and a filter set of convergence points of each algorithm during training.
The experimental result proves that the filter set obtained by training the algorithm has larger quality gain in the aspect of the reconstruction performance of the test image.
The test images reconstructed using the filter set trained using the DN-SFU algorithm herein are subjectively much better than the traditional conjugate gradient algorithm.
The second experiment is based on a data set Caltech-101, the data set is composed of natural images, is divided into 101 classes, and is one of general data sets used for evaluating algorithm efficiency in the field of pattern recognition. The general identification experiment test parameters are used, namely 30 images of each class are taken as training images, 3060 images are obtained in the training set, and the rest images are used as test sets. All images of the data set are preprocessed into gray scale images, and are scaled to 150 × 150 pixels in a manner of zero padding. The deconvolution network model set in this experiment was a first layer of 15 profiles using filters of 7 × 7 size, and a second layer of 4 profiles using filters of 21 × 21 size. After the training of the network is completed, the test image is used for extracting a characteristic diagram through the network, and the first layer is adoptedThe feature map is used as an input parameter of the classifier. The classifier adopts a general SPM[13]The classification method comprises the steps of respectively extracting feature vectors of all first-layer feature maps of one map through SPM, adding the vectors, averaging to obtain a final unique feature vector of the map, and using the final unique feature vector as the final classification input of the support vector machine. This experimental method and reference [7 ]]The described method is consistent. The recognition results are shown in table 2.
TABLE 2
Table 2 demonstrates that the filter set obtained by training using the method of the present disclosure can effectively extract image features and is beneficial to improving the classification accuracy. For fairness, the methods listed herein all use a single feature for recognition, and the multi-feature comprehensive recognition method has more excellent performance on the data set.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and all simple modifications, equivalent variations and modifications made to the above embodiment according to the technical spirit of the present invention still belong to the protection scope of the technical solution of the present invention.
Claims (2)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510046974.8A CN104866900B (en) | 2015-01-29 | 2015-01-29 | A kind of deconvolution neural network training method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510046974.8A CN104866900B (en) | 2015-01-29 | 2015-01-29 | A kind of deconvolution neural network training method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104866900A CN104866900A (en) | 2015-08-26 |
| CN104866900B true CN104866900B (en) | 2018-01-19 |
Family
ID=53912720
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510046974.8A Active CN104866900B (en) | 2015-01-29 | 2015-01-29 | A kind of deconvolution neural network training method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104866900B (en) |
Families Citing this family (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11074492B2 (en) | 2015-10-07 | 2021-07-27 | Altera Corporation | Method and apparatus for performing different types of convolution operations with the same processing elements |
| CN105678249B (en) * | 2015-12-31 | 2019-05-07 | 上海科技大学 | Face Recognition Methods for Different Image Quality of Registered Face and Face to Be Recognized |
| CN106326346A (en) * | 2016-08-06 | 2017-01-11 | 上海高欣计算机系统有限公司 | Text classification method and terminal device |
| WO2018076130A1 (en) * | 2016-10-24 | 2018-05-03 | 中国科学院自动化研究所 | Method for establishing object recognition model, and object recognition method |
| CN108073876B (en) * | 2016-11-14 | 2023-09-19 | 北京三星通信技术研究有限公司 | Facial analysis device and facial analysis method |
| EP3330898B1 (en) * | 2016-12-01 | 2025-04-23 | Altera Corporation | Method and apparatus for performing different types of convolution operations with the same processing elements |
| CN106682730B (en) * | 2017-01-10 | 2019-01-08 | 西安电子科技大学 | network performance evaluation method based on VGG16 image deconvolution |
| WO2018145308A1 (en) * | 2017-02-13 | 2018-08-16 | Nokia Technologies Oy | Filter reusing mechanism for constructing robust deep convolutional neural network |
| US11003989B2 (en) * | 2017-04-27 | 2021-05-11 | Futurewei Technologies, Inc. | Non-convex optimization by gradient-accelerated simulated annealing |
| US10657446B2 (en) | 2017-06-02 | 2020-05-19 | Mitsubishi Electric Research Laboratories, Inc. | Sparsity enforcing neural network |
| CN109033107B (en) * | 2017-06-09 | 2021-09-17 | 腾讯科技（深圳）有限公司 | Image retrieval method and apparatus, computer device, and storage medium |
| CN107545263B (en) * | 2017-08-02 | 2020-12-15 | 清华大学 | Object detection method and device |
| CN107993217B (en) * | 2017-12-22 | 2021-04-09 | 北京奇虎科技有限公司 | Video data real-time processing method and device, and computing device |
| CN108012191B (en) * | 2017-12-22 | 2020-08-14 | 北京奇虎科技有限公司 | Video data processing method and device, computing equipment and medium |
| CN108805036B (en) * | 2018-05-22 | 2022-11-22 | 电子科技大学 | Unsupervised video semantic extraction method |
| WO2019222936A1 (en) * | 2018-05-23 | 2019-11-28 | 富士通株式会社 | Method and device for training classification neural network for semantic segmentation, and electronic apparatus |
| CN108985317B (en) * | 2018-05-25 | 2022-03-01 | 西安电子科技大学 | Image classification method based on separable convolution and attention mechanism |
| CN108846385B (en) * | 2018-07-10 | 2021-11-09 | 浪潮通用软件有限公司 | Image identification and correction method and device based on convolution-deconvolution neural network |
| CN109272118B (en) * | 2018-08-10 | 2020-03-06 | 北京达佳互联信息技术有限公司 | Data training method, device, equipment and storage medium |
| CN109727195B (en) * | 2018-12-25 | 2023-06-13 | 四川康吉笙科技有限公司 | Image super-resolution reconstruction method |
| CN109784490B (en) | 2019-02-02 | 2020-07-03 | 北京地平线机器人技术研发有限公司 | Training method, device and electronic device for neural network |
| US11526736B2 (en) * | 2019-08-15 | 2022-12-13 | Intel Corporation | Methods, systems, articles of manufacture and apparatus to map workloads |
| CN110782396B (en) * | 2019-11-25 | 2023-03-28 | 武汉大学 | Light-weight image super-resolution reconstruction network and reconstruction method |
| CN113657352B (en) * | 2020-03-19 | 2025-03-04 | 蚂蚁区块链科技(上海)有限公司 | A method, device and equipment for extracting facial features |
| CN113033704B (en) * | 2021-04-22 | 2023-11-07 | 江西理工大学 | Intelligent judging method and system for copper converter converting copper-making final point based on pattern recognition |
| CN113505865B (en) * | 2021-09-10 | 2021-12-07 | 浙江双元科技股份有限公司 | Sheet surface defect image recognition processing method based on convolutional neural network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101853490A (en) * | 2010-04-21 | 2010-10-06 | 中国科学院半导体研究所 | A Bionic Image Restoration Method Based on Human Visual Characteristics |
| CN102546128A (en) * | 2012-02-23 | 2012-07-04 | 广东白云学院 | Method for multi-channel blind deconvolution on cascaded neural network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9799098B2 (en) * | 2007-04-24 | 2017-10-24 | Massachusetts Institute Of Technology | Method and apparatus for image processing |
| WO2010003041A2 (en) * | 2008-07-03 | 2010-01-07 | Nec Laboratories America, Inc. | Mitotic figure detector and counter system and method for detecting and counting mitotic figures |
-
2015
- 2015-01-29 CN CN201510046974.8A patent/CN104866900B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101853490A (en) * | 2010-04-21 | 2010-10-06 | 中国科学院半导体研究所 | A Bionic Image Restoration Method Based on Human Visual Characteristics |
| CN102546128A (en) * | 2012-02-23 | 2012-07-04 | 广东白云学院 | Method for multi-channel blind deconvolution on cascaded neural network |
Non-Patent Citations (4)
| Title |
|---|
| Adaptive deconvolutional networks for mid and high level feature learning;Zeiler M D etal.;《Computer Vision(ICCV),2011IEEE International Conference on》;20111231;第2018-2025页 * |
| Breast image feature learning with adaptive deconvolutional networks;Jamieson AR etal.;《SPIE Medical Imaging》;20121231;第1506-1518页 * |
| 反卷积网络图像表述与复原;陈扬钛;《中国优秀硕士学位论文全文数据库 信息科技辑》;20120415;第25-66页 * |
| 基于稀疏表示模型的图像解码方法;施云惠 等;《北京工业大学学报》;20130331;第39卷(第3期);第420-424页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104866900A (en) | 2015-08-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104866900B (en) | A kind of deconvolution neural network training method | |
| KR102796191B1 (en) | Method for optimizing neural networks | |
| Kothari et al. | Trumpets: Injective flows for inference and inverse problems | |
| CN113077388B (en) | Data-augmented deep semi-supervised over-limit learning image classification method and system | |
| Cui et al. | Deep network cascade for image super-resolution | |
| CN110648292B (en) | A high-noise image denoising method based on deep convolutional network | |
| Yang et al. | Legonet: Efficient convolutional neural networks with lego filters | |
| CN113255437A (en) | Fault diagnosis method for deep convolution sparse automatic encoder of rolling bearing | |
| CN109977989B (en) | A method of processing image tensor data | |
| CN112150354B (en) | A Single Image Super-Resolution Method Based on Joint Contour Enhancement and Denoising Statistical Priors | |
| CN113177882A (en) | Single-frame image super-resolution processing method based on diffusion model | |
| CN104050507B (en) | Hyperspectral image classification method based on multilayer neural network | |
| CN110428045A (en) | Depth convolutional neural networks compression method based on Tucker algorithm | |
| CN103400402A (en) | Low-rank structure-based sparse compressive sensing MRI (Magnetic Resonance Imaging) image reconstruction method | |
| CN110909801A (en) | Data classification method, system, medium and device based on convolutional neural network | |
| CN108985457B (en) | Deep neural network structure design method inspired by optimization algorithm | |
| CN109146061A (en) | The treating method and apparatus of neural network model | |
| CN112132760B (en) | Image restoration method based on learnable differentiable matrix inverse and matrix decomposition | |
| CN116228575B (en) | Infrared spectrum restoration method based on frequency domain enhanced Transformer network | |
| CN116168197B (en) | Image segmentation method based on Transformer segmentation network and regularization training | |
| CN116797456B (en) | Image super-resolution reconstruction method, system, device and storage medium | |
| CN112801104A (en) | Image pixel level pseudo label determination method and system based on semantic segmentation | |
| Lacey et al. | Stochastic layer-wise precision in deep neural networks | |
| CN117095217A (en) | Multi-stage comparative knowledge distillation method | |
| CN105184742A (en) | Image denoising method of sparse coding based on Laplace graph characteristic vector |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| OL01 | Intention to license declared | ||
| OL01 | Intention to license declared | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20150826 Assignee: Beijing Feiwang Technology Co.,Ltd. Assignor: Beijing University of Technology Contract record no.: X2024980041978 Denomination of invention: A training method for deconvolution neural network Granted publication date: 20180119 License type: Open License Record date: 20241226 |
|
| EE01 | Entry into force of recordation of patent licensing contract | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20150826 Assignee: BEIJING MOQIU TECHNOLOGY CO.,LTD. Assignor: Beijing University of Technology Contract record no.: X2025980023083 Denomination of invention: A training method for deconvolution neural network Granted publication date: 20180119 License type: Open License Record date: 20250917 |
|
| EE01 | Entry into force of recordation of patent licensing contract |