1979 results about "Autoencoder" patented technology

Embodiments of the present invention relate to systems and methods for improving the training of machine learning systems to recognize certain objects within a given image by supplementing an existing sparse set of real-world training images with a comparatively dense set of realistic training images. Embodiments may create such a dense set of realistic training images by training a machine learning translator with a convolutional autoencoder to translate a dense set of synthetic images of an object into more realistic training images. Embodiments may also create a dense set of realistic training images by training a generative adversarial network (“GAN”) to create realistic training images from a combination of the existing sparse set of real-world training images and either Gaussian noise, translated images, or synthetic images. The created dense set of realistic training images may then be used to more effectively train a machine learning object recognizer to recognize a target object in a newly presented digital image.

The invention provides an image inpainting method and system based on antagonistic generation neural network, which comprises the following steps: firstly, constructing a self-encoder convolutional neural network (including encoder and encoding discriminator), a decoder (generator) convolutional neural network, a discriminator convolutional neural network, a global discriminator, and a local discriminator; Then, different loss functions are constructed for the five networks, and the whole network is trained by step-by-step training. Finally, when the network training is completed, the defect image is put into the network for repairing, and the result graph generated by the decoder (generator) is the final repairing result graph. The invention has the advantages that: the potential constraint of the image is kept and the image is thinned; an end-to-end image restoration network is implemented. The dependence of the repair network on the missing position mask information of the image iseliminated. The robustness in practical application is improved.

The present invention provides a method for repairing face defect images based on an auto-encoder and generative adversarial networks. Through combination of the auto-encoder and the generative adversarial networks, the method comprises the following steps of: (1) performing face data set defect preprocessing; (2) employing the data set after processing to train the auto-encoder to allow the auto-encoder to reach the optimal state; (3) employing the data set after processing to train the condition generative adversarial networks to allow the condition generative adversarial networks to reach the optimal state; (4) inputting a defect image to be repaired into the trained encoder to generate a face image to be repaired; and (5) inputting the image to be repaired to the condition generative adversarial networks to generate much clearer and more natural restored face image. The method improves the restoring definition of the defect face area and the fidelity of the defect content, avoids the pseudomorphism of the defect area edges to the maximum extent, restricts the generation direction of the defect area and generates much clearer and more natural restoring effect.

The invention discloses a session emotion autoanalysis method based on depth learning and belongs to the natural language processing and data mining field. According to the method, voice and text expression is learned on the basis of a de-noising auto-encoder, further through a depth learning method, depth fusion of two types of expressions is realized to acquire unified high level expression, and emotion analysis is carried out on the basis of high level expression after fusion. Through the method, depth fusion of acoustic and text characteristics is realized, and emotion classification accuracy is improved.

A system and method are provided to detect, analyze and digitally remove makeup from an image of a face. An autoencoder-based framework is provided to extract attractiveness-aware features to perform an assessment of facial beauty.

The invention discloses a face image aging synthesis method based on a feature separation adversarial network, belongs to the technical field of computer vision, and relates to a face image aging synthesis method. The method comprises the steps: supposing that the age features of a face are distributed in a Manifold, extracting the hidden features of an original image through an Auto-encoder, adding a corresponding age condition, finally generating an adversarial network through the condition, synthesizing a face image of a specific age, and maintaining the identity corresponding to an original face image. The innovativeness of the invention lies in that the method achieves the constraint of the hidden features outputted by an Encoder through the intra-class distance measure, enables the hidden features to be exclusively correlated with the identity, and removes the information correlated with the age, so as to guarantee that the age of the generated image is matched with a target agein a better way after the age is added. The method can be used for the face image aging synthesis, the cross-age face recognition and beatifying software.

The invention discloses a speech emotion recognition method. The method includes the steps that firstly, a speech signal is converted into a spectrogram to serve as initial input; secondly, a deep convolutional neural network is trained to automatically extract emotion features; thirdly, a stack type auto-encoder is trained for each kind of emotions, and all the stack type auto-encoders are fused to automatically construct membership functions of an emotion fuzzy set; fourthly, the features obtained in the second step are subjected to feature optimization by means of the fuzzy optimal theory in the third step; fifthly, emotion classification recognition is conducted by means of a Softmax classifier. The method takes abstract fuzzy properties of speech emotion information into consideration, the extracted emotion features are subjected to selective fuzzy optimization to improve the significance of the features, fuzzy membership functions in the fuzzy theory are automatically constructed by means of the concept of deep neural network layer-by-layer training, and the problem that the proper membership functions in the fuzzy theory are difficult to select and determine is solved.

The invention discloses a convolutional variational auto-encoder neural network-based finger vein identification method and system. The system comprises an image collection module, an image preprocessing module, an image feature extraction module, an image training module and an image identification module. The identification method comprises the steps of obtaining a finger vein image of a to-be-identified user; performing image preprocessing on finger vein image information, and extracting a finger vein region-of-interest (ROI) image; extracting finger vein feature codes in a finger vein ROIthrough the convolutional variational auto-encoder neural network; and inputting the feature codes to a full connection network for performing identification processing, thereby identifying identity information of the to-be-identified user. The finger vein features can be effectively extracted; the noise redundancy is improved; and the identification precision of the finger vein identification system is remarkably improved.

The invention discloses a hierarchical network attack identification and unknown attack detection method based on deep learning, which comprises the following steps of: learning behavior modes of normal flow and abnormal flow by using an auto-encoder so as to judge whether the flow to be detected is the normal flow or the abnormal flow by using the learned auto-encoder; using a deep neural networkas a discrimination model to discriminate whether the attack type of the to-be-detected flow is a certain known attack type or an unknown type; and fusing the results of the auto-encoder and the discrimination model to complete network attack classification and unknown attack detection. According to the method, normal flow can be identified, known abnormal flow can be classified, novel unknown network attacks can be detected, and the detection accuracy is improved by combining the advantages of the generation model and the discrimination model.

The invention discloses an abnormity detection method for satellite telemetry data through an adversarial network autoencoder, and the method comprises the steps: breaking the limitation of a traditional empirical model, and employing a pure data driving model; on the basis of a variational autoencoder, introducing a confrontation network idea, using a bidirectional LSTM (Long Short Term Memory) (Long-short term memory network) as a discriminator, and judging whether satellite telemetry data is abnormal or not by using errors of reconstructed data and original data; aiming at the redundancy problem of a satellite sensor, the conventional situation is broken through, and a Markov distance is used for measuring a reconstruction error. In combination with periodicity of satellite orbit operation, a dynamic threshold determination method based on a periodic time window is provided. The method has the advantages that pure data driving is adopted, expert experience is not needed, and the method can be suitable for various occasions; By combining the respective advantages of the variational auto-encoder and the generative adversarial network, the proposed network has the characteristics of high training speed and relatively easy convergence; eliminating redundant data influence between satellite telemetry data by adopting a Mahalanobis distance. According to the periodicity of the satellite, the dynamic threshold method based on the periodic time window is provided, and the misjudgment rate is reduced.

The invention discloses a hyperspectral image classification method based on deep learning, and belongs to the technical field of remote sensing image processing. The method comprises the steps that 1, dimension reduction treatment on a hyperspectral image is achieved by obtaining a data sample, conducting layer-by-layer training on an autoencoder network and further adjusting an initial weight value obtained through pre-training by adopting a BP algorithm; a data cube in each pixel neighbourhood in the hyperspectral image is taken as input of a convolutional neural network, a ground object type corresponding to a pixel serves as expected output of the convolutional neural network, the convolutional neural network is trained, the trained convolutional neural network acts on the whole hyperspectral image, and a final high-precision classification result is obtained. According to the method, the defects that in a traditional hyperspectral image classification problem, details are discarded in the dimension reduction process, space information is lost in the classification process, and the classification precision is low are overcome, the good classification precision is achieved, and the method is suitable for classification of various hyperspectral images.

The invention belongs to the technical field of computer digital image processing, and particularly relates to an image deblurring method based on an aggregation expansion convolutional network. The method comprises the steps of constructing a deep neural network, generating a network based on a condition countermeasure, wherein the network comprises a generator and a discriminator, the generatorstructure uses a stacked self-encoder module, and the self-encoder module is connected with a jump through a self-encoder structure, a residual error module is used on a construction module, residualerror module uses a residual network and multi-channel aggregation expansion convolution, and the discriminator uses a five-layer convolutional neural network; training the deep neural network: usingfuzzy image data set in public and real scenes, using image content loss function and a countermeasure loss function to train the deep neural network constructed in the previous step, and using a trained network model to carry out deblurring processing on a blurred image. According to the method disclosed by the invention, the deblurring effect can be ensured, a blurred image can be quickly and efficiently restored to a clear image, and the image deblurring efficiency can be greatly improved.

The invention discloses an image retrieval method and apparatus, an image information acquisition method and apparatus, an image identification method and apparatus, an image identification system, a method and apparatus used for calculating an image feature value, and an electronic device. The image retrieval method comprises the steps of extracting a local feature of a to-be-retrieved image; calculating a feature value of the local feature by adopting a pre-trained depth autoencoder network model; performing matching on the feature value and feature values of registered images in an image retrieval database; and selecting a registered image, which meets a preset condition, as a retrieval result of the to-be-retrieved image according to a matching result. By adopting the method, a depth autoencoder network can effectively keep distance information and a distinguishing capability between the feature values in a process of performing compressed representation on the local feature, so that the image retrieval precision can be effectively improved, the workload of rearrangement and filtration can be reduced, and the retrieval efficiency can be enhanced.

Disclosed is a data analysis and cybersecurity method, which forms a time-based series of behavioral features, and analyzes the series of behavioral features for attack detection, new features derivation, and / or features evaluation. Analyzing the time based series of behavioral features may comprise using a Feed-Forward Neural Networks (FFNN) method, a Convolutional Neural Networks (CNN) method, a Recurrent Neural Networks (RNN) method, a Long Short-Term Memories (LSTMs) method, a principal Component Analysis (PCA) method, a Random Forest pipeline method, and / or an autoencoder method. In one embodiment, the behavioral features of the time-based series of behavioral features comprise human engineered features, and / or machined learned features, wherein the method may be used to learn new features from historic features.

The invention relates to an event trigger word extraction method based on auto-encoder fusion document information, which comprises the following steps: generating a training set by using unlabeled free text corpora, and training a GRU model to construct an auto-encoder; pre-processing and labeling the training corpus, and extracting to-be-identified words; obtaining a document vector of the to-be-identified word in a document where the to-be-identified word is located by the auto-encoder, and taking the document vector as a global feature of the to-be-identified word; expressing word vectorsand entity types of the to-be-identified words in a distributed manner to serve as local features of the to-be-identified words; carrying out vector splicing on the global feature and the local feature to obtain context features of the to-be-identified word; inputting the context features to the Bi-GRU model to perform multi-classification to identify whether the to-be-identified word is an eventtrigger word and the corresponding event type of the to-be-identified word.

Methods and systems for embedding a network in a latent space include generating a representation of an input network graph in the latent space using an autoencoder model and generating a representation of a set of noise samples in the latent space using a generator model. A discriminator model discriminates between the representation of the input network graph and the representation of the set of noise samples. The autoencoder model, the generator model, and the discriminator model are jointly trained by minimizing a joint loss function that includes parameters for each model. A final representation of the input network graph is generated using the trained autoencoder model.

The invention discloses a large-compression-ratio satellite remote sensing image compression method based on a deep self-encoding network, and mainly aims to solve the problem of low compression ratio in the prior art. The method comprises the following implementation steps: cascading and stacking a plurality of self-encoders to construct the deep self-encoding network; inputting a group of training image data to the deep self-encoding network, and training the network to obtain optimized network parameters in order to obtain a deep compression network and a deep decompression network; transmitting a remote sensing image to be compressed into the deep compression network to obtain high-order sparse features, and quantifying and encoding the features to obtain final compressed code streams; and inversely quantifying and encoding the received code streams to obtain the high-order sparse features, and transmitting the high-order sparse features to the deep decompression network, wherein a final output of the network is a decompressed remote sensing image. Image processing and deep learning technologies are combined, so that large-ratio compression of satellite remote sensing data is realized. Only simple forward transmission operation is required in compressing and decompressing processes, so that high timeliness is achieved, and the storage and transmission burdens of massive remote sensing data are relieved.

The invention discloses a VAE (Variational Autoencoder)-based voice conversion method under a non-parallel corpus training condition. Under a non-parallel text condition, Bottleneck features are extracted through a deep neural network, and then, based on a VAE model, learning and modeling of a conversion function are realized, and in a conversion stage, conversion from multiple speakers to multiple speakers can be realized. The VAE-based voice conversion method under non-parallel corpus training has the advantages in three aspects: 1) dependence on a parallel text is relieved, and any alignment operation is not needed during a training process; 2) multiple source-target speaker pair conversion systems can be integrated to one conversion model, and multi-to-multi conversion is realized; and3) the multi-to-multi conversion system under the non-parallel text condition can provide a techncial support for enabling the technology to go to actual voice interaction.

The invention discloses a remote sensing image content description method based on variational self-attention reinforcement learning, and belongs to the crossing field of computer vision and natural language processing. The method comprises the following steps: pre-training a convolutional neural network fused with a variational auto-encoder by using a remote sensing image classification data set;extracting spatial features and semantic features of the remote sensing image by using a pre-trained convolutional neural network; fusing the spatial features with contextual information using self-attention; describing the data set by using remote sensing image content, decoding spatial features and semantic features by using Transformer, fusing the features, and outputting text description of the remote sensing image content; and improving text description quality by using reinforcement learning; pre-training a convolutional neural network by utilizing the remote sensing image classification data set and fusing the variational auto-encoder, and the quality of the remote sensing image content description text is optimized by using a self-attention mechanism, feature fusion and reinforcement learning.

The invention discloses a multi-to-multi speech conversion method based on a text encoder under non-parallel text conditions. The method includes a training part and a conversion part, a conditional variation self-encoder, the generative adversarial network (VAWGAN) and a text encoder (Text-Encoder) are used to implement a speech conversion system, sentence embedding representing semantics is added to the VAWGAN, the speech quality and personality similarity of the converted speech can be better improved, and high-quality speech conversion is realized. The method is advantaged in that dependence on parallel texts is relieved, speech conversion under the non-parallel text conditions is realized, not any alignment process is required in the training process, versatility and practicability ofthe speech conversion system are improved, the conversion system of multiple source / target speaker pairs can be integrated into one transformation model, that is, multi-talker-to-multi-talker conversion is realized, and the method has good application prospects in the fields of film dubbing, speech translation, speech synthesis and the like.

The invention discloses a voice converting method based on deep learning, and belongs to the technical field of voice signal processing. According to the invention, the method includes the following steps: configuring a voice encoding and decoding device AHOcoder as a feature extraction terminal and a voice synthesis terminal, training voice features by using the deep leaning method to separatelyobtain deep features of a source speaker and a target speaker, also obtaining the capability of decoding the deep features to original features, mapping the source speaker and the target speaker by using a BP neural network, thus realizing voice conversion. According to the invention, the method stitches the original features of voice, the combined feature parameters obtained from stitching are deemed to include the dynamic features of the voice features of the speaker, the training of the deep neural network is accelerated by pre-training the deep autoencoders, and by converting the deep features, the method herein obtains quality converted voice even when less voice materials are trained. The method also supports offline learning, and saves computing resources and memory of terminal devices.

Autoencoder-based, semi-supervised approaches are used for anomaly detection. Defects on semiconductor wafers can be discovered using these approaches. The model can include a variational autoencoder, such as a one that includes ladder networks. Defect-free or clean images can be used to train the model that is later used to discover defects or other anomalies.

The embodiment of the invention provides a training sample data expansion method and device based on a variational auto-encoder, and relates to the technical field of big data. The method comprises the steps of obtaining an original sample; inputting the original sample into the encoder of the variational autoencoder, wherein the encoder of the variational autoencoder comprises two neural networks, the two neural networks output Mu and sigma respectively, and Mu and sigma are both functions of the original sample; according to the square of the Mu and sigma, namely sigma 2, generating a randomnumber of corresponding Gaussian distribution; randomly sampling the standard normal distribution to obtain a sampling value epsilon, and determining a sampling variable Z according to the sampling value epsilon and the random number of the Gaussian distribution; and inputting the sampling variable Z into a decoder of the variational autoencoder, decoding the sampling variable Z by the decoder ofthe variational autoencoder, and then outputting similar samples of the original samples, and taking the similar samples as extended samples. Therefore, the technical scheme provided by the embodiment of the invention can solve the problems that the time and labor are wasted and the efficiency is low when sample data is manually expanded in the prior art.

The invention relates to a stacked SAE (Sparse Autoencoder) deep neural network-based bearing fault diagnosis method. The first layer of a network is applied to the qualitative judgment of a bearing fault, that is, the first layer of the network is applied to the fault type judgment of the bearing fault; and the second layer of the network is applied to the quantitative judgment of the bearing fault, that is, the second layer of the network is applied to the severity judgment of the bearing fault. According to the method of the invention, empirical mode decomposition (EMD) and an autoregressive (AR) model are combined together to perform pre-processing on original bearing signals, the parameters of the AR model are extracted and are adopted as the input of the network, and therefore, the input dimensions of the network can be greatly reduced, the simplification of calculation can be facilitated, and the training and testing of the network can be accelerated; a deep neural network on which the method of the invention is based can further automatically extract features of the input and qualitatively and quantitatively determine the bearing fault automatically, and therefore, the diagnostic accuracy of the method of the present invention can be ensured, and at the same time, dependence on signal processing expertise can be decreased, manual judgment is not required, the consumption of manpower can be decreased; and thus, the method has a higher practical value in the era of big data.

The invention discloses an image anomaly detection method based on a variational auto-encoder, in particular to an anomaly detection method integrating a variational auto-encoder and support vector data description, solving the problems of separation of two stages of anomaly detection and feature extraction, limitation of anomaly detection performance and incapability of coping with high-dimensional and large-scale anomaly detection tasks in traditional anomaly detection in the prior art. The image anomaly detection method comprises the steps of collecting image data; performing data set division and data preprocessing; constructing an anomaly detection model based on the variational auto-encoder; training an anomaly detection model; calculating a threshold value for distinguishing normaland abnormal image data according to the trained model; and judging whether the to-be-detected image is an abnormal image or not by using the trained model. According to the image anomaly detection method, support vector data description is adopted to perform distance constraint on the features extracted by the variational auto-encoder, and the extracted features are more suitable for anomaly detection, and the memory complexity is low, and the image anomaly detection method can be applied to high-dimensional and large-scale anomaly detection tasks.

The invention discloses a title generation method based on a variational neural network subject model, belonging to the technical field of natural language processing. This method automatically learnsthe document topic hidden distribution vector by variational self-encoder, and combines the document topic hidden distribution vector and the document representation vector learned by multi-layer neural network with attention mechanism, so as to express the comprehensive and deep semantics of the document on the topic and global level, and to construct a high-quality title generation model. Thismethod uses the multi-layer encoder to learn the more comprehensive information of the document, and improves the effect of summarizing the main idea of the full text of the title generation model; the topic implicit distribution vector of VAE learning is utilized, and the document content is represented in the abstract level of topic. The topic implicit distribution vector and the document information learned by the multi-layer encoder are combined with the deep semantic representation and context information to construct a high quality title generation model by using the attention mechanism.

The invention discloses a natural language intention understanding method in man-machine interaction.The method comprises the steps that intention labeling is conducted on text natural language instruction data, and each sentence of text is labeled with an intention; the text is vectorized, on the basis of a traditional text vector space model, information of parts of speech of a text instruction is fused, and a new text representation model, namely, a vector space model of the parts of speech is defined; a stackable denoising auto-encoder is applied to natural language instruction intention understanding, and the high-order characteristic of the instruction is extracted; at last, training and prediction are conducted through a support vector machine, and intention understanding of the natural language instruction is achieved.According to the natural language intention understanding method in man-machine interaction, more semantic information in the natural language instruction can be excavated, the recognition rate of intention understanding is increased, the stackable denoising auto-encoder is adopted, random noise is added during the training process, the actual application scene is more approached, and a model obtained from training has higher generalization capacity.