166 results about "Multi layered perceptron" patented technology

The invention belongs to the technical field of computers, and discloses a machine translation method and system based on a generative adversarial neural network. The method comprises the following steps that: on the basis of an original machine translation generation network, a discrimination network which generates network countermeasure with the original machine translation generation network is imported; a translation used for judging a target language is from a training parallel corpus and is a network machine translation result of the original machine translation generation network; and the discrimination network adopts a multi-layer sensor feedforward neural network model to realize binary classification. The system comprises the discrimination network, a generation network, a mono-lingual corpus and a parallel corpus. While manually annotated bilingual parallel corpus resources are fully utilized, and mono-lingual corpus resources also can be fully utilized to carry out semi-supervised learning; and the mono-lingual corpus resources are very rich and can be easily obtained, and the problem that required training corpora required by the neural network machine translation model are not sufficient is solved.

The invention discloses a non-standard character recognition method based on a convolution neural network and a support vector machine. The method comprises steps of 1, acquiring image signals of non-standard characters to serve as sample data; 2, establishing a convolution neural network and carrying out initialization; 3, passing the trained sample data through the convolution neural network so as to finish forward propagation; 4, carrying out error calculation and gradient calculation on a multi-layered perceptron in the step 3, and if errors are converged, extracting characteristic data and entering the step 6, or else, entering the step 5; 5, using a back propagation algorithm to propagate the errors and the gradients obtained in the step 4 to a network base layer through the convolution neural network layers by layers, judging whether the grid base layer is an input layer, and if yes, entering the step 3, or else, continuing to judge whether the next layer is the input layer until the input layer is determined and entering the step 3; 6, transmitting the characteristic data to a support vector machine for training and establishing a non-standard character recognition training model; and 7, inputting to-be-recognized non-standard character signals into the non-standard character recognition training model for recognition.

The invention relates to a point cloud registration model and method combining an attention mechanism and a three-dimensional graph convolutional network, and the model is a three-branch Siamese architecture, and comprises a Dejector model and a Descriptor model. The Detector model is used for extracting attention features of points and constructing an attention mechanism; the Descriptor model isused for generating an expression of a three-dimensional depth feature to represent the three-dimensional depth feature of the point, and learning and judging the depth feature of the point cloud. Themethod comprises the following steps: carrying out model training, and constructing a loss function to train a model by using a failure align triplet loss, so as to effectively extract attention features and descriptor features from a point cloud; after model training, carrying out the point cloud registration. According to the method, the key points and the three-dimensional depth features of each key point can be automatically extracted, in the three-dimensional graph convolutional network, the multi-layer perceptron MLP is combined with the graph convolutional network GCN, a new point cloud feature extraction module is designed, more point cloud features with identification significance can be extracted, and the accuracy of point cloud registration is improved.

The invention provides a Bayesian Decision Theory foreground extraction method combined with reflected illumination. The Bayesian Decision Theory foreground extraction method comprises the steps of appointing a point light source located on a foreground object by a user, carrying out gray level matching on an image, converting and imitating point light source illumination, strengthening image edge information, obtaining an illumination function according to before-after conversion comparison, filtering waves, reducing noise, dividing the image through a watershed algorithm, calculating a sectional drawing parameter through a Bayes formula, imitating an alpha value function curve through a multi-layer perception device, integrating the illumination function and a color distribution function, and completing extraction of the foreground object. The user is only required to appoint the position of the point light source and not required to preset edge information of a foreground and a background, the requirement for user interaction is reduced, meanwhile, time complexity of the used algorithms is series, and the defects that a common sectional drawing algorithm is large in calculated quantity and low in processing speed are avoided. Due to the facts that the illumination function is introduced and the alpha value is matched by the perception device, an accurate and complete extraction result can be obtained for the foreground object with complicated edges, and particularly for the foreground object similar colors of the edge and the ground.

The invention relates to a method for detecting the pulmonary artery blood pressure by using a heart sound analysis method of a multilayer feedforward network. The method comprises the steps of: extracting a heart sound signal; preprocessing the heart sound signal by fast Fourier transform and normalized average Shannon energy distribution to obtain heart sound signal features; filtering the heart sound signal features by using a principal component analysis method; and performing training learning on the filtered hear sound signal features by using a perceptron neural network or a multilayer perceptron feedforward neural network to obtain an optimal neural network between the heart sound signal features and a pulmonary artery blood pressure value. The heart sound signal features are signal features of a first heart sound and a second heart sound; the heart sound signal features comprise crest frequency, average frequency, crest amplitude, average amplitude, duration and time interface of heart sounds. By adopting the method for detecting the pulmonary artery blood pressure by using the heart sound analysis method of the multilayer feedforward network, provided by the invention, the pulmonary artery blood pressure can be easily and conveniently detected at low cost in a non-intrusive and no risk manner.

The invention discloses a video character recognition method based on a multi-modal feature fusion deep network. The video character recognition method is a deep learning target recognition multi-modal fusion algorithm provided specially for the target recognition problem of multi-modal character video feature data. According to the video character recognition method, a network structure of the algorithm is composed of a plurality of single-modal multi-layer sensor identification modules and a multi-modal feature fusion module. The video character recognition method comprises the steps of: preprocessing multi-modal data generated by a video, training a plurality of deep networks by using the preprocessed different modal data, on the basis, subjecting features generated by a plurality of sub-networks to weighted fusion, and combining a feature weighted fusion module with models of different modals to achieve a better identification result. By adopting the video character recognition method, a multi-modal feature set weight fusion strategy is used for constructing a video figure target recognizer on a public video figure data set (iQIYI-VID-2019) for multi-modal features generated bythe preprocessed video, multi-model integration is not needed, and the average precision mean value of a single model reaches 89.52%.

The present invention discloses a speech emotion recognition method through fusion of feature assessment and a multi-layer perceptron. The method comprises the steps: S1, extracting multi-dimensional emotion feature parameters of a training speech set corresponding to various emotion states, and obtaining an original feature set; S2, performing rating ordering of various emotion feature parameters in an original feature set, and obtaining a feature set after ordering; S3, obtaining a plurality of feature subsets with different quantities from the feature set after ordering, using a multi-layer perceptron to perform classification of each feature subset, and selecting an optimal feature subset according to a classification result; and S4, using the multi-layer perceptron to train an emotion classification model for the optimal feature subset, and performing emotion recognition of the speech to be recognized through the classification model obtained through training. The realization method is simple, the speech emotion recognition method through fusion of feature assessment and the multi-layer perceptron can fuse the feature assessment and multi-layer perceptron to realize emotion recognition, and the emotion recognition precision and the efficiency are high.

The invention discloses a complex chemical process fault diagnosis method based on deep learning multi-model fusion. According to the invention, two neural networks automatically extract fault features from two aspects respectively, then fuse the features and input the features into a multi-layer perceptron (MLP) for further feature compression and extraction, and finally output a diagnosis result. Features are extracted through a convolutional neural network (CNN) and a long-term and short-term memory (LSTM) respectively, and the features finally extracted by the network have space and time characteristics at the same time; and the final diagnosis is carried out by integrating the characteristics of the two aspects, so that the problem of large calculated amount of the existing traditional diagnosis technology is solved, and the technical problem that the fault diagnosis cannot be accurately carried out in the complex chemical process due to the fact that the characteristics extractedby a single network are too one-sided is also solved.

The invention discloses a many-to-many voice conversion method and system based on speaker style feature modeling. Firstly, a multi-layer sensor and a style encoder are added to a StarGAN neural network, effective extraction and constraint of speaker style features are realized, the defect of limited speaker information carried by one-hot vectors in a traditional model is overcome, then, an adaptive instance normalization method is adopted to realize full fusion of semantic features and speaker personality features, so that the network can learn more semantic information and speaker personality information, furthermore, a lightweight network module SKNet is introduced into the generator residual error network, so that the network can adaptively adjust the size of a receptive field according to multiple scales of input information, the weight of each feature channel is adjusted through an attention mechanism, the learning ability of frequency spectrum features is enhanced, and the details of the frequency spectrum features are refined.

The invention discloses a traditional Chinese medicine syndrome diagnosis method and device based on deep learning and an attention mechanism. The method comprises the steps of obtaining medical record data with symptoms of a to-be-diagnosed patient; converting the medical record data into vector data; inputting the vector data into a trained traditional Chinese medicine syndrome diagnosis model,and outputting a diagnosis result. The traditional Chinese medicine syndrome diagnosis model comprises an attention mechanism module constructed by using a matrix mapping layer, an activation functiontan h and softmax, and a prediction deep neural network constructed by using a multilayer perceptron and an activation function sigmoid. The method comprises the following steps: firstly, performingone-hot coding mapping on medical record data with symptoms of a patient, and converting the medical record data into vector data; adopting a traditional Chinese medicine syndrome diagnosis model of amultilayer perceptron combining deep learning and supervised learning to diagnose the syndrome of a patient, so that the diagnosis and prediction accuracy of the traditional Chinese medicine syndromeof the patient can be improved.

The invention discloses an interest point recommendation method based on a space-time sequence and social embedding ranking, and the method comprises the steps: firstly proposing a mixed depth model based on a horizontal convolution filter, a vertical convolution filter and a multilayer perceptron model, and capturing the user preference and the impact of a space-time sequence mode; secondly, modeling the social relationship of the user by adopting a metric learning method; then, adopting a unified framework based on matrix decomposition to integrate personal interests of the users, sign-in sequence modes and social information of the users; and finally, optimizing a target loss function by adopting a BPR standard, fitting a partial order relationship of the user on the interest point pair, and finally generating an interest point recommendation list. According to the invention, the interest point recommendation method integrating the hybrid depth model and metric learning is constructed, so that the final recommendation is more accurate.

The invention discloses a graph convolutional network system and a 3D object detection method based on the graph convolutional network system. The system comprises a shape semantic extraction module used for modeling geometric positions of points in point cloud features of an image; a multi-layer perceptron which is connected with the shape semantic extraction module and is used for extracting multi-level semantic features by using a multi-layer graph convolutional neural network and filtering the multi-level semantic features by using an attention mechanism; a proposal generator which is connected with the multi-layer perceptron and is used for summarizing the multi-level semantic features and weighting to generate a primary proposal; and a proposal reasoning module which is connected with the proposal generator and is used for predicting a 3D bounding box and a semantic category of the object in the image by utilizing the global semantic features and the primary proposal. According to the method, the detection performance of the whole graph convolutional network system is effectively improved, the precision of 3D object detection is improved, and the interpretability of the deep network is higher.

The invention provides a three-dimensional point cloud processing method, device and equipment. The method comprises the following steps: acquiring point cloud data comprising a plurality of points; inputting the point cloud data into a pre-trained convolutional neural network; the convolutional neural network comprises a convolution module for geometric feature information sharing; for each pointin the point cloud data, obtaining a neighbor point of the point in the Euclidean space based on the convolution module, and determining a neighbor point of the point in the feature value space basedon the neighbor point of the point in the Euclidean space; aggregating the neighbor points of the points in the Euclidean space and the neighbor points of the points in the feature value space to obtain aggregated features; and performing feature learning on the aggregated features by using a multi-layer perceptron, and performing maximum pooling operation on the dimensions of the neighbor pointsto obtain output features. A convolution structure is directly constructed to process three-dimensional point cloud data, so that the problems of data missing and data redundancy are effectively relieved.

The invention discloses a dynamic linearization system of a power amplifier based on mode recognition. The dynamic linearization system comprises a feature vector generation module, a multilayer sensor module, a predistortion coefficient lookup table module, a vector signal module, a transmitting link module, a power amplifier module and a feedback loop module. The invention further discloses a method using the dynamic linearization system. The dynamic linearization system disclosed by the invention adopts a new standard, that is, an amplitude modulation to amplitude modulation (AM/AM) mode, to recognize a working state of the power amplifier, and uses a recognition result of a multilayer sensor neural network on the power amplifier AM/AM to index different predistortion coefficients so asto achieve the dynamic linearization of the power amplifier. Compared with the traditional power lookup tables and power indexing algorithms, the method has the advantages of being able to provide better linearization performance. In addition, the dynamic linearization system disclosed by the invention is also capable of solving the linearization problem under a dynamic condition of the power amplifier caused by many other non-power factor changes such as temperature, frequency, bandwidth, signal type, etc.

The invention discloses an optimization method of a hybrid bat algorithm and a multilayer perceptron optimization method, and the method comprises the steps: firstly, improving an original reverse learning strategy in the aspect of population reconstruction, and proposing disturbance multi-strategy reverse learning, so as to reconstruct an effective diversified search space; secondly, in the aspect of global exploration and search, by introducing a whale optimization algorithm and improving the whale optimization algorithm, providing a self-adaptive constraint step length whale optimization algorithm so as to make up for the defects of an original bat algorithm in the aspect of global exploration capacity; thirdly, in the aspect of local mining search, providing a bat algorithm based on Cauchy variation and dynamic correction by introducing Cauchy variation and designing a dynamic correction strategy, so that the local search capability of the original bat algorithm is improved; and fourthly, under the synergistic effect of the three strategies, effectively improving the precision and stability of an optimization result by the new algorithm obtained by optimization; finally, applying the new algorithm to the multi-layer perceptron training problem, and obtaining relatively high classification precision.

The invention discloses a method for synthesizing a virtual viewpoint image by using implicit neural scene representation on the basis of multi-view three-dimensional cross-view loss, and is suitable for the field of computer vision. The method comprises the following steps: acquiring an image data set which needs to generate a virtual viewpoint; preprocessing the training image data set, and performing feature point extraction and matching on the input training image data set based on a feature matching algorithm Sift in the preprocessing stage; processing the obtained training image data and the extracted feature point information, and inputting the processed training image data and extracted feature point information into a multi-layer perceptron network for training; inputting test image data into the trained multi-layer perceptron network, and then obtaining a tested rendered image through volume rendering; and generating a virtual viewpoint image based on the trained multilayer perceptron network. Therefore, the data volume of the neural network during training fitting scene representation is reduced, and centralized sampling is performed in combination with the image depth information, so that the operation speed and performance of the neural scene representation can be improved, and a high-quality virtual viewpoint image is generated.

Some embodiments are associated with a system and method for deep learning unsupervised anomaly prediction in Internet of Things (IoT) sensor networks or manufacturing execution systems. The system and method use an unsupervised predictive GAN model with multi-layer perceptrons (MLP) as generator and discriminator.

The invention discloses a numerically-controlled machine tool thermal error compensation method and system based on workpiece size data. The method comprises the steps: carrying out the analysis of aprocess capability index (Cpk) based on a machining process, and obtaining a key temperature point from the actual measurement temperature data of cutting work; according to the key temperature pointand workpiece size data machined under the machine tool load condition, a numerical control machine tool thermal error model based on workpiece size detection data is constructed through a multi-layerperceptron neural network (MLP), then obtaining the motion compensation amount of a motion axis corresponding to workpiece size features under the actual machining condition, and achieving thermal error compensation through the external coordinate zero offset function of a numerical control machine tool. According to the method, the thermal deformation of the machine tool and the workpiece causedby the machining process is considered, and the thermal error of the machine tool under the actual machining condition is effectively compensated.

The invention relates to the technical field of medical artificial intelligence, in particular to an embryo development potential prediction method and system, equipment and a storage medium. The method comprises the steps: inputting an embryo initial image into a blastocyst prediction model, and obtaining an embryo feature vector; inputting the embryo feature vector into a bidirectional long-short-term memory network to obtain embryo development features; based on a cross-modal feature fusion mechanism, obtaining fusion features according to clinical data and the embryonic development features; and inputting the fusion features into a first multi-layer sensor, and predicting to obtain the embryo pregnancy rate. According to the invention, multi-focal-segment embryo videos shot in the early stage are analyzed, and the fusion features with the space-time characteristic is obtained by utilizing a multi-focal-segment selection model and a time transfer model, so that the pregnancy rate of the embryo cultured in vitro is predicted in real time, and the prediction accuracy is improved; meanwhile, by predicting the blastocyst forming probability and the euploid probability, doctors are assisted in early embryo screening, and therefore the labor cost is reduced.

The invention discloses a pedestrian re-identification three-dimensional data set construction method and device based on a neural radiation field. The method comprises the following steps: S1, carrying out image acquisition on a pedestrian to be input through a group of cameras with different visual angles; s2, sampling through camera rays in a scene to generate a three-dimensional space position point set, and converting the observation direction of a camera corresponding to the three-dimensional space position point set into a three-dimensional Cartesian unit vector; s3, inputting the three-dimensional space position point set and the observation direction, converted into the three-dimensional Cartesian unit vector, of the three-dimensional space position point set into a multi-layer sensor, and outputting corresponding density and color; according to the pedestrian re-identification three-dimensional data set construction method and device based on the neural radiation field, a brand-new pedestrian re-identification data set construction method is provided, and a new thought for data set construction is provided. Compared with a traditional data set construction method, the data acquisition mode is more direct and clear through images and spatial positions acquired by multiple devices.