213results about How to "Improve the effect of the model" patented technology

The invention discloses a dialog strategy online realization method based on multi-task learning. According to the method, corpus information of a man-machine dialog is acquired in real time, current user state features and user action features are extracted, and construction is performed to obtain training input; then a single accumulated reward value in a dialog strategy learning process is split into a dialog round number reward value and a dialog success reward value to serve as training annotations, and two different value models are optimized at the same time through the multi-task learning technology in an online training process; and finally the two reward values are merged, and a dialog strategy is updated. Through the method, a learning reinforcement framework is adopted, dialog strategy optimization is performed through online learning, it is not needed to manually design rules and strategies according to domains, and the method can adapt to domain information structures with different degrees of complexity and data of different scales; and an original optimal single accumulated reward value task is split, simultaneous optimization is performed by use of multi-task learning, therefore, a better network structure is learned, and the variance in the training process is lowered.

The invention relates to a text abstract generation method based on a sentence association attention mechanism, and belongs to the technical field of natural language processing. The method comprisesthe following steps: firstly, encoding a document by using a layered bi-directional long short-term memory Bi-LSTM network to obtain sentence semantic vectors, then analyzing an association relationship among sentences by virtue of a gating network to realize sentence-level importance and redundancy evaluation, and finally, providing a decoding algorithm based on a sentence association attention mechanism to generate an abstract. When a neural network abstract generation framework is constructed, sentence relevance analysis is fused, and the evaluation capacity of the model for sentence importance and redundancy in an original text is improved. According to the method, the performance of the generative abstract is effectively improved, and a relatively good effect is achieved on the current ROUGH evaluation index.

The invention relates to the field of robot vision, pattern recognition and machine learning, in particular to an underwater robot target detection method, a system and related equipment, aiming at improving the robustness of target detection technology to underwater target occlusion, deformation and illumination changes. The object detection method of the invention comprises the following steps:obtaining an original image to be detected; normalizing the pixel value of the original image to be detected, and obtaining the image to be detected after preprocessing; the preprocessed image being input into the target detection network for detection, and the bounding frame of the region of interest and the probability of belonging to each target class being obtained; according to the bounding box of ROI and the probability of belonging to each target class, the improved non-maximum suppression algorithm being used to obtain the bounding box and the class of the target object, wherein, a deformable convolution neural network is used to extract feature map in the target detection network, and the candidate region method is used to detect the target. The detection method of the invention improves the detection precision under the condition of guaranteeing the speed.

The invention discloses a voice synthesis method and a device. The voice synthesis method comprises steps of performing text characteristic extraction on a text to be synthesized to obtain the context characteristic information, obtaining a pre-generated model, wherein the pre-generated model is generated by training according to the context characteristic information of the training sample and converted acoustic parameter, and the converted acoustic parameters comprise a plurality of rhythm level fundamental frequency parameters, determining the model output parameter corresponding to the context characteristic information according to the model, wherein the model output parameters comprise a plurality of the rhythm level fundamental frequency parameters, performing the fundamental frequency reconstruction on the plurality of rhythm level fundamental frequency parameter, and synthesizing voice according to the parameter after the fundamental frequency reconstruction and the other parameters in the model output parameters. The method can improve the performance result of the synthesized speech.

The present invention relates to a method and a system for predicting the load of a power distribution network. The method comprises the steps of obtaining a unsupervised training sample set, a supervised training sample set and a test sample set according to the time information and on the basis of the historical load influence factor of the power distribution network and the historical load value of a to-be-predicted area; according to the unsupervised training sample set, subjecting a DBN model layer in a pre-established load prediction model to unsupervised training layer by layer, wherein the load prediction model includes a DBN model layer and a linear neural network layer; adopting network parameters obtained through during the unsupervised training step as the network parameter initial values of the load prediction model; according to the supervised training sample set, subjecting the load prediction model to supervised training, and obtaining an optimal load prediction model; testing the test sample set by using the optimal load prediction model so as to obtain a load prediction value of the to-be-predicted area. The present invention can realize the high-precision loadprediction of the intelligent power grid environment under the influence of various factors.

The invention discloses a pyramid network Chinese herbal medicine identification method based on an attention mechanism, which comprises the following steps: 1) constructing a Chinese herbal medicinedata set, and making a Chinese herbal medicine training set and a Chinese herbal medicine test set; 2) constructing a feature fusion structure block based on a channel attention mechanism, and introducing a competition attention module; 3) adding a spatial attention mechanism to the feature fusion structure block of the pyramid network, adjusting the two information flows by using a spatial collaborative attention module, and fusing the two adjusted information flows as output; 4) constructing a pyramid network based on an attention mechanism, and training by using a Chinese herbal medicine training set; and 5) transmitting the pictures in the Chinese herbal medicine test set to the trained network model to identify the Chinese herbal medicine types corresponding to the pictures, thereby improving the Chinese herbal medicine identification accuracy and performance, assisting related industrial personnel to identify the Chinese herbal medicines, and facilitating non-professionals to identify the Chinese herbal medicines.

The invention discloses a scene semantic segmentation method based on deep learning. The method comprises a training stage and a testing stage; the training stage comprises the steps: employing Resnet101 for pre-training on a COCO data set to obtain a pre-training model, and then loading the pre-training model into a constructed convolutional neural network to extract a low-level feature image; performing high-level feature extraction and feature fusion on the low-level feature image through a feature enhancement network, an adaptive deformable cavity space convolution pooling pyramid networkand a feature attention network in sequence, and finally outputting a semantic segmentation Mask graph through an up-sampling operation, and obtaining a convolutional neural network semantic segmentation model weight; the test stage comprises the steps: inputting a PASCAL VOC 2012 or Cityscapes test data set into the weight of the convolutional neural network semantic segmentation model, and obtaining a predicted semantic segmentation Mask graph. According to the method, the boundary contour precision of the target image and the accuracy of scene semantic segmentation can be improved.

The invention aims to provide a method and device used for training a language model according to a corpus sequence. The corpus sequence used for training the target language model is acquired, initial order information of the target language model is set as the current training order, and the following operations are carried out through iteration in combination with the highest order information of the target language model till the current training order exceeds the highest order information, wherein the operations include that according to the current training order, a smoothing algorithm corresponding to the target language model is determined; according to the corpus sequence, the target language model is trained through the smoothing algorithm to acquire an updated target language model; the current training order is updated. In comparison with the prior art, the method and device have the advantages that different smoothing algorithms are adopted for language models with different orders according to the characteristics of the language models with different orders, the advantages of different smoothing algorithms are played, and thus better model establishment effects can be achieved. Furthermore, the method and device can be combined with voice identification, and thus the accuracy of the voice identification can be improved.

The invention discloses a zinc floatation state dividing method based on isomerism textural features. Zinc floatation image textural features are extracted by combining a gray-level co-occurrence matrix algorithm which has a good effect on high-frequency band textural features and a Gauss Markov random field algorithm which has a good modeling effect on low-and-medium-frequency texture images, and the zinc floatation image textural features are subjected to Gauss normalization to serve as a textural feature vector. In an integrated clustering algorithm, partitional clustering with high efficiency is conducted firstly to eliminate the influences of noise points and outliers, then a hierarchical clustering algorithm with high clustering quality and high stability is adopted to combine clustering centers output through partitional clustering, and then a final clustering result is obtained. Experiments prove that the extracted textural feature quantity has high mode separability, and foam in different states can be distinguished with the integrated clustering algorithm; furthermore, the method can be directly realized on a computer and is low in cost, high in efficiency and easy to implement.

The invention discloses a parameter optimization control method of semiconductor advance process control (APC). In semiconductor technological process, a traditional method uses a linear prediction model for the optimization control method of batch process. The parameter optimization control method of the semiconductor advance process control uses an optimized back propagation (BP) neural network prediction model based on genetic algorithm, optimizes the initial weight values and threshold values of the neural network through the genetic algorithm, uses selecting operation, probability crossover and mutation operation and the like according to the fitness function F corresponding to each chromosome, and outputs the optimum solution finally to determine the optimum initial weight value and the threshold value of the BP neural network. The performance of the BP neural network is improved with an additional momentum method and variable learning rate learning algorithm being used, so that the BP neural network after being trained can predict the non-linear model well. The genetic algorithm in the method has good global searching ability, a global optimal solution or a second-best solution with good performance is easy to obtain, and the genetic algorithm well promotes the improvement of modeling ability of the neural network.

The invention provides a scenic spot recommendation method based on hybrid supervised learning. The scenic spot recommendation method comprises the steps of obtaining historical tourist touring data;constructing a scenic spot knowledge graph; performing corresponding attribute sub-graph extraction on the scenic spot knowledge graph according to the attribute category of the scenic spot; generating a scenic spot sequence; training the scenic spot sequence and mapping the scenic spot sequence into a low-dimensional vector space to generate a feature vector; adding and averaging the vectors of each scenic spot under different attributes to obtain a fused semantic feature vector of each scenic spot; learning tourist vectors and scenic spot potential vectors; carrying out matrix decompositionon the tourist vector and the fused semantic features to obtain a first interaction vector; obtaining a second interaction vector of the tourist vector and the scenic spot potential vector by using amulti-layer perceptron; splicing the first interaction vector and the second interaction vector and performing normalization processing to obtain a score of the tourist for the scenic spot; ranking the scores of the tourists for the scenic spots from high to low, and obtaining a top _ k scenic spot recommendation list by taking the first K scenic spots with the highest scores.

The invention provides a personalized scenic spot recommendation method based on a knowledge graph and long and short-term preferences of a user. The method comprises the following steps: preprocessing a historical tourist scenic spot sequence of a tourist, and carrying out scenic spot-coding conversion; using node2vec to randomly walk to obtain a scenic spot sequence, and using Skip-gram model in word2vec to obtain feature vectors of tourists and scenic spots; adding bias to the feature vectors of the scenic spots to serve as input of a GRU network, and then utilizing the GRU network to train and output the potential vector of each scenic spot; allocating different weights to the scenic spots, multiplying the weight of each scenic spot by the potential vector of the scenic spot, accumulating to obtain the long-term preference of the current tourist, splicing the long-term preference of the current tourist and the current preference of the tourist, and multiplying the spliced preference by the weights to obtain a final vector; and performing dot product operation on the final vector and the current preference of the tourist to obtain an estimated score of the scenic spot, performing normalization processing on the estimated score of the scenic spot to obtain a prediction probability of each scenic spot, and taking scenic spots corresponding to the former K scores to obtain a top_k scenic spot recommendation list.

The invention discloses a streaming phonetic transcription system based on a self-attention mechanism. The streaming phonetic transcription system based on the self-attention mechanism comprises a feature front-end processing module, a self-attention audio coding network module, a self-attention prediction network module and a united network module. The feature front-end processing module is usedfor receiving an input acoustic feature and converting into a vector with specific dimensionality; the self-attention audio coding network module is connected with the feature front-end processing module and is used for receiving the processed acoustic feature and obtaining an coded acoustic state vector; the self-attention prediction network module is used for generating a language state vector according to an input prediction mark of the last moment; and the united network module is connected with the self-attention audio coding network module and the self-attention prediction network module, and is used for combining with an acoustic state and a language state and calculating the probability of a new prediction mark. The invention provides a streaming feedforward voice coder based on the self-attention mechanism, so that the calculation efficiency and the precision of a traditional voice coder are improved.

The invention belongs to the technical field of solid mineral exploration and three-dimensional geological modeling, and specifically relates to a method for extracting stratum boundary line to carryout three-dimensional modeling by utilizing exploration line profile. The method comprises the following steps: preparing and processing basic data, and carrying out translation, rotation and other operations on a profile to erect a two-dimensional profile map at the actual position of the three-dimensional space; extracting different geologic body boundaries on the section by utilizing the vertical geologic section; and establishing an integral three-dimensional geologic model by using the extracted geologic body boundary line according to the idea of breaking the whole into parts, carrying out block modeling and merging the whole parts. According to the method, the problem that the three-dimensional geology difficulty of a complex geology background research area is high is solved, the modeling precision is improved, the modeling effect is optimized, and technical support is provided for digitalization and quantification of mineral exploration.

The embodiment of the invention discloses an acoustic model establishing method, an acoustic model establishing device, a speech synthesis method, a speech synthesis device, a facility and a storage medium. The acoustic model establishing method comprises the following steps: acquiring a phoneme sequence sample of a plurality of training samples from a corpus, and acquiring the context feature ofeach phoneme and the duration of each phoneme in the phoneme sequence sample, wherein the rhotic accent phoneme in the phoneme sequence sample is split to two phonemes; extracting acoustic features from the training samples; and by adopting the phoneme sequence sample, training the acoustic model by taking the context feature of each phoneme and the duration of each phoneme in the phoneme trainingsample as the input of the acoustic model and the acoustic features as the output of the acoustic model, so that the pretrained acoustic model is obtained. The modeling performance of the rhotic accent is good, the synthesis of the rhotic accent can be well realized, then the rhotic accent not appearing in the corpus can be synthesized, and meanwhile, the recording cost of linguistic data in thecorpus can be reduced.

The invention discloses an industrial soft measurement method based on a bionic intelligent ant colony algorithm, comprising the following steps of: (1) determining key variables used in the process of soft measurement and collecting variable data from the historical database of the key variables as training samples when a system is normal; (2) normalizing the variable data of the training samples in a soft measurement intelligent processor to enable the mean value of the processed key variables to be 0 and the variance to be 1; and (3) establishing the function of a RBF (Radial Basis Function) neural network to obtain a soft measurement model, training the RBF neural network by using the normalized variable data, and then determining the hidden layer node number and the primary function center of the RBF neural network by using the bionic intelligent ant colony algorithm. The method of the invention has the advantages of convenient determination of control parameters and wide range of application and can be generalized in various industrial processes; and besides, the invention has good effect of adopting the RBF neural network to model, high data fitting precision and simple and convenient operation and avoids complicated mechanism modeling.

The invention discloses a hot topic detection method based on weighted LDA and improved Single-Pass clustering algorithm. The hot topic detection method of the Pass clustering algorithm comprises thefollowing steps: preprocessing text data, including Chinese word segmentation, stop word removal and feature word weighting; modeling the text data by utilizing a weighted LDA topic model, realizing feature dimension reduction by mining hidden topic information in the text data, and filtering and denoising a vectorized result; subjecting text vectorization result processed by LDA topic model weighted by feature words to improved Single-Pass clustering algorithm to carry out clustering; and calculating a hot value of the topic cluster by utilizing the topic cluster scale and the topic cluster compactness, and identifying the hot topic. The detection method has the advantages of being low in algorithm complexity, low in dependency on text input time sequence and the like.

The invention discloses an improved Bouc-Wen model hysteresis modeling method, and belongs to the technical field of control. In the method, the Bouc-Wen model is taken as a back-end network part of a fuzzy neural network, parameters of a Bouc-Wen model can be adaptively adjusted according to a neural network, and asymmetric hysteresis loops related to frequency and amplitude of a piezoelectric ceramic micro-positioning platform are enabled, so that high-precision hysteresis modeling is realized. The method comprises the following steps of: deriving a discretized Bouc-Wen parameter model equation; constructing an improved Bouc-Wen model; measuring and obtaining data required by modeling according to the piezoelectric ceramic micro-positioning platform; and using a gradient descent method to obtain input and output data pairs. According to the method, the modeling effect of the model on the frequency-dependent and amplitude-dependent asymmetric hysteresis loop is greatly improved, and a foundation is laid for design and practical application of a controller behind a piezoelectric ceramic micro-positioning platform.

The invention relates to the video field, in particular to the content analysis of the video and the object detection field. The purpose of the invention is to solve problem that the moving object segmentation is easily affected by illumination changes in the application of video monitoring, such as abrupt change of solar illumination in daytime, automobile light at night, a great amount of falsealarm can be generated by using the traditional method. Two key technologies are utilized in the invention for realizing the above purpose. One key technology is to take a space-time video block as abasic process unit, thus apparent spatial information and time motion information are simultaneously utilized to carry out background modeling and prospective detection segmentation. The other key technology is to effectively capture background modeling by utilizing online sub-space learning method. The method can be used in the systems for processing and analyzing the video content, which need tocarry out background modeling and prospective detection, such as video monitoring system.

The invention relates to the field of digital image post-processing, in particular to a feature enhancement-based residual neural network and an image deblocking effect method. The residual neural network based on feature enhancement introduces a local residual unit, a global feature enhancement unit and a local feature enhancement unit. The three basic units promote each other, so that the learning ability and the modeling ability of the target neural network are greatly enhanced. Accurate mapping from a low-quality image with a blocking effect to a high-quality image can be established for an image blocking removal effect problem, and finally, a JPEG compressed image with given quality can be processed through the established effective mapping to obtain a high-quality image. According tothe image deblocking effect method, the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) of the image can be remarkably improved, the efficiency, quality and robustness of imagedeblocking are greatly improved, and the method has profound significance in the field of image post-processing.

The invention discloses a scene classification method based on a nonparametric space judgment hidden Dirichlet model. The scene classification method mainly overcomes the defect that an existing classification method does not contain scene space information. The scene classification method is implemented in the following steps of (1) inputting images, (2) extracting image block features, (3) initializing nonparametric space judgment hidden Dirichlet model parameters, (4) establishing the nonparametric space judgment hidden Dirichlet model, and (5) classifying image scenes. The scene classification method based on the nonparametric space judgment hidden Dirichlet model utilizes image blocks containing space information, can describe the image scenes more abundantly, and improves the accuracy rate of image scene classification.

The invention discloses an iteration interpolation method based on face triangle mesh adaptive subdivision and Gauss wavelet. First, performing triangulation to a face model through a mesh optimized face subdivision method to obtain an optimal triangle; determining whether a spatial triangle is intersected or not, calculating depth information of an interpolation point through a two-dimension Gauss wavelet function on the basis of status of three vertexes of the spatial triangle to obtain a complete three-dimension coordinate of the interpolation point; obtaining the two-dimension coordinate (x, y) of the interpolation point, after finishing determining the two-dimension coordinate (x, y) of the interpolation point, performing recovery to a z axis of the interpolation point through the two-dimension Gauss wavelet function; determining the value of the interpolation point at the z axis and determining the optimal value of m according to the three vertexes of x, y, and z. The beneficial effects of the invention are: failure in shaping a recognizable three-dimension face model due to insufficient three-dimension characteristic points is effectively avoided, and modeling effect is impressive.

The invention discloses a method of correcting multipath errors of GNSS positioning and attitude determination. The method comprises steps of short baseline solution and multipath calculation, multipath sky hemisphere model establishment based on trend surface analysis and real-time multipath correction. The trend surface analysis method is used to fit multipath spatial distribution characteristics in a sky grid, the high-frequency multipath modeling ability by the multipath modeling method based on spatial repeatability is enhanced, and the positioning and attitude determination precision isenhanced; requirements on the scale of the sky grid are reduced, and the calculation efficiency is improved; and the method is also applicable to dynamic scenarios with static and multipath environments unchanged.

The invention discloses an MCS (Mesoscale Convection System) identification and tracking method based on image anchor-frame-free detection, which comprises the following steps of: step 1, preprocessing infrared brightness temperature data of an original stationary satellite, carrying out mesoscale convection system marking on an infrared cloud picture obtained after processing, and then randomly dividing a training set, a verification set and a test set; step 2, constructing an instance segmentation network based on no anchor frame, the network being used for extracting image features, detecting a mesoscale convection system and segmenting specific instances; step 3, performing training set image enhancement, using a transfer learning supervised training instance to segment the convolutional neural network, and automatically learning network parameters; step 4, performing mesoscale convection system detection and segmentation on the geostationary satellite infrared nephograms at adjacent moments by using the trained model; and step 5, realizing the tracking of the mesoscale convection system according to a related target matching principle.

The invention belongs to the field of digital image processing, and particularly relates to a residual neural network based on hole convolution and a two-stage image demosaicing method. According to the invention, a shallow feature extraction unit, a local residual unit and a deep feature extraction unit are introduced based on a residual neural network; the interaction of the three basic units greatly enhances the learning ability and modeling ability of the target neural network. Accurate mapping from the mosaic image to the RGB color image can be established for the image demosaicing problem, and finally the mosaic image in the Bayer CFA mode can be processed through the established effective mapping to obtain the RGB color image; meanwhile, a two-stage image demosaicing model is introduced, prior information is fully utilized, the modeling capability of the network is improved, and the understanding space is optimized; by means of the image demosaicing method, the peak signal-to-noise ratio of the image can be remarkably increased, the image demosaicing efficiency, quality and robustness are greatly improved, and the image demosaicing method has far-reaching significance in thefield of image processing.

The invention belongs to the technical field of image enhancement, and discloses a low-illumination image enhancement method based on an improved depth separable generative adversarial network, and the method comprises the steps: constructing an improved depth separable convolutional generative adversarial network model; training the constructed deep separable convolution generative adversarial network model; and performing low-illumination image enhancement by using the trained depth separable convolution generative adversarial network model. The model parameter quantity can be greatly reduced and the calculation complexity can be reduced while the low-degree image enhancement effect is ensured, so that the problem of insufficient memory in the current research can be solved. Depth separable convolution is introduced and improved, so that the invention is also suitable for a low-illumination image enhancement task while model parameters are reduced, and the calculation efficiency is improved. Compared with a low-illumination image enhancement algorithm with the same calculation complexity and parameter model quantity level, the method has obvious superiority in effect.

The invention discloses a face recognition method which belongs to the technical field of face recognition. The face recognition method comprises the following steps: extracting face characteristics by adopting a fuzzy mean clustering method and forming a training sample set, and grouping the training sample set according to the face characteristics; and then establishing a parallel neural network and synthesizing the output of the parallel neural network, thereby acquiring a final recognition result; and updating training samples during a recognition process by a classifier. By adopting the face recognition method, the problem that a traditional neural network is low in recognition speed and low in recognition precision during a process of processing a mass of image data sources is solved, and the adaptive ability of a system is promoted.

The invention belongs to the technical field of knowledge maps, and discloses a knowledge map information representation learning method, a system, equipment and terminal, and the knowledge map information representation learning method comprises the steps: carrying out the preprocessing according to a path constraint resource distribution method; calculating the reliability of all paths, and outputting the reliability to a training set and a test set; initializing the model and setting parameters; generating a triple according to an iterator, and randomly replacing head and tail entities; calculating a loss function of the triple according to the score function; calculating a loss function of an additional path according to the path reliability; performing parameter optimization by using an Adam method; and performing model verification by using entity prediction and relation prediction. According to the method, rich path information contained in the knowledge graph is considered, the modeling effect of entities and relationships is improved, the modeling of the relationships can be optimized by inputting vectors into a complex plane and using rotation to represent the vectors, and the method can be used for link prediction and recommendation systems.

An event subject identification method comprises the following steps: identifying an entity in a target text by adopting a predetermined entity identification model; marking the identified entity in the target text by adopting a first predetermined symbol to obtain a marked target text; obtaining an embedded vector of each character in the target text according to the marked target text; inputtingthe obtained embedded vector of each character in the target text into a named entity recognition prediction model to obtain an output label corresponding to each character in the target text; and according to the obtained output label corresponding to each character in the target text, identifying an event main body in the target text. According to the invention, the identification accuracy canbe improved.

The invention brand-new proposes a patent knowledge graph construction method based on an object agent database, and the method comprises the steps: firstly achieving the storage of basic classes through employing a source class in an object agent, and then completing the modeling of a relation class through employing four agent classes. On the basis, the construction of the query is realized by utilizing the cross-class query statement. Compared with a current knowledge graph mode, the multi-level semantic association can be better expressed, the connection operation of the data table is reduced, and the query efficiency is improved.