46results about How to "Reduce model size" patented technology

An artificial neural network compression coding device comprises a memory interface unit, an instruction buffer memory, a controller unit and an operation unit, wherein the operation unit is used for performing corresponding operation on data from the memory interface unit according to the instruction of the controller unit. The operation unit mainly performs three steps of: 1, multiplexing an input neuron with weight data; 2, performing an addition tree operation for adding the weighted output neurons after the first step through an addition tree or obtaining a biased output neurons through adding the output neurons and a bias; and 3, executing a function activation operation, and obtaining a final output neuron. The invention further provides an artificial neural network compression coding method. The artificial neural network compression coding device and the artificial neural network compression coding method have advantages of effectively reducing size of an artificial neural network model, improving data processing speed of the artificial neural network, effectively reducing power consumption and improving resource utilization rate.

A system and method for training a language classifier are disclosed that may include obtaining an initial dictionary-based classifier model, stored in a computer memory, the model including a plurality of classifier n-grams; pruning away selected ones of the n-grams that do not significantly affect a performance of the classifier model; adding, to the model, selected supplemental n-grams that increase the effectiveness of the classifier model at identifying a language of a text sample, thereby growing the classifier model; and enabling the adding step to include adding n-grams of varying order, thereby enabling the provision of a variable-order model.

The invention discloses a license plate detection model based on an improved YOLOv3 network and a construction method. The improved YOLOv3 network is used for inputting a license plate image and extracting three feature maps of different scales; carrying out up-sampling on the obtained three feature maps with different scales, scaling depth features to the same proportion, then carrying out down-sampling, and carrying out decoding through a constructed convolution layer to generate a feature map after feature enhancement; performing feature aggregation on the generated feature maps of the three different scales after feature enhancement and the feature maps of the three different scales extracted from the YOLOv3 feature extraction network to generate a feature pyramid, and obtaining an improved license plate detection model of the YOLOv3 network; and training the license plate detection model to obtain a final model. According to the method, the detection speed is greatly improved, thepyramid multi-scale feature network is introduced to enhance the features of the backbone network and generate a more effective multi-scale feature pyramid, and the features are better extracted fromthe input image.

The invention discloses a CNN (Convolutional Neural Network)-based voiceprint recognition method for anti-record attack detection. The CNN-based voiceprint recognition method comprises the following steps: step S101, acquiring to-be-detected voice frequency and establishing a voiceprint recognition data set; step S102, carrying out character extraction on the voice frequency of the voiceprint recognition data set, wherein extracted characters comprise a character MFCC (Mel Frequency Cepstrum Coefficient) and a bottleneck layer character; step S103, establishing a CNN by combining MobileNet andUnet; step S104, inputting the voiceprint recognition data set to the CNN for training; step S105, inputting the bottleneck layer character to the trained CNN by using testing voice frequency, thus obtaining a testing score for judging real talk or record voice frequency. The CNN-based voiceprint recognition method disclosed by the invention combines the characteristics of two models of the Unetand the MobileNet, has lower model complexity, i.e., lower model size, smaller computation resource loss and higher recognition accuracy rate, and can be transplanted and applied to a mobile phone side and an embedded device.

The invention discloses a light human body action recognition method based on deep learning. The method comprises the steps of firstly constructing a light-weight deep learning network (SDNet) combining a shallow network and a deep network, wherein the network comprises a shallow multi-scale module and a deep network module, and constructing a light-weight human body action recognition model basedon deep learning based on the network; in the model, firstly utilizing the SDNet to perform feature extraction and representation on the space-time double flow; utilizing a time pyramid pooling layerto aggregate the video frame level features of the time stream and the space stream into the video level representation; obtaining a recognition result of the space-time double flow to the input sequence via a full connection layer and a softmax layer, and finally fusing the double-flow result in a weighted average fusion mode to obtain a final recognition result. By adopting the light human bodyaction recognition method based on the deep learning, the model parameter quantity can be greatly reduced on the premise of ensuring that the recognition precision is not reduced.

The invention discloses a method for carrying out segmentation and depth-of-field rendering on a monocular portrait based on WNET. A WNET network parameter model is constructed by superposing UNET, the trained WNET network parameter model is loaded into a mobile client, preliminary segmentation of a portrait picture is realized in the mobile client, and a mask image subjected to preliminary segmentation is zoomed to the size of an original image by adopting a bilinear interpolation method; morphological operation is carried out on the mask image, corrosion and expansion processing is carried out on the mask image, connected regions with edges not in a preset region are removed, remaining connected regions are reserved, edge refinement processing is carried out to obtain a portrait mask image, and the separated foreground and the background after Gaussian blur are synthesized to obtain a depth-of-field rendering image; according to the method, the calculation amount and the model size can be greatly reduced, and the portrait segmentation precision is improved, so that portrait depth-of-field rendering of the monocular camera of the mobile terminal is realized.

The invention relates to the technical field of language processing, in particular to a language model training method in a self-reconstruction mode, which comprises the following steps of: S1, extracting at least one sentence to be trained from a pre-training text, segmenting the sentence to be trained into single word sequences, and mapping corresponding single sub-sequences into a text matrix through position coding; s2, establishing a neural network structure in combination with a transformer model and a self-attention mechanism; s3, taking the text matrix as an input sample of a neural network structure, and taking the transformer model as a parameter to train and optimize to obtain a target function; and S4, repeating the steps S1 to S3 to update the target function until a set optimization condition is reached so as to obtain a pre-training model. The invention also provides a system and a computer readable medium.

This invention relates to context-dependent Chinese phone identifying and modeling method, which applies initial consonant right-dependent and final sound left dependent modeling method including: a, creating a context-dependent basic modeling unit by relating the initial consonant with the adjacent right final sound and relating the final sound with its adjacent left initial consonant, b, utilizing the state clustering method to train the model parameters to get an initial HMM, c, utilizing the sub-space clustering method to compress the HMM to generate a final model.

The invention discloses a globe country image identification method based on the convolutional neural network. The method comprises steps that firstly, a globe country image data set is constructed through collecting various types of images of various countries on commonly-used teaching globes by means of data acquisition and data enhancement, and multiple images of the different countries from different spatial locations and angles under different lighting conditions and different focusing conditions are collected; secondly, each image in the data set is compressed and preprocessed; and thirdly, a new convolutional neural network model is designed based on characteristics of classic convolutional neural network models MobileNet and DenseNet, the new model is trained based on the collecteddata set, the model is made to learn the image characteristics of each country on the globe, and classification is further carried out. The method is advantaged in that the characteristics of the models MobileNet and DenseNet are integrated in the designed identification model, and relatively high identification accuracy and relatively low model complexity are realized.

A system and method for bounded model checking of computer programs includes providing a program having at least one reachable property node. The program is decomposed for bounded model checking (BMC) into subproblems by creating a tunnel based on disjunctive control paths through the program. A reduced BMC sub-problem obtained using BMC unrolling, while using path constraints imposed by the at least one tunnel. For the reachable property node, determining a quantifier-free formula (QFP) in a decidable subset of first order logic. Satisfiability of the QFP is checked, independently and individually, to determine whether the QFP is satisfiable for the subproblem. The decomposing is continued until the a BMC bound is reached.

The invention discloses an SCUC model power flow constraint feasible region boundary identification method based on rank judgment. The method comprises the following steps: generating a complete branch active power flow inequality constraint set according to the safety operation requirements of a power grid in an SCUC problem; constructing an equation boundary corresponding to the inequality constraint; obtaining all feasible region vertexes and corresponding equality boundary sets thereof; counting a feasible region vertex set corresponding to each equation boundary; judging whether each equation is a boundary constraint or not according to the vertex, and obtaining the boundary constraints of all equation forms; and restoring the boundary equality constraints into a feasible region boundary inequality constraint set. According to the method, the model scale of the unit commitment optimization problem can be effectively reduced, the calculation time of unit commitment optimization is shortened, and then the reliability and robustness of optimization problem solving are improved; and by applying the method, the energy-saving dispatching calculation speed of the power grid can be increased, the energy-saving dispatching effect can be improved, the operation loss of the power grid can be reduced, carbon emission can be reduced, and better energy-saving and environment-friendly effects can be realized.

A pitching arrangement for a model helicopter, comprising a securing base, a plurality of servers supported by the securing base, a main connecting shaft communicated to the servers, and has a supporting ring, a control device operatively provided between the servers, and is axially mounted to the main connecting shaft, and a plurality of connecting rods connecting the servers and the control device through the main connecting shaft, wherein the supporting ring is mounted on the main connecting shaft, wherein the connecting rods are connected to the supporting ring of the main connecting shaft.

The invention discloses a lightweight neural network single-image defogging method based on multi-scale convolution. The method comprises the following steps: (1) dividing a data set into a training set, a verification set and a test set; (2) training the proposed network model by using the training set and the verification set; (3) testing the trained model by using the test set; and (4) evaluating the model by adopting a measurement standard. According to the method, the multi-scale defogging module is specially designed and used for extracting the picture information under each scale, whichhas an important influence on the final defogging result; model parameters are reduced as much as possible under the condition that information of all scales is fully considered, the size of the model is compressed, and then the calculation complexity of the model is reduced; a defogging effect with a higher peak signal-to-noise ratio is achieved, and meanwhile defogging of a single picture of transmissivity and global atmospheric light does not need to be estimated.

The invention discloses a multi-level whole-process monitoring method for power equipment. The method comprises the following steps: rendering and displaying a three-dimensional scene in a designatedarea of a two-dimensional picture of the power equipment for displaying an online operation state; switching to a three-dimensional scene of corresponding equipment by activating equipment pictures inthe two-dimensional picture; wherein the three-dimensional scene is sequentially divided into a plurality of levels of three-dimensional scene models according to the connection relationship of the power equipment; the switching among the three-dimensional scene models of multiple levels is realized by activating the movable parts with the hierarchical link information in the three-dimensional scene models, and the display of the equipment information in the three-dimensional scene models of multiple levels is realized through instantiated variable transmission. According to the multi-level whole-process monitoring method for the power equipment, the whole-process monitoring from the whole structure of the power equipment to the details of the single equipment is realized, and the equipment perception capability, the defect discovery capability, the state management and control capability and the emergency disposal capability of operation and maintenance personnel on the power equipment of a converter station are improved.

The invention provides a test device and a test method for a long-distance water diversion channel circulating water flow freezing model. The test device comprises a U-shaped flow channel, a water tank, a water pump, and a variable frequency control cabinet. The U-shaped flow channel comprises two sections of straight water conveyance flow channels and a section of 180 DEG arc-shaped water conveyance flow channel. According to the test device and the test method for the long-distance water diversion channel circulating water flow freezing model, a circulating flowing water flow is formed in the U-shaped flow channel by the water tank and the water pump; the water flow speed can be adjusted by the variable frequency control cabinet to guarantee the running stability of the water flow; and long-distance water diversion channel water conveying process in winter and cold environment can be simulated, and a model test platform is provided for researching the freezing characteristic law andthe ice melting mechanism of long-distance channel engineering under different water delivery flow rates.

The invention particularly relates to an intelligent terminal video analysis algorithm combining edge calculation and deep learning. According to the intelligent terminal video analysis algorithm combining edge calculation and deep learning, a moving target detection model is carried on edge end equipment, original video data is processed in real time, and a moving target in a video is analyzed, so that the delay problem of a network is effectively reduced; meanwhile, the purpose of video compression is achieved by reserving key frame information including a background and a moving target, sothat a large amount of storage space is saved, and the searching time of positioning abnormal behaviors is shortened. According to the intelligent terminal video analysis algorithm combining edge calculation and deep learning, the video analysis process completed in a rear-end processor in the traditional process is replaced, and the training model obtained by utilizing deep learning is carried inthe cameras, so that each camera is a microprocessor and has the capability of processing the video stream in real time, and the storage pressure and the network transmission pressure are greatly reduced.

The invention discloses a pedestrian detection method based on multi-scale self-attention feature fusion, and relates to the field of artificial intelligence. The method comprises the following steps: (1) acquiring data of a pedestrian detection image; (2) designing the size of a pedestrian in the detection image; (3) dividing positive and negative samples of a pedestrian data set; and (4) building a pedestrian detection model. According to the method, the Faster R-CNN is adopted for building a pedestrian detection framework, and a multi-scale feature fusion network model is provided, so that more and more effective feature information can be extracted and overfitting can be avoided; a GPU with good performance is used for training, so that the training speed is greatly increased; the receptive field is expanded, so that small targets can be detected and the resolution ratio is not reduced; and the method is very suitable for accurate and rapid detection of pedestrians.

The invention discloses a retaining wall soil pressure model test device under the plane strain condition and a test method thereof. The problem that the simulation of a test device in the prior art is inconsistent with actual engineering is solved. The retaining wall soil pressure model test device and the test method thereof have the beneficial effects of realizing the determination of the relation between soil pressure and lateral displacement under different displacement modes, realizing the assumption of plane strain in a complete sense, greatly reducing the model size and simplifying thetest steps. According to the scheme, the retaining wall soil pressure model test device comprises a U-shaped structure, a main wall body on one side, a side wall and a plurality of propulsion mechanisms, wherein a plurality of pressure sensors are arranged on the surface of the main wall body, the side wall surrounds the U-shaped structure, and the U-shaped structure is internally provided with fillers; and the propulsion mechanisms are correspondingly connected with the main wall body, and the propulsion mechanisms are arranged up and down to enable the upper half section and/or the lower half section of the main wall body to rotate or translate relative to a soil retaining base under the drive of the two groups of propulsion mechanisms.

The embodiment of the invention provides a model training method, a model training device and an entity extraction method, and relates to the technical field of data processing. The model training method is applied to electronic equipment, and comprises the steps of firstly, obtaining training data and training a preset model according to the training data to obtain a first training model; secondly, obtaining a second training model according to part of parameters of the first training model; and then, training the second training model according to a loss function between the first training model and the second training model to obtain an entity extraction model. Through the setting, the small-scale second training model realizes the function of the first training model, so that the modelscale is reduced on the premise of not reducing the accuracy, and the model prediction and reasoning efficiency is improved.

The invention discloses a Chinese rhythm hierarchy prediction method based on self-attention. The method comprises the steps of learning a large number of unlabeled texts to obtain word vectors of single words, converting a to-be-predicted text into a word vector sequence by utilizing the word vectors, inputting the word vector sequence into a trained rhythm level prediction model, and outputtingword positions and rhythm levels of the text. According to the method, Chinese rhythm hierarchy prediction is carried out by using a rhythm hierarchy prediction model, characteristics of character granularity are used as input while prediction performance is guaranteed, dependence on a word segmentation system and possible negative effects are avoided, the model directly models the relation between any two characters in a text through a self-attention mechanism, and parallel calculation can be achieved. Pre-training is carried out by using additional data to improve the model performance, so that each rhythm level of the to-be-processed text can be accurately predicted at the same time, and wrong transmission is avoided.

The invention discloses a neural network filter pruning method based on a batch feature heat map. The method is mainly used for reducing the model memory space and improving the model reasoning speed.The method comprises the following steps: loading and finely adjusting a pre-training model on a given data set; generating a batch feature heat map of each layer of the model; obtaining Mask of eachfilter based on the gray threshold to score the filters; performing random non-repetitive extraction on the given data set to update the score of the filter; realizing pruning of each layer of filterby taking the score of the filter as a measurement criterion; and re-training the pruned model to recover the precision, and the like. According to the invention, the problems of large storage capacity and low reasoning speed of the neural network model are solved, so that the pruned neural network model can be applied to a resource-limited scene under the condition of extremely low precision reduction.

The invention relates to an SAR image ship target identification method. According to the SAR image ship target identification method, training data are preprocessed through a proposed inter-class sample imbalance processing technology (including up-sampling processing based on data enhancement and a method for generating batches in proportion), and the training data sent into a network are kept in inter-class balance while the diversity of the training data is improved; through the provided dense residual network used for SAR image ship target identification, reutilization of original features can be realized while more new features are learned; through a proposed loss function based on center loss, network model parameters are adjusted, and simultaneous optimization of intra-class compactness and inter-class separability is realized. A recognition result on OpenSARShip shows that the designed dense residual network has higher accuracy, smaller model size and smaller calculation amount compared with a conventional neural network.

The invention discloses an attribute recognition device, method and system and a neural network for recognizing object attributes. The attribute recognition device uses a neural network to recognize an attribute of an object, the neural network including an extraction sub-network, a determination sub-network, and a recognition sub-network composed of at least two recognition branches. The attribute recognition apparatus includes: a unit that extracts a feature from an input image using an extraction sub-network, in which the extracted feature can be used for recognizing an attribute of an object by all recognition branches in the recognition sub-network; a unit that determines at least two identification branches from the identification sub-network based on the input object category by using the determination sub-network; and a unit for recognizing, from the input image, an attribute of an object belonging to the input object category on the basis of the extracted feature, using at least the determined recognition branch. According to the present invention, the model size of the neural network for identifying the attributes of the object is greatly reduced.

Systems and methods are disclosed for a centrosymmetric convolutional neural network (CSCNN), an algorithm/hardware co-design framework for CNN compression and acceleration that mitigates the effects of computational irregularity and effectively exploits computational reuse and sparsity for increased performance and energy efficiency.