204results about How to "Improve inference speed" patented technology

The invention relates to a hardware framework for carrying out reasoning acceleration by aiming at a convolution neural network, and a working method thereof. The hardware framework comprises a preprocessing unit, a hardware acceleration unit and a storage unit, wherein the preprocessing unit is used for preprocessing an original image frame which is originally input; the hardware acceleration unit is used for reading the preprocessed original image frame to be convoluted, a convolution kernel coefficient and an offset parameter for convolution, executing fully connected layer calculation after convolution is finished, and outputting a calculation characteristic judgment result after the fully connected layer calculation is finished; the storage unit is used for storing the original imageframe which is originally input, the convolution kernel coefficient, the offset parameter, output data obtained by each convolution and the output data of the fully connected layer. According to the hardware framework, the problems that a traditional processor is low in speed and high in time delay, real-time reasoning can not be realized and the like are solved, and a new solution is provided fordesigning the processor which carries out the reasoning calculation by aiming at the CNN (Convolution Neural Network).

The invention belongs to the technical field of image target detection, and particularly discloses a fabric defect detection method based on a YOLO v4 improved algorithm, and the method introduces a latest lightweight attention module (CA) on a backbone network, not only can capture cross-channel information, but also can capture direction perception and position perception information, the network can carry out emphasized detection on an interested target, and a deformable convolutional network (DCN) is added to enhance the adaptability of the network to flaws with variable shapes, so that the detection accuracy is improved. And for the feature fusion part, adaptive weighted fusion (ASFF) is used on the basis of an original path aggregation network, so that features extracted by each feature layer are fused with different weights before prediction, meanwhile, a cross-stage local network structure (CSP) is used for replacing partial convolution of the feature fusion part, and under the condition that the speed is ensured, and the fabric defect detection accuracy of the network is greatly improved.

Methods for faster statistical inference in computation based recognition problems on highly parallel processors with multiple cores on-a-chip are disclosed, which include: selectively flattening levels of the recognition network to improve inference speed (improving the recognition model); selectively duplicating parts of the recognition network to minimize a critical section in atomic accesses to as few as one atomic instruction (improving the recognition procedure); and combining weight and source port into one 32-bit word to minimize the number of atomic operations. These methods have been implemented on an NVIDIA GTX 280 processor in a Large Vocabulary Continuous Speech Recognition (LVCSR) embodiment, and achieve more than a 10× speed up compared to a highly optimized sequential implementation on an Intel Core i7 processor.

The invention relates to a monocular vision scene depth estimation method based on deep learning, and the method comprises the steps of employing a VGG-13 network model, employing a depth separable convolution layer to replace a standard convolution layer so as to reduce the model parameter quantity, and obtaining a network model which can be used for obtaining a parallax image; inputting the monocular image into the trained network model to generate the disparity maps of multiple scales, and generating a single disparity map consistent with the input image in scale by combining multi-scale fusion and disparity map smoothing; and generating a corresponding depth image according to the geometric transformation relationship between the disparity map and the depth map in the multi-view geometry. The method has the beneficial effects that the simple and easily available binocular visible light image is used for training the network model without acquiring the high-cost real depth data, andthe standard convolution is replaced with the depth separable convolution, so that the parameter quantity of the network model can be reduced to one seventh of the previous parameter quantity, and the reasoning speed of the model is increased.

The invention discloses a lightweight multi-speaker voice synthesis system and electronic equipment. The system comprises a text feature extraction and normalization module, a speaker feature extraction module, a feature fusion module and a voice generation module. The text feature extraction and normalization module is used for carrying out encoding and feature extraction on to-be-processed textinformation by adopting a lightweight encoder, carrying out duration prediction on each word or phoneme corresponding to text deep features output by the lightweight encoder by adopting a lightweightduration prediction network, and carrying out length normalization processing to obtain regular text features with length equal to that of a target Mel spectrum. The speaker feature extraction moduleis used for generating features capable of representing a target speaker. The feature fusion module is used for fusing the features of the target speaker with the regular text features. The voice generation module is used for carrying out deep feature extraction, dimension mapping and residual error integration on the fused features and generating voice. The system supports multi-speaker voice synthesis and is high in synthesis speed.

The invention relates to a high-precision direction signboard target extraction method based on point cloud. The method comprises the steps of acquiring point cloud data, RGB picture information and track information of a signboard, and thinning RGB pictures according to the track information to form a picture data set, and extracting position information and attribute information of the signboardfrom the picture data set through a target detection deep learning model; detecting the shape attribute of the signboard through another deep learning model, associating the shape attribute of the signboard into the point cloud data through the track information, and storing the point cloud data as a point cloud data set in a classified manner; converting the point cloud data in the point cloud data set into a point cloud picture, and predicting contour information of the point cloud picture by utilizing a semantic segmentation deep learning model; back-projecting the point cloud picture intothe point cloud data set according to a mapping relationship, and finely extracting shape points with inaccurate contour points according to the strength information of the point cloud data; and associating and storing the attribute information and the form points through the position information. According to the invention, the time for manually extracting the traffic elements in the signboard is reduced.

The invention discloses a case reasoning device and the method based on a case tree, wherein, the device comprises a case tree memory module, an information input module, a search module, and an information output module; the method comprises the following steps: first, judging whether the similarity of the root node of the case tree and the feature information of the problem accords with the set value; if according with the set value, then implementing the second step; otherwise, implementing the fourth step; second, judging whether the similarity of the child nodes in the case tree accords with the set value in order; if in conformity with the set value, then judging whether the similarity of the subordinate child nodes of the child nodes and the feature information of the problem accords with the set value; otherwise, finishing the judgment to the similarity of the subordinate child nodes; third, repeating the second step until finding all the nodes with the similarity according with the set value; fourth, outputting the reason and solution of the problem corresponding to the nodes with the similarity in conformity with the set presentation threshold value in the nodes. The invention has the advantages that the reasoning speed is improved greatly and the method is convenient.

The invention provides a small-target human body detection method based on balanced sampling and nonlinear feature fusion, and the method comprises the steps: carrying out the fusion and enhancement of the features of all scales in a training process, and enabling the scale where a small-target human body is located to obtain sufficient human body semantic information; using an instance balance sampling strategy for positive samples during subsequent loss calculation, it is guaranteed that human bodies of different scales can be trained in the same way, an intersection-parallel ratio balance sampling strategy is used for negative samples, it is guaranteed that enough samples are difficult to participate in training, and finally the classification capacity of the network is improved; and aknowledge distillation training method is adopted in a training strategy, so that the model size is compressed while the precision is ensured, and the reasoning speed is increased.

The invention discloses an SAR image ship detection method and system based on lightweight deep learning, wherein the method comprises the steps: carrying out the preprocessing of a large-size SAR image, and selecting a training sample; introducing a Ghost module and Ghost Bottleneck to upgrade the YOLOv5s to obtain a preliminary lightweight model of the YOLOv5s; on the basis of the preliminary lightweight model, utilizing the network pruning and knowledge distillation of a traditional model lightweight algorithm to realize further lightweight of the model; using a TensorRT reasoning optimizer for carrying out reasoning acceleration on the lightweight YOLOv5s model, and deploying the lightweight YOLOv5s model on NVIDIA Jetson TX2; cutting large-size SAR images to be detected, and sequentially sending the cut large-size SAR images to the model to complete detection; and synthesizing a detection result, and screening a prediction frame on the final large-size SAR image by using NMS non-maximum suppression. On the premise of satisfying acceptable precision loss, the parameter quantity and the floating point operand of the model are compressed, and the detection speed is improved.

The invention relates to a language model training method based on knowledge distillation, a text classification method, a language model training device based on knowledge distillation, a text classification device, electronic equipment and a non-temporary computer readable storage medium. The language model training method based on knowledge distillation comprises a first word vector layer parameter determination step and a language model training step. The text classification method comprises the steps of obtaining a to-be-classified text; based on the to-be-classified text, obtaining a keyword code list of the to-be-classified text through extraction; obtaining a word vector of each keyword corresponding to the to-be-classified text through a language model according to the keyword code list; and through the text classification layer, obtaining a classification result of the to-be-classified text. According to the method, a knowledge distillation method is adopted, the dependence on a labeled sample is reduced while the accuracy of the model is reserved, and the reasoning speed is increased by simplifying the structure of the model, so that the applicability and reliability ofthe text classification method in an intelligent auxiliary secret setting system are improved.

The invention relates to a pedestrian trajectory prediction method based on Informer, and belongs to the technical field of computer vision and automatic driving. The method comprises the following steps: (1) carrying out position coding on a track sequence, and inputting the track sequence into an Informer-encoder to obtain a feature vector; (2) enabling a hidden variable predictor to generate hidden variables according to the feature vectors; (3) generating a track key point according to the hidden variable, and initializing a position corresponding to an Informer-decoder by using the track key point; and (4) carrying out position coding on the initialization sequence of the Informer-decoder, generating a prediction track by combining the hidden variables again, and calculating a loss function. According to the method, an Informer self-attention distillation technology, a probability sparse self-attention mechanism and a generative decoder are adopted as core technologies of a basic network, trajectory key points are predicted based on a hidden state, and then the trajectory key points are used for initializing the position corresponding to the decoder. The method can be used for trajectory prediction of an automatic driving vehicle for pedestrians, vehicles and other agents, helps the vehicle to make decisions better, and protects traffic safety.

The invention relates to the technical field of computer vision and intelligent traffic, and discloses a small target pedestrian detection method based on an improved YOLO algorithm. The method comprises the following steps of: firstly, making a small target pedestrian data set by using KITTI and INRIA data sets; secondly, re-clustering a pre-selected box by adopting a k-means algorithm based on the data set; thirdly, on the basis of a YOLO-V3 model, using a Mish activation function for replacing ReLU, simplifying a feature extraction network , and using a PANet structure for feature fusion; and finally, optimizing a loss function, and calculating a coordinate error by using CIoU. Compared with a YOLO-V3 model, the improved algorithm has the advantages that the network reasoning speed is increased by 3.2 AP and 20.8%, and the algorithm has certain practicability in small-target pedestrian detection tasks.

The invention discloses a debugging method for a RDFS (Resource Description Framework Schema) body in a distributed environment. According to the invention, the debugging method comprises the steps of: firstly, performing compression treatment on RDFS body data by applying a dictionary coding technology to reduce storage space of the data and speed up body ratiocinating and debugging speed; then realizing a RDFS body ratiocination algorithm based on a parallel distributed calculation framework Spark, and storing the ratiocinated body debugging information after compression in an HBase distributed database to prepare for debugging information inquiry in a body debugging stage. The debugging method for the RDFS body in the distributed environment realizes a distributed gain storage updating algorithm based on the Spark framework; in order to obtain contained explanation in the distributed environment, the traditional methods are improved, and two ways of obtaining the contained explanation are realized.

The invention discloses a dense crowd counting method based on a multi-scale feature pyramid network, and aims to solve the problems of density map estimation and crowd counting in a complex and crowded scene. A feature pyramid fusion module is utilized to effectively capture multi-scale context information in the crowd image, and better feature expression is further obtained; focusing a high-density position in the crowd image by using a feature attention perception module to reduce background interference; restoring the image to the original size by using a bilinear difference value; according to the method, the problem that crowd counting is difficult due to large head scale change, serious crowd shielding and poor illumination conditions in a dense scene is effectively solved, the robustness of the model to noise is improved, accurate crowd counting and high-quality prediction of the density map can be achieved, and the crowd counting accuracy is improved. Due to the fact that the grouping convolution module is utilized, the reasoning speed is high.

The invention discloses a pedestrian intention multi-task identification and trajectory prediction method under the view angle of an intelligent automobile, and the method comprises the steps: carrying out the multi-task identification of the pedestrian intention and trajectory prediction according to different kinds of spatio-temporal context information captured in an environment, including visual feature information and non-visual feature information, through a novel neural network architecture, employing a hybrid method, and carrying out the multi-task identification of the pedestrian intention in the view angle of the intelligent automobile; performing joint visual space and dynamic reasoning on each information source by using a feedforward network and a loop architecture, fusing visual information and non-visual information of m historical time steps at T time, classifying current states or actions of pedestrians at the time T, predicting future crossing intentions, outputting actions and intention probabilities at the time T, and obtaining a crossing intention prediction result; the model also predicts a trajectory from time T to time T + n. The method comprehensively considers global spatio-temporal context information of a traffic environment where pedestrians are located, comprises five kinds of visual and non-visual information sources, improves the accuracy of pedestrian crossing intention prediction, and has the advantages of being small in occupied memory amount, high in reasoning speed, complementary in associated task performance and the like.

Methods for faster statistical inference in computation based recognition problems on highly parallel processors with multiple cores on-a-chip are disclosed, which include: selectively flattening levels of the recognition network to improve inference speed (improving the recognition model); selectively duplicating parts of the recognition network to minimize a critical section in atomic accesses to as few as one atomic instruction (improving the recognition procedure); and combining weight and source port into one 32-bit word to minimize the number of atomic operations. These methods have been implemented on an NVIDIA GTX 280 processor in a Large Vocabulary Continuous Speech Recognition (LVCSR) embodiment, and achieve more than a 10× speed up compared to a highly optimized sequential implementation on an Intel Core i7 processor.

The embodiment of the invention discloses a calculation graph execution method, computer equipment and a storage medium, and the calculation graph execution method comprises obtaining a binary instruction executed by an artificial intelligence processor corresponding to a calculation graph according to an operation instruction of a fusion operator when a universal processor compiles the calculation graph with the fusion operator.

The invention provides a neural network model reasoning method and device, electronic equipment and a readable medium, and relates to the technical field of neural networks. The method comprises the following steps of preprocessing a to-be-identified picture by adopting an assembly line method; loading a pre-trained neural network model into a neural network reasoning acceleration engine; and reasoning the preprocessed to-be-identified picture by adopting the loaded neural network reasoning acceleration engine to obtain a reasoning result. Picture preprocessing is subjected to parallel processing through an assembly line method, CPU resources are fully utilized, the reasoning performance is improved, a neural network reasoning acceleration engine is adopted for reasoning the acceleration engine, and the reasoning speed of a neural network model is increased.

The invention provides a convolutional neural network compression method based on channel number search, and the method comprises the steps: firstly selecting a target image data set for image recognition, and dividing the target image data set into a training set and a test set; secondly, inputting the training set into a convolutional neural network for training, and outputting an importance index of each channel corresponding to the convolutional neural network; comparing the importance index value with a set threshold value, and abandoning a channel corresponding to the importance index lower than the threshold value to obtain an improved convolutional neural network; and finally, replacing the convolution layer of the improved convolutional neural network with the deep convolution layer to obtain a lightweight network model, and inputting the test set into the lightweight network model to verify the identification performance of the lightweight network model. According to the method, the lightweight convolutional neural network model is constructed by combining the search of the number of channels with the improvement of the network convolution mode, so that the parameters ofthe network model are greatly reduced, and the operation speed of the model is increased.

The invention relates to the technical field of smart home, and discloses a cleaning robot control method and system. The method comprises the steps of respectively carrying out the global position distance obtaining, local obstacle avoidance information collection and self angle collection, building a global coordinate system, obtaining the position information of each obstacle, carrying out thepreprocessing of obstacle data, and obtaining the position information of each obstacle; representing the indoor environment map through a grid method, matching and designing the sizes of sub grids, obtaining a global total path and splitting the global total path into a continuous grid point set, establishing a local coordinate system, obtaining local obstacle position information, controlling and adjusting the advancing direction of the robot, and executing sequential traversal of the grid points according to the global total path. The cleaning robot can be controlled to carry out full-coverage sweeping, obstacles can be found in real time, the obstacle avoidance capacity is high, computing resources can be saved, the actual working interval can be traversed efficiently at low energy consumption, collision-free and low-path-repetition-rate efficient sweeping is achieved, and high practical value and wide application prospects are achieved.

The invention discloses a cloth defect detection method and system based on a deep neural network, and belongs to the technical field of pattern recognition. The method comprises the steps that a defect cloth image training set is used for training a deep neural network model, labels are defect types and real frame position information, the deep neural network model is composed of a backbone network and a detection network, and the backbone network is used for extracting three feature maps with different scales from defect cloth images; the detection network includesthree detection sub-networks and the detection result fusion module, wherein the three detection sub-networks are the same in structure. Each detection sub-network is used for detecting a defect type and prediction frame position information from the feature map, and consists of three dense connecting blocks, and the feature channel connection between the dense blocks is used for enhancing feature transfer, and the detection result fusion module is used for performing non-maximum suppression on the prediction result to obtain a final prediction frame and a defect type, and inputting to-be-detected cloth into the traineddeep neural network model to obtain a detection result, so that the type and the position of the defect in the cloth can be detected more quickly and accurately.

The invention provides a roadside laser radar target detection method, which comprises the steps: selecting multiple frames of background point cloud data in different time periods from data acquiredby a roadside laser radar as background data, performing rasterization processing on the background data, performing statistics on grid features, and performing calculation to obtain grid average statistical features as a background grid statistical table; for the actually measured original point cloud data, carrying out rasterization processing with the same grid size as the background data, performing statistics to obtain grid statistical features corresponding to the original point cloud data, and carrying out background filtering in combination with the background grid statistical table to obtain non-background point cloud data; and inputting the non-background point cloud data into a constructed multi-scale voxel three-dimensional detection network, and outputting a detection resulttensor of the target, the result tensor comprising category information and bounding box information of the target. According to the method, a large number of invalid points are filtered through background filtering, the training and reasoning time of the network is remarkably shortened, interference of a large number of background points is avoided, and the precision of a detection result is improved.

The invention relates to the field of neural network algorithm hardware implementation, and discloses a quantitative calculation method and system for a convolutional neural network. The quantitativecalculation method comprises the steps of: allowing all calculation layers of a convolutional neural network to be respectively matched and quantized in a multi-valued quantification mode and a multi-bit quantification mode according to the calculation precision and calculation capability requirements, allowing the calculation layers after multi-bit quantification to be mapped to a high-precisionarray, and carrying out high-precision calculation; and mapping the calculation layers after multi-bit quantification to a high-calculation-power array, performing high-calculation-power calculation,and completing calculation of the convolutional neural network according to a high-precision calculation result and a high-calculation-power calculation result in combination with non-calculation layers. According to the invention, the reasoning speed of the convolutional neural network is increased; the accuracy is ensured; meanwhile, the network power consumption is reduced as much as possible;and high practical value and wide application prospect are achieved.

The invention discloses a speech recognition method and device based on a self-attention mechanism and a memory network, relates to the technical field of artificial intelligence, and can solve the problems of poor speech recognition effect and low efficiency caused by certain limitations of an existing model in calculation complexity and accuracy during speech recognition at present. The method comprises the following steps: updating an encoder structure and a decoder structure of an RNN-Transducer model according to the self-attention mechanism and the memory network LSTM; extracting voice sequence features and text sequence features of the target voice; and determining a target text label corresponding to the target voice based on the voice sequence features and the text sequence features by using the updated RNN-Transducer model. The method is suitable for online voice recognition, for example, the method can be applied to scenes such as dialogue robots, online education and real-time conference systems.

The invention discloses a convolution kernel similarity pruning-based recurrent neural network model compression method and belongs to the technical field of computer electronic information. The method comprises the steps of: loading a pre-trained recurrent neural network model into a compressed recurrent neural network for training the pre-trained recurrent neural network model so as to obtain aweight matrix-initialized recurrent neural network model; calculating the L2 norms of each convolution kernel in the recurrent neural network model, sorting the L2 norms, and selecting convolution kernels in a norm pruning rate range and pruning the convolution kernels; and calculating the weight center of the convolution kernel of the pruned pre-trained recurrent neural network model, selecting convolution kernels in a similarity pruning rate P2 range and pruning the convolution kernels, performing gradient updating on a weight matrix corresponding to the convolution kernels, and pruning parameters in the updated weight matrix to obtain a compressed recurrent neural network model. According to the recurrent neural network model compression method provided by the invention, the large recurrent neural network model is effectively compressed while accuracy loss in a pruning process is reduced.

The invention discloses a sample classification method, system and medium based on a multi-sample reasoning neural network. The method comprises the following steps: (1) establishing a multi-sample reasoning neural network MSIN; (2) taking several training samples of different sample fields as input values, inputting them into the multi-sample reasoning neural network MSIN, and training the multi-sample reasoning neural network MSIN for a specified number of rounds; after each round of training, inputting the verification samples to the MSIN for testing, and saving the MSIN which minimizes thetotal loss function of the MSIN as the final network. Step (3): taking a plurality of test samples from different sample domains as input values of the multi-sample reasoning neural network MSIN, inputting the input values into the trained multi-sample reasoning neural network, and outputting the sample categories corresponding to the test samples or the sample domains in which the test samples are located.

The invention relates to the technical field of image recognition, and particularly discloses a vehicle-mounted liquid crystal screen light guide plate defect visual detection method based on a target detection network, the method comprises the following steps: collecting a light guide plate image; preprocessing the image; establishing and training a first-stage target detection network, wherein the first-stage target detection network comprises a trunk feature extraction sub-network, a feature enhancement sub-network and a classification and regression sub-network; and defect detection: obtaining four effective feature layers through an image pyramid of the feature fusion sub-network by using a feature fusion mode, transmitting the four effective feature layers through the classification and regression sub-network to obtain a prediction result of the defect of the light guide plate, and displaying the prediction result in an upper computer. According to the method, the problems of imbalance of positive and negative samples, small defect detection rate, detection efficiency and the like are solved, a defect positioning task and defect existence classification are completed at the same time, a visual result is provided at the same time, and industrial application is achieved and realized.

The invention relates to the technical field of artificial intelligence, and particularly relates to a speech enhancement method and device based on a convolutional neural network, equipment and a medium. The speech enhancement method based on the convolutional neural network comprises the following steps: acquiring a time domain oscillogram of speech to be denoised and a speech enhancement model, wherein the speech enhancement model comprises a Gabor convolution layer, a simple recursion layer, a feature masking layer and a deconvolution layer which are connected in sequence; carrying out Gabor transformation on the time domain oscillogram through a complex filter, and extracting Gabor transformation features; inputting the Gabor transformation features into a simple recursion layer for prediction so as to determine a masking vector corresponding to a feature masking layer; filtering the Gabor transformation features according to the masking vector through the feature masking layer to obtain denoised Gabor transformation features; and restoring the denoised Gabor transformation features through a deconvolution layer to obtain a target denoised voice. According to the speech enhancement method based on the convolutional neural network, the model calculation efficiency and accuracy can be effectively improved.

The invention discloses a fully homomorphic encryption deep learning reasoning method and system based on an FPGA. The method comprises the steps of obtaining a ciphertext encrypted by a homomorphic encryption algorithm and a coded plaintext; obtaining the multiplication depth, the data processing scale and the network layer of the initial deep learning network; determining a value range of a coefficient module factor of the polynomial according to the term number of the plaintext polynomial, determining the number of values according to the multiplication depth, and determining a coefficient module according to the coefficient module factor selected by the error parameter; determining the weight and deviation of the network layer according to the item number and coefficient module of the polynomial and the data processing scale so as to obtain a packaging strategy of the network layer; according to the packaging strategy and the plaintext, selection of the item number and the coefficient module of the polynomial is judged, a network layer is optimized, an inference model is constructed, and therefore a ciphertext inference result is output to the ciphertext. An accelerator is integrally designed for the combination of homomorphic encryption and a deep learning network by using an FPGA, so that the reasoning speed of homomorphic encryption data on the deep learning network is increased.

The invention relates to a target detection algorithm based on a wheat image, and the detection algorithm comprises the following steps: 1, a first stage is a training stage of the algorithm, and 2, asecond stage is a real-time reasoning stage of the algorithm. The YOLOv5 algorithm after wheat data set optimization has high precision and high reasoning speed, real-time reasoning can be achieved,offline training of the YOLOv5 algorithm can also be completed on a single GPU (such as GTX1080ti or RTX2080Ti), and the requirement for hardware is greatly reduced.