-
公开(公告)号：US11941087B2
公开(公告)日：2024-03-26
申请号：US17165640
申请日：2021-02-02
发明人： Xiuming Yu , Wei Wang , Jing Xiao
IPC分类号： G06F16/00 , G06F7/24 , G06F18/213 , G06F18/214 , G06F18/22 , G06F18/2413 , G06F18/2431
CPC分类号： G06F18/24137 , G06F7/24 , G06F18/213 , G06F18/214 , G06F18/22 , G06F18/2431
摘要： Provided is an unbalanced sample data preprocessing method, which includes: a data acquisition request is received and initial data is acquired according to the data acquisition request, and the initial data is classified according to a preset classification rule to obtain first-class sample sets and second-class sample sets; characteristics of K first sample points extracted are analyzed to obtain a new data characteristic of the first-class sample sets; a new data label of the first-class sample sets is generated according to a first label corresponding to the first-class sample sets; a ratio between the number of first-class sample sets and the number of second-class sample sets is calculated; and new data of the first-class sample sets is generated according to the new data characteristic and the new data label, and the amount of new data is adjusted according to the ratio to increase the number of first-class sample sets.
-
2.发明授权公开(公告)号：US11487952B2
公开(公告)日：2022-11-01
申请号：US16637274
申请日：2019-06-26
摘要： The present disclosure relates to the technical field of natural language understanding, and provides a method, a terminal and a medium for generating a text based on a self-encoding neural network. The method includes: obtaining a text word vector and a classification requirement of a statement to be input; reversely inputting the text word vector into a trained self-encoding neural network model to obtain a hidden feature of an intermediate hidden layer of the self-encoding neural network model; modifying the hidden feature according to a preset classification scale and the classification requirement; defining the modified hidden feature as the intermediate hidden layer of the self-encoding neural network model, and reversely generating a word vector corresponding to an input layer of the self-encoding neural network model by the intermediate hidden layer; and generating the corresponding text, according to the generated word vector.
-
公开(公告)号：US20220343100A1
公开(公告)日：2022-10-27
申请号：US17238832
申请日：2021-04-23
发明人： Xinyi Wu , Yiwei Wang , Tian Xia , Peng Chang , Mei Han , Jing Xiao
摘要： A method for cutting or extracting video clips from a video, including the audio content relevant to points of particular interest, and combining the same for instruction or training on particular points; a computing device applying the method extracts text information from the spoken audio content of a video to be cut and obtains multiple paragraph segmentation positions as candidates for inclusion in a desired and finished presentation by analyzing the information from text representing the spoken audio content, the analysis being carried out by a semantic segmentation model. Candidate items of text are obtained by isolating pieces of text according to the paragraph segmentation positions. Time stamps of the candidate text segments are acquired, and candidate video clips are obtained by cutting the video according to the acquired time stamps.
-
4.发明授权公开(公告)号：US11386499B2
公开(公告)日：2022-07-12
申请号：US16084993
申请日：2017-09-30
发明人： Jianzong Wang , Chenyu Wang , Jin Ma , Jing Xiao
摘要： Disclosed are a car damage picture angle correction method, an electronic device, and a readable storage medium. The method includes: after receiving a car damage picture to be classified and identified, identifying a rotation category corresponding to the received car damage picture by using a pre-trained picture rotation category identification model; determining a rotation control parameter corresponding to the identified rotation category according to a pre-determined mapping relation between rotation categories and rotation control parameters, the rotation control parameter including a rotation angle and a rotation direction; and rotating the received car damage picture according to the determined rotation control parameter, so as to generate an angle-normal car damage picture. The disclosure can perform car damage picture angle correction more comprehensively and more effectively with no need to artificially perform angle identification on a car damage picture and to manually rotate the picture, thereby achieving a higher efficiency and accuracy.
-
5.发明申请公开(公告)号：US20210224567A1
公开(公告)日：2021-07-22
申请号：US16097291
申请日：2017-08-31
发明人： Jianzong Wang , Jin Ma , Zhangcheng Huang , Tianbo Wu , Jing Xiao
摘要： A deep learning based license plate identification method, device, equipment, and storage medium. The deep learning based license plate identification method comprises: extracting features of an original captured image by using a single shot multi-box detector to obtain a target license plate image; correcting the target license plate image to obtain a corrected license plate image; identifying the corrected license plate image by using a bi-directional long short-term memory model to obtain target license plate information. When the deep learning based license plate identification method performs license plate identification, the identification efficiency is high and the accuracy is higher.
-
公开(公告)号：US11062699B2
公开(公告)日：2021-07-13
申请号：US16348807
申请日：2017-08-31
发明人： Hao Liang , Jianzong Wang , Ning Cheng , Jing Xiao
IPC分类号： G10L15/00 , G10L15/20 , G10L17/00 , G10L15/14 , G06F17/16 , G06F17/18 , G06N3/08 , G10L15/16 , G10L15/02
摘要： A speech recognition method, comprising: acquiring speech data to be recognized; extracting a Filter Bank feature and a Mel-Frequency Cepstral Coefficient (MFCC) feature in the speech data; using the MFCC feature as input data of a trained Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) model, acquiring a first likelihood probability matrix outputted by the trained GMM-HMM model; using the Filter Bank feature as an input feature of a trained long short-term memory (LSTM) model which has a connection unit, acquiring a posterior probability matrix outputted by the LSTM model; using the posterior probability matrix and the first likelihood probability matrix as input data of a trained HMM model, acquiring an second likelihood probability matrix; and acquiring a target word sequence corresponding to the speech data to be recognized from a phoneme decoding network according to the second likelihood probability matrix.
-
公开(公告)号：US11030998B2
公开(公告)日：2021-06-08
申请号：US16097850
申请日：2017-08-31
发明人： Hao Liang , Jianzong Wang , Ning Cheng , Jing Xiao
摘要： An acoustic model training method, a speech recognition method, an apparatus, a device and a medium. The acoustic model training method comprises: performing feature extraction on a training speech signal to obtain an audio feature sequence; training the audio feature sequence by a phoneme mixed Gaussian Model-Hidden Markov Model to obtain a phoneme feature sequence; and training the phoneme feature sequence by a Deep Neural Net-Hidden Markov Model-sequence training model to obtain a target acoustic model. The acoustic model training method can effectively save time required for an acoustic model training, improve the training efficiency, and ensure the recognition efficiency.
-
8.发明申请公开(公告)号：US20210165970A1
公开(公告)日：2021-06-03
申请号：US16637274
申请日：2019-06-26
摘要： The present disclosure relates to the technical field of natural language understanding, and provides a method, a terminal and a medium for generating a text based on a self-encoding neural network. The method includes: obtaining a text word vector and a classification requirement of a statement to be input; reversely inputting the text word vector into a trained self-encoding neural network model to obtain a hidden feature of an intermediate hidden layer of the self-encoding neural network model; modifying the hidden feature according to a preset classification scale and the classification requirement; defining the modified hidden feature as the intermediate hidden layer of the self-encoding neural network model, and reversely generating a word vector corresponding to an input layer of the self-encoding neural network model by the intermediate hidden layer; and generating the corresponding text, according to the generated word vector.
-
9.发明申请公开(公告)号：US20200294509A1
公开(公告)日：2020-09-17
申请号：US16759384
申请日：2018-07-06
发明人： Yuanzhe Cai , Jianzong Wang , Ning Cheng , Jing Xiao
摘要： A method and apparatus for establishing a voiceprint model, a computer device, and a storage medium are described herein. The method includes: collecting speech acoustic features in a speech signal to form a plurality of cluster structures; calculating an average value and a standard deviation of the plurality of cluster structures and then performing coordinate transformation and activation function calculation to obtain a feature vector; and obtaining a voiceprint model based on the feature vector.
-
10.发明申请公开(公告)号：US20200294240A1
公开(公告)日：2020-09-17
申请号：US16759383
申请日：2018-07-13
发明人： Jianzong Wang , Chenyu Wang , Jin Ma , Jing Xiao
摘要： A method and apparatus for training a semantic segmentation model, a computer device, and a storage medium are described herein. The method includes: constructing a training sample set; inputting the training sample set into a deep network model for training; inputting the training sample set into a weight transfer function for training to obtain a bounding box prediction mask parameter; and constructing a semantic segmentation model.
-
-
-
-
-
-
-
-
-
搜索结果
55 条记录 (耗时 0.02 秒)
