CN108875592A - A kind of convolutional neural networks optimization method based on attention - Google Patents
A kind of convolutional neural networks optimization method based on attention Download PDFInfo
- Publication number
- CN108875592A CN108875592A CN201810519139.5A CN201810519139A CN108875592A CN 108875592 A CN108875592 A CN 108875592A CN 201810519139 A CN201810519139 A CN 201810519139A CN 108875592 A CN108875592 A CN 108875592A
- Authority
- CN
- China
- Prior art keywords
- data
- attention
- model
- feature
- noise
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The present invention is to provide a kind of convolutional neural networks optimization method based on attention model.The noise data of submarine target is segmented first, extracts its MFCC for every section of noise data, the purpose is to target noise data are become to the vector quantization data of fixed length.Then, the vector quantization data of obtained fixed length are spliced by the arrangement position and its sequential relationship of hydrophone in experimentation, it forms a complete period water and listens a gust feature, then listen a gust feature to change into corresponding picture to be input in trained network as input data set in the water of formation again.The present invention in the interpretation of result of service condition and modifies and optimizes to model to model by testing, and deep learning obtains the promotion of 10%-15% to Underwater Targets Recognition discrimination.
Description
Technical field
The present invention relates to a kind of Underwater targets recognitions.
Background technique
Underwater acoustic signal processing and sonar technique be one quickly grow, demand driving power is powerful, application prospect is extremely wide
Subject, be an important component part of modern sonar system and hydroacoustic electronic warfare, all the time by many scholars, engineering
The very big concern of technical staff.Various submarine targets all have its own distinctive noise characteristic, can be used for target identification, divide
Class, status monitoring etc. application field.Due to the complexity of marine environment and the particularity in underwater sound signal channel, to make an uproar from target
A kind of validity feature expression that can reflect target substantive characteristics and be able to satisfy underwater long-range detection requirement again is extracted in acoustical signal,
The always problem in this field.
Traditional Underwater Targets Recognition solution is mostly started with from the low-frequency information of submarine target, the DEMON based on target
The generic that analysis obtains target is carried out with LOFAR spectrum.
Summary of the invention
The purpose of the present invention is to provide one kind quickly to carry out dimensionality reduction operation to model to reduce over-fitting wind
The convolutional neural networks optimization method based on attention model of danger.
The object of the present invention is achieved like this：
(1) cutting of noise data
The cutting of the noise data is segmented as unit of 100ms to the noise data of submarine target；
(2) extraction of data characteristics
The extraction of the data characteristics is to extract its MFCC, the extraction process of MFCC such as formula for every section of noise data
For：
In formula, f is target noise frequency, and mel (f) is the mel cepstrum coefficients that shown target noise extracts；
(3) multichannel data splices
After submarine target noise data is segmented and carries out the extraction of data characteristics, in conjunction with the group battle array position of hydrophone, group
At corresponding vector matrix, meanwhile, the data of different time series data composition different batches, every section of noise data passes through previous step
Obtained MFCC is spliced by the exhaust position of multichannel hydrophone and the sequential relationship of every section of noise data, forms depth
Learning model input data set；
(4) pondization operation is accelerated based on attention model
Submarine target noise data is handled using convolution model, attention is defined as in the processing method of convolutional coding structure
Characteristic pattern in corresponding convolutional layer after convolution operation, the dimensionality reduction direction of characteristic pattern is determined by the method for principal component analysis, is being mentioned
It takes and carries out pond on the basis of the feature of convolution kernel；
(6) attention weighting connection
The feature for being extracted convolutional neural networks using convolution kernel feature, i.e. the last layer are as a result, be based on attention model
It is weighted processing.
After obtaining feature by convolution operation in convolutional neural networks, classification can be done using these features.It is theoretical
On say, directly can go to train classifier using all obtained features extracted, but if can directly face greatly if training
Calculation amount challenge.The inspiration that researcher obtains from biology has using the mankind in image recognition and classification " static
Property " attribute, aggregate statistics are carried out to the feature of different location, these summary statistics features not only have lower dimension, together
When can also prevent over-fitting.But the problem of traditional pondization strategy is that all pond strategies are fixed, i.e., pondization operates
It is only simple when participating in model training that dimensionality reduction rambunctiously is carried out to upper one layer of convolution results.The invention proposes a kind of combinations
The pond and splicing construction of convolutional neural networks and attention model are divided into interlayer attention pond model and towards full articulamentum
Merging features attention split-join model.Dimensionality reduction operation quickly can be carried out to reduce the over-fitting wind of model to model
Danger.
The invention proposes pond and the splicing constructions of a kind of combination convolutional neural networks and attention model, are divided into interlayer
The attention split-join model of attention pond model and the merging features towards full articulamentum.Interlayer attention pond model into
The characteristics of row pondization fully considers current convolution kernel when operating carries out effective pond according to the data characteristic of convolution kernel and characteristic pattern
Change, the operation of this pondization not only can the scale in raising pond by a relatively large margin can be with thus to the quick dimensionality reduction of characteristic pattern progress
Prevent the over-fitting easily caused on lesser data set using deep learning model.Meanwhile with convolution kernel and characteristic pattern
Data characteristic as attention pondization operation carry out the feature of image can be kept not lose during quick dimensionality reduction
It loses.
Traditional Underwater Targets Recognition solution is mostly started with from the low-frequency information of submarine target, the DEMON based on target
The generic that analysis obtains target is carried out with LOFAR spectrum.The present invention is using attention model to the pond of convolutional neural networks model
Layer optimizes, and carries out analysis identification using audio-frequency information of the convolutional neural networks to submarine target, is extracted using sound characteristic
Vectorization method of the method as model learning input data believes the vector quantization information processing of the audio data of generation at audio frequency
Input data of the thermodynamic chart of breath as model training.By testing to model in the interpretation of result of service condition and to model
It modifies and optimizes, deep learning obtains the promotion of 10%-15% to Underwater Targets Recognition discrimination.
The present invention is using the method for MFCC vector splicing multichannel hydrophone input and merely using MFCC and other classifiers
Classification results comparing result.Wherein list is apparently higher than using the method for MFCC vector joining method fusion multichannel hydrophone
The pure method that sound is described using MFCC.Compared to minimum recognition accuracy, present invention engagement MFCC vector used is spelled
The method connect improves nearly by 16.1% than the accuracy rate of conventional method.
Detailed description of the invention
Fig. 1 is the convolutional neural networks application framework based on attention model；
Fig. 2 is the interlayer pondization operation based on attention model；
Fig. 3 is the full connection layer operation based on attention model；
Fig. 4 is the convolutional neural networks structure based on attention model；
Fig. 5 is the convolution kernel size Comparative result based on attention model；
Fig. 6 is data set preprocessing process；
Fig. 7 is that data set extracts result example.
Specific embodiment
It illustrates below and the present invention is described in more detail.
Training dataset is pre-processed first.Spy directly is done to noise data using Mel-Frequency with previous
Sign extracts difference, and the present invention is first segmented the noise data of submarine target as unit of 100ms, for every section of noise number
According to its MFCC is extracted, the purpose is to target noise data are become to the vector quantization data of fixed length.Then, by the arrow of obtained fixed length
Quantized data is spliced by the arrangement position and its sequential relationship of hydrophone in experimentation, forms a complete period
Water listens a gust feature, then listens a gust feature to change into corresponding picture to be input to training net as input data set in the water of formation again
In network.
(1) cutting of noise data
The noise data obtained from experimental situation is institute after the multichannel hydrophone group battle array for being placed in different location different angle
The underwater target noise of measurement, by taking 16 road water listen battle array as an example, single channel hydrophone time of measuring 6min, frequency band 25.6kHz, sample rate
All signals for 65536Hz, acquisition are voltage value.Since the unsuitable convolutional neural networks of the data of single hydrophone are direct
It is handled, while considering that the architectural characteristic of convolutional network, the present invention carry out noise data acquired in multiple groups hydrophone
There is the processing being directed to.The present invention considers hydrophone and obtains the positional relationship of data and the sequential relationship of data itself, right first
Data are cut, and then opsition dependent is integrated with sequential relationship.Cutting unit will fully consider hydrophone characteristic.Hereafter divide
Singulation position is set to 100ms, to form the input data set of the Underwater Targets Recognition of convolutional coding structure.
(2) extraction of data characteristics (by taking MFCC as an example)
Present invention use can merge a variety of noise characteristic extracting methods, such as LPCC, PLP, hereafter by taking MFCC as an example,
Shown in the extraction process of MFCC such as formula (4), in specifically used this method, the acquisition of underwater noise used by considering is set
Standby acquisition characteristic carrys out setup parameter.By 512 frames of every section of noise point in this example, 31 triangular filters are utilized.The present invention is mentioning
While taking single order MFCC (16 filters), it is contemplated that the static characteristic that MFCC can only obtain noise is unable to get noise number
According to behavioral characteristics, while extracting single order MFCC result again extract target noise first-order difference MFCC (15 scales
Point), behavioral characteristics of the first-order difference MFCC as target noise, the extraction submarine target for allowing the present invention finer and smoother is made an uproar
The feature of sound data.It is combined finally by by all characteristic lines, to obtain one 128 for the noise data after every section of cutting
The feature vector of × 1 dimension.
In formula, f is target noise frequency, and mel (f) is the mel cepstrum coefficients that shown target noise extracts.
(3) multichannel data splices
Data splicing is the significant process that the present invention combines actual experiment, and actually data splicing is exactly one by hydrophone
Process of the position in conjunction with target noise sequential relationship.Specific feature extracting method will be pressed after the segmentation of submarine target noise data
Afterwards, in conjunction with the group battle array position of specific hydrophone, corresponding vector matrix is formed, meanwhile, different batches of different time series data composition
Secondary data, such as data of all hydrophones of first group of 100ms form batch (Batch) data, second group of 100ms's
All hydrophone data form the data of another Batch.Every section of noise data is by previous step by obtained MFCC by more
The sequential relationship of the exhaust position of road hydrophone and every section of noise data is spliced, to form deep learning mode input
Data set.
(4) pondization operation is accelerated based on attention model
The reduction process that the present invention accelerates pondization to operate using attention model.Attention mechanism, which will be one, to be encoded
The method that device-decoder architecture is freed from the internal attribute of regular length.The attention operation that the present invention defines refers to
It is the operation trend by keep model in step before in current operation.The present invention is handled underwater using convolution model
Target noise data, attention is defined as the feature in corresponding convolutional layer after convolution operation in the processing method of convolutional coding structure
Figure, the dimensionality reduction direction of characteristic pattern is determined by the method for principal component analysis, to provide effective finger for adjacent pondization operation
It leads.The present invention fully considers the effect of upper one layer of convolution kernel carrying out Chi Huashi, on the basis of extracting the feature of convolution kernel into
Row pond.
(5) attention weighting connection
The feature that the present invention utilizes convolution kernel feature to extract convolutional neural networks, i.e. the last layer result.Convolutional Neural
The full articulamentum of network plays the role of " classifier " in entire convolutional neural networks.If convolutional layer, pond layer and swash
The operations such as function layer living are if initial data to be mapped to hidden layer feature space, and full articulamentum then plays the " distribution that will be acquired
Formula character representation " is mapped to the effect in sample labeling space.The apish pathways for vision of convolutional layer extracts feature, full articulamentum one
As be responsible for classification or return, since full articulamentum can lose some feature locations information.The core operation connected entirely is exactly square
Battle array vector product, essence is exactly by a feature space linear transformation to another feature space.Any dimension of object space is all
Think to will receive source space per one-dimensional influence.The feature that the present invention utilizes convolution kernel feature to extract convolutional neural networks,
That is the last layer is as a result, be weighted processing based on attention model, to the characteristics of not only having considered process of convolution but also remain spy
The location information of sign.
Main feature and content of the invention is as follows：
(1) pondization operation is accelerated based on attention model
Encoder-Decoder structure shows advanced level in multiple fields, but this structure indicates list entries
For the internal representation of regular length.The length of list entries is limited, also causes model to the performance of especially long list entries
It is deteriorated.Focus can be freed from preceding n fixed sequence program using attention model, to accomplish to be concerned about desired pass
N preamble sequence of note.Attention mechanism is one and liberates coder-decoder structure from the internal attribute of regular length
Method out.By keeping model to the intermediate output of each step in list entries treatment process as a result, training pattern learns
Input how is selectively paid close attention to, and it is connected with the item in output sequence.People when carrying out observation image,
In fact it is not that once just each position pixel of entire image has been seen, is to focus onto figure according to demand mostly
The specific part of picture.And the image study that the mankind can observe before will observe what image attention power should be concentrated to following
Position, the reduction process that the present invention accelerates pondization to operate using attention model.
(2) full articulamentum (fully connected layers, FC) plays " classification in entire convolutional neural networks
The effect of device ".If the operations such as convolutional layer, pond layer and activation primitive layer are that initial data is mapped to hidden layer feature space
If, " the distributed nature expression " that full articulamentum then plays the role of to acquire is mapped to sample labeling space.Actually make
In, full articulamentum can be realized by 1 × 1 convolution operation to reach quick calculated result, meanwhile, 1 × 1 convolution kernel can play one
A effect across channel polymerization plays the purpose for reducing parameter so may further play the role of dimensionality reduction (or rising dimension).
The apish pathways for vision of convolutional layer extracts feature, and full articulamentum is generally responsible for classifying or return, since full articulamentum can be lost
Lose some feature locations information.
The core operation connected entirely is exactly matrix-vector product, and essence is exactly by a feature space linear transformation to another
A feature space.Any dimension of object space all thinks to will receive every one-dimensional influence of source space.The present invention utilizes convolution kernel
The feature that feature extracts convolutional neural networks, i.e. the last layer are as a result, be weighted processing based on attention model, thus both
The characteristics of considering process of convolution remains the location information of feature again.
Accelerate pondization operation for based on attention model, calculation method is as follows.
Model uses AoCLIt indicates, shown in calculation method such as formula (1)
In formula, LiI-th of pond of L layer is indicated as a result, eigVector (ki) indicate kthiA extracted spy of convolution kernel
Levy vector, Area (ki) indicate kthiThe region that a convolution kernel is covered.
IoCiThe attention model influenced based on convolution kernel is represented, then shown in its calculation method such as formula (2).
In formula, wiFor a weight matrix, specific gravity of every one-dimensional characteristic in entire characteristic pattern is recorded,To connect entirely
Model treatment before layer is as a result, m is the number of convolution kernel.
wiCalculating be calculated with multilayer perceptron model, shown in method such as formula (3).
wi=f (ki,yi) (4)
In formula, f () indicates a feedforward neural network.Feedforward neural network is input with convolution kernel, with kiRepresent it
Corresponding convolution kernel, yiIndicate label corresponding to this feature.
Claims (1)
1. a kind of convolutional neural networks optimization method based on attention model, it is characterized in that：
(1) cutting of noise data
The cutting of the noise data is segmented as unit of 100ms to the noise data of submarine target；
(2) extraction of data characteristics
The extraction of the data characteristics is to extract its MFCC for every section of noise data, and the extraction process of MFCC such as formula is：
In formula, f is target noise frequency, and mel (f) is the mel cepstrum coefficients that shown target noise extracts；
(3) multichannel data splices
After submarine target noise data is segmented and carries out the extraction of data characteristics, in conjunction with the group battle array position of hydrophone, phase is formed
The vector matrix answered, meanwhile, the data of different time series data composition different batches, every section of noise data is obtained by previous step
MFCC spliced by the exhaust position of multichannel hydrophone and the sequential relationship of every section of noise data, formed deep learning mould
Type input data set；
(4) pondization operation is accelerated based on attention model
Submarine target noise data is handled using convolution model, attention is defined as accordingly in the processing method of convolutional coding structure
Characteristic pattern in convolutional layer after convolution operation determines the dimensionality reduction direction of characteristic pattern by the method for principal component analysis, rolls up extracting
Pond is carried out on the basis of the feature of product core；
(6) attention weighting connection
The feature for being extracted convolutional neural networks using convolution kernel feature, i.e. the last layer based on attention model as a result, carried out
Weighting processing.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810330123 | 2018-04-13 | ||
| CN201810330123X | 2018-04-13 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108875592A true CN108875592A (en) | 2018-11-23 |
Family
ID=64335012
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810519139.5A Pending CN108875592A (en) | 2018-04-13 | 2018-05-28 | A kind of convolutional neural networks optimization method based on attention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108875592A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109299396A (en) * | 2018-11-28 | 2019-02-01 | 东北师范大学 | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model |
| CN109344920A (en) * | 2018-12-14 | 2019-02-15 | 汇纳科技股份有限公司 | Customer attributes prediction technique, storage medium, system and equipment |
| CN109376804A (en) * | 2018-12-19 | 2019-02-22 | 中国地质大学（武汉） | Based on attention mechanism and convolutional neural networks Classification of hyperspectral remote sensing image method |
| CN109902399A (en) * | 2019-03-01 | 2019-06-18 | 哈尔滨理工大学 | Rolling bearing fault recognition methods under a kind of variable working condition based on ATT-CNN |
| CN110110796A (en) * | 2019-05-13 | 2019-08-09 | 哈尔滨工程大学 | A kind of analysis method of the marine ships time series data based on deep learning |
| CN110222574A (en) * | 2019-05-07 | 2019-09-10 | 杭州智尚云科信息技术有限公司 | Production operation Activity recognition method, apparatus, equipment, system and storage medium based on structuring double fluid convolutional neural networks |
| CN110490892A (en) * | 2019-07-03 | 2019-11-22 | 中山大学 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
| CN111401548A (en) * | 2020-03-03 | 2020-07-10 | 西北工业大学 | L off line spectrum detection method based on deep learning |
| CN111652041A (en) * | 2020-04-14 | 2020-09-11 | 河北地质大学 | Hyperspectral band selection method, device and apparatus based on depth subspace clustering |
| CN113222899A (en) * | 2021-04-15 | 2021-08-06 | 浙江大学 | Method for detecting, segmenting and classifying liver tumors through CT (computed tomography) based on deep learning |
| CN113888430A (en) * | 2021-09-30 | 2022-01-04 | 北京达佳互联信息技术有限公司 | Image processing method and device and model training method and device |
| US11681914B2 (en) | 2020-05-08 | 2023-06-20 | International Business Machines Corporation | Determining multivariate time series data dependencies |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682574A (en) * | 2016-11-18 | 2017-05-17 | 哈尔滨工程大学 | One-dimensional deep convolution network underwater multi-target recognition method |
| CN107180247A (en) * | 2017-05-19 | 2017-09-19 | 中国人民解放军国防科学技术大学 | Relation grader and its method based on selective attention convolutional neural networks |
-
2018
- 2018-05-28 CN CN201810519139.5A patent/CN108875592A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682574A (en) * | 2016-11-18 | 2017-05-17 | 哈尔滨工程大学 | One-dimensional deep convolution network underwater multi-target recognition method |
| CN107180247A (en) * | 2017-05-19 | 2017-09-19 | 中国人民解放军国防科学技术大学 | Relation grader and its method based on selective attention convolutional neural networks |
Non-Patent Citations (3)
| Title |
|---|
| MD. MONIRUZZAMAN等: "Deep Learning on Underwater Marine Object Detection: A Survey", 《ADVANCED CONCEPTS FOR INTELLIGENT VISION SYSTEMS》 * |
| 梁斌等: "基于多注意力卷积神经网络的特定目标情感分析", 《计算机研究与发展》 * |
| 王念滨等: "基于卷积神经网络的水下目标特征提取方法", 《系统工程与电子技术》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109299396B (en) * | 2018-11-28 | 2020-11-06 | 东北师范大学 | Convolutional neural network collaborative filtering recommendation method and system fusing attention model |
| CN109299396A (en) * | 2018-11-28 | 2019-02-01 | 东北师范大学 | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model |
| CN109344920A (en) * | 2018-12-14 | 2019-02-15 | 汇纳科技股份有限公司 | Customer attributes prediction technique, storage medium, system and equipment |
| CN109344920B (en) * | 2018-12-14 | 2021-02-02 | 汇纳科技股份有限公司 | Customer attribute prediction method, storage medium, system and device |
| CN109376804A (en) * | 2018-12-19 | 2019-02-22 | 中国地质大学（武汉） | Based on attention mechanism and convolutional neural networks Classification of hyperspectral remote sensing image method |
| CN109376804B (en) * | 2018-12-19 | 2020-10-30 | 中国地质大学（武汉） | Hyperspectral remote sensing image classification method based on attention mechanism and convolutional neural network |
| CN109902399A (en) * | 2019-03-01 | 2019-06-18 | 哈尔滨理工大学 | Rolling bearing fault recognition methods under a kind of variable working condition based on ATT-CNN |
| CN109902399B (en) * | 2019-03-01 | 2022-07-19 | 哈尔滨理工大学 | Rolling bearing fault identification method under variable working condition based on ATT-CNN |
| CN110222574A (en) * | 2019-05-07 | 2019-09-10 | 杭州智尚云科信息技术有限公司 | Production operation Activity recognition method, apparatus, equipment, system and storage medium based on structuring double fluid convolutional neural networks |
| CN110222574B (en) * | 2019-05-07 | 2021-06-29 | 杭州智尚云科信息技术有限公司 | Production operation behavior identification method, device, equipment and system based on structured double-current convolutional neural network and storage medium |
| CN110110796A (en) * | 2019-05-13 | 2019-08-09 | 哈尔滨工程大学 | A kind of analysis method of the marine ships time series data based on deep learning |
| CN110110796B (en) * | 2019-05-13 | 2020-12-18 | 哈尔滨工程大学 | Deep learning-based marine vessel time sequence data analysis method |
| CN110490892A (en) * | 2019-07-03 | 2019-11-22 | 中山大学 | A kind of Thyroid ultrasound image tubercle automatic positioning recognition methods based on USFaster R-CNN |
| CN111401548A (en) * | 2020-03-03 | 2020-07-10 | 西北工业大学 | L off line spectrum detection method based on deep learning |
| CN111401548B (en) * | 2020-03-03 | 2022-03-22 | 西北工业大学 | Lofar line spectrum detection method based on deep learning |
| CN111652041A (en) * | 2020-04-14 | 2020-09-11 | 河北地质大学 | Hyperspectral band selection method, device and apparatus based on depth subspace clustering |
| US11681914B2 (en) | 2020-05-08 | 2023-06-20 | International Business Machines Corporation | Determining multivariate time series data dependencies |
| CN113222899A (en) * | 2021-04-15 | 2021-08-06 | 浙江大学 | Method for detecting, segmenting and classifying liver tumors through CT (computed tomography) based on deep learning |
| CN113888430A (en) * | 2021-09-30 | 2022-01-04 | 北京达佳互联信息技术有限公司 | Image processing method and device and model training method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108875592A (en) | A kind of convolutional neural networks optimization method based on attention | |
| CN110807365B (en) | Underwater target identification method based on fusion of GRU and one-dimensional CNN neural network | |
| Doan et al. | Underwater acoustic target classification based on dense convolutional neural network | |
| Peng et al. | Efficient speech emotion recognition using multi-scale cnn and attention | |
| Mielke et al. | A method for automated individual, species and call type recognition in free-ranging animals | |
| CN109473120A (en) | A kind of abnormal sound signal recognition method based on convolutional neural networks | |
| CN109493874A (en) | A kind of live pig cough sound recognition methods based on convolutional neural networks | |
| CN109800700A (en) | A kind of underwater sound signal target classification identification method based on deep learning | |
| CN106782511A (en) | Amendment linear depth autoencoder network audio recognition method | |
| CN109637522A (en) | A kind of speech-emotion recognition method extracting deep space attention characteristics based on sound spectrograph | |
| CN110148408A (en) | A kind of Chinese speech recognition method based on depth residual error | |
| CN110111797A (en) | Method for distinguishing speek person based on Gauss super vector and deep neural network | |
| CN108520753A (en) | Voice lie detection method based on the two-way length of convolution memory network in short-term | |
| Ibrahim et al. | Transfer learning for efficient classification of grouper sound | |
| CN110287770A (en) | Individual goal matching and recognition method in a kind of water based on convolutional neural networks | |
| Sun et al. | Underwater single-channel acoustic signal multitarget recognition using convolutional neural networks | |
| CN109559755A (en) | A kind of sound enhancement method based on DNN noise classification | |
| Wei et al. | A method of underwater acoustic signal classification based on deep neural network | |
| Schröter et al. | Segmentation, classification, and visualization of orca calls using deep learning | |
| CN115294994A (en) | Bird sound automatic identification system in real environment | |
| CN115393933A (en) | Video face emotion recognition method based on frame attention mechanism | |
| Buchanan et al. | Deep convolutional neural networks for detecting dolphin echolocation clicks | |
| Ng et al. | Teacher-student training for text-independent speaker recognition | |
| Zhang et al. | Spectrogram transformers for audio classification | |
| Tang et al. | Transound: Hyper-head attention transformer for birds sound recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |