CN109376576A - The object detection method for training network from zero based on the intensive connection of alternately update - Google Patents
The object detection method for training network from zero based on the intensive connection of alternately update Download PDFInfo
- Publication number
- CN109376576A CN109376576A CN201810951609.5A CN201810951609A CN109376576A CN 109376576 A CN109376576 A CN 109376576A CN 201810951609 A CN201810951609 A CN 201810951609A CN 109376576 A CN109376576 A CN 109376576A
- Authority
- CN
- China
- Prior art keywords
- attention
- training
- module
- longitudinal
- object detection
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of based on object detection method of the intensive connection from zero training network is alternately updated, and collection target image first makes data set, and is labeled；Carry out intensively being connected to again alternately update module, bounded can deformation convolution module, decouple based on the attention power module, spatial scaling feature pyramid module and variable segment that combine from the transverse and longitudinal of attention and channel attention the training of convolution module, obtain training pattern；Training image is obtained, data set is obtained, detection identification finally is carried out to target image using data set and training pattern.The present invention will also promote to realize that deformation, posture, dimensional variation is big and has the high-precision identification for blocking underwater movement objective efficiently from the development of zero training study mechanism from distorted movement target detection and follow under deep learning angle auxiliary water.
Description
Technical field
The present invention relates to the alternatings being oriented to based on attention to update underwater distorted movement of the intensive connection from zero training network
Object detection method belongs to Intelligent Information Processing and object detection and recognition technical field.
Background technique
Object detection and recognition is link important in vision system, and target detection technique is in video monitoring, intelligence machine
The fields such as people's navigation, automatic Pilot, gesture recognition, Shape-memory behavior have broad application prospects.Object detection and recognition is to close
Reason utilizes and protection marine resources, permanently effective multi-angle monitoring ocean, is also cultivation fishery, marine fishing and Fish behavior
The offers basic data and information support such as analysis.However for underwater distorted movement target, since marine optics is at slice
Part restricts, and vulnerable to illumination, visual angle, block, the factors such as form and dimensional variation influence so that pole occurs for the external appearance characteristic of target
Big variation, to bring great challenge to image detection and identification.
In recent years, target detection achieves quantum jump, has benefited from deep learning --- mainly convolutional neural networks
(Convolution Neural Network, CNN) and candidate region (Region Proposal) algorithm.Target detection and knowledge
Other main stream approach includes: conventional target detection, deep learning target detection based on Region Proposal and based on returning
The deep learning target detection of method.State-of-the-art target detection identification network is very dependent on this kind of big in Imagenet at present
Classification task is trained to obtain feature extraction network and then training objective detection identification on type categorized data set in advance, but due to classification
It will lead to study deviation with difference of the detection identification on loss function and classification, model fine tuning can alleviate this deviation but not
This deviation can be fundamentally solved, and pre-training model is moved into the bigger detection identification neck of difference domain from classification task
Domain is more difficult.The DSOD method proposed gives a very good solution thinking within 2017, utilizes the outstanding ladder of DenseNet
Degree transmission mechanism can not depend on pre-training disaggregated model start from scratch training detection identification network.
Target detection is extremely important a part in computer vision field, blocks deformation, background between object
Complexity, illumination variation, dimensional variation etc. are urgent problems to be solved in detection process.It is i.e. existing in the prior art mainly to ask
Topic: (1) since marine optics image-forming condition restrict, and vulnerable to illumination, visual angle, block, the factors such as form and dimensional variation influence,
So that the problem of greatly variation, occurs for the external appearance characteristic of target；(2) since image object lacks contextual information, in image object
In the case where partial occlusion or deformation occurs, the problem of will lead to target detection mistake.
Summary of the invention
In view of the above-mentioned problems, the object of the present invention is to provide a kind of alternating based on attention guiding update intensive connection from
The underwater distorted movement object detection and recognition method of zero training network, to realize for distorted movement target's feature-extraction, spy
Sign refining, the unification of attention, it is intended to quickly excavate and detect in the underwater observation data low from magnanimity, high speed, value density
Distorted movement target, to make up the deficiencies in the prior art.
In order to achieve the above objectives, the present invention adopts the following technical scheme:
A kind of underwater distorted movement target inspection that the intensive connection of alternating update based on attention guiding trains network from zero
Survey and recognition methods, comprising the following steps:
(1) it collects target image and makes data set, and be labeled；
(2) it intensive connection alternating update module: in order to not depend on classification pre-training model from zero training, avoids classifying and examine
Study deviation and data set cross-border issue caused by difference of the identification on loss function and classification are surveyed, especially for from waterborne
Data set is transitioned into the problem of underwater data collection, using intensive connection alternating update module, relies on its outstanding gradient conveyer
System, efficient feature extraction and feature refining effect, included attention effect realize feature extraction, feature refining, pay attention to
The unification of power；
(3) bounded can deformation convolution module: in order to overcome the detection of deformation target to identify difficult problem, handing in intensive connection
For use after update module bounded can deformation convolution module, can deformation convolution the limitation of receptive field is not present, there is study sense
Target area can be effectively paid close attention to the modified receptive field of the variation of input picture by the effect of open country offset, and
Can deformation convolution there is very strong adaptive faculty for target deformation, the detection effect of network will not restrict by target deformation；
(4) based on the attention power module combined from the transverse and longitudinal of attention and channel attention: in order to preferably optimize transmitting
The intensive connection feature that alternately update module is extracted, using based on the attention combined from the transverse and longitudinal of attention and channel attention
Module is laterally intended to pay close attention to the relationship between area-of-interest and hard objectives, is longitudinally intended to pay close attention to the important of different channel characteristics
Property, reinforce important feature, weaken inessential feature, be characterized extraction module and convey superior feature, is drawn using attention force characteristic
Lead feature transmission, dominant carry out characteristic optimization；
(5) spatial scaling feature pyramid module: accuracy of identification is detected to improve, in network header feature pyramid module
It is middle to replace up-sampling using spatial scaling, the fusion between different resolution feature is carried out under the premise of not destructive characteristics, together
When greatly reduce the calculation amount of network header；
(6) variable segment decouples convolution: to solve the problems, such as that the detection from zero training identifies that network convergence is slow and common
Convolution is difficult to distinguish the problem of difference and class inherited in class, using the amplitude and angle of decoupling convolution, uses amplitude measure class
Interior difference, angle measure class inherited, and direct study amplitude and angle is taken to replace being fitted；
(7) training image is obtained, and carries out scale amplification to it and hide at random to obtain { In, n=1,2 ..., N }；
(8) the target detection identification of model training: { I is usedn, n=1,2 ..., N and trained model to target
Image carries out detection identification.
Further, the intensive connection alternating update module in the step (2):
X expression characteristic pattern, k represent the number alternately updated, and i represents i-th layer in certain alternating update, and W is represented wait instruct
Experienced weight, * represent convolution operation, and g represents nonlinear function, m, and l is the index of summation sign.
Further, the bounded in the step (3) can deformation convolution module:
X indicates that characteristic pattern, p indicate the p of characteristic pattern X0Position and p=p after the offset of position receptive field0+ Δ p, wherein
Δ p is the variable that can learn, and q indicates the position for four integral points being located at around p, and G (q, p) is the interpolation for asking p and q, is passed through
The mode of interpolation finds out the characteristic pattern after offset.
Further, in the step (4) based on the attention mould combined from the transverse and longitudinal of attention and channel attention
Block, wherein lateral attention:
F (x)=Wfx
G (x)=Wgx
yi=γ oi
Wherein, x is the characteristic pattern of input, and W is weight to be trained, and f, g, h is respectively three kinds for laterally paying attention to power module
Feature extraction mode, γ are the significance level of lateral attention, are a trainable variable, yiLaterally to pay attention to power module
Output；
Longitudinal attention are as follows:
S=σ (W2δ(W1Z))
Wherein, Z is the feature in global average pond, and H, W are characterized the ranks size of figure, and U is to input longitudinal attention
Characteristic pattern, σ, δ are nonlinear function, and W is weight to be trained, and s is longitudinal output for paying attention to power module；
Finally, lateral attention and longitudinal attention are merged:
Y=(X+yi)*(1+s)
Wherein, X is the input that transverse and longitudinal pays attention to power module, and Y is the output that transverse and longitudinal pays attention to power module.
Further, the spatial scaling feature pyramid module in the step (5):
Wherein I is characterized figure, and before LR is spatial scaling, after SR is spatial scaling, r be the step-length of conversion, x, y, and c divides
Not Wei column coordinate, row coordinate, depth coordinate.
Further, the variable segment in the step (6) decouples convolution:
Wherein, x is the characteristic pattern of input, and w is weight to be trained, and β, ρ are that can train variable, and E is to seek desired value.
Further, the size 7 for the characteristic pattern that the multiple dimensioned training method in the step (7) is finally obtained using network
~10 times be used as network inputs, this is conducive to network for the target detection precision of different scale, at the same to the pixel of image into
Row blocks at random, and the detection accuracy of identification of network can be improved in this in the training process, promote the entirety of network attention target and
It is not a certain position.
Beneficial effects of the present invention:
The present invention from distorted movement target detection and follow under deep learning angle auxiliary water, also will promote efficiently from
The development of zero training study mechanism realizes that deformation, posture, dimensional variation is big and has the high-precision for blocking underwater movement objective to know
Not.It can be inclined to avoid classifying and detecting study caused by difference of the identification on loss function and classification from zero training study mechanism
Difference and data set cross-border issue, and the intensive connection alternating update module proposed also achieves feature extraction, feature refines,
The unification of attention.Can deformation convolution sum transverse and longitudinal pay attention to power module effectively overcome deformation target detection know
Not, while using the transmission of attention guide features, play the role of further refining feature.Spatial scaling feature pyramid module
The fusion of different resolution feature can be carried out under the premise of not destructive characteristics, improve the precision of detection identification.Variable segment
Decoupling convolution can solve that slow from the detection identification network convergence of zero training and common convolution is difficult to distinguish difference and class in class
Between difference the problem of.Finally, multiple dimensioned and hiding at random training method can be improved network for target scale and block
Robustness.
Detailed description of the invention
Fig. 1 is overall flow figure of the invention.
Fig. 2 is that marine environment is to be detected in embodiment 1 and identifies image.
Fig. 3 is intensive connection of the invention alternately update module figure.
Fig. 4 is that bounded of the invention can deformation convolution module figure.
Fig. 5 is the attention module map of the invention combined from the transverse and longitudinal of attention and channel attention.
Fig. 6 is spatial scaling feature pyramid module map of the invention.
Fig. 7 is multiple dimensioned and random hiding training method schematic diagram of the invention.
Fig. 8 is the detection recognition result figure in example 1.
Specific embodiment
To keep the purpose of the present invention, embodiment and advantage relatively sharp, with reference to the accompanying drawing and pass through specific embodiment
It is next that present invention be described in more detail.
Embodiment 1: being detection identification object with the dynamic of ocean underwater environment Mesichthyes.
The specific flow chart of the present embodiment is as shown in Figure 1.
Specifically using one section as shown in Fig. 2, under the marine environment that Shandong Province's aquafarm is shot in the present embodiment
In the daytime fish movement video (1920*1080 pixel, 25 frame per second) is as to be detected and identification video.
Following steps should be described in detail in conjunction with attached drawing and concrete outcome, and should be general in summary of the invention
The step of condition.
Step 1: a large amount of fish image making data sets that will be taken in aquafarm, mark all fishes in image
Position, type；
Step 2: as shown in figure 3, alternately update module realizes feature extraction, feature refining, attention by intensively connection
Unification:
X indicate characteristic pattern, k represent alternately update number, i represent certain time alternately update in i-th layer, W represent to
Trained weight, * represent convolution operation, and g represents nonlinear function.
Step 3: in order to overcome the detection of deformation target to identify difficult problem, it can deformation using bounded after step 2
Convolution module, as shown in figure 4, can deformation convolution be not present receptive field limitation, have the function of learn receptive field deviate, with
The modified receptive field of the variation of input picture, can effectively pay close attention to target area, and can deformation convolution for target
Deformation has very strong adaptive faculty, and the detection effect of network will not be restricted by target deformation, and formula is as follows:
X indicates that characteristic pattern, p indicate the p of characteristic pattern X0Position and p=p after the offset of position receptive field0+ Δ p, wherein
Δ p is the variable that can learn, and q indicates the position for four integral points being located at around p, and G (q, p) is the interpolation for asking p and q, is passed through
The mode of interpolation finds out the characteristic pattern after offset.
Step 4: used after step 3 based on the attention power module combined from the transverse and longitudinal of attention and channel attention,
As shown in figure 5, preferably to optimize the intensive connection feature that alternately update module is extracted of transmitting, wherein lateral attention:
F (x)=Wfx
G (x)=Wgx
yi=γ oi
Wherein x is the characteristic pattern of input, and W is weight to be trained, and f, g, h is respectively three kinds of spies for laterally paying attention to power module
Extracting mode is levied, it is a trainable variable, y that γ, which is the significance level of lateral attention,iLaterally to pay attention to the defeated of power module
Out.
Longitudinal attention are as follows:
S=σ (W2δ(W1Z))
Wherein, Z is the feature in global average pond, and H, W are characterized the ranks size of figure, and U is to input longitudinal attention
Characteristic pattern, σ, δ are nonlinear function, and W is weight to be trained, and s is longitudinal output for paying attention to power module.
Finally, lateral attention and longitudinal attention are merged:
Y=X+s*X+yi
Wherein, X is the input that transverse and longitudinal pays attention to power module, and Y is the output that transverse and longitudinal pays attention to power module.
Step 5: three times by Module cycle of the step 2 into step 4, then using spatial scaling feature pyramid mould
Block improves detection accuracy of identification, replaces up-sampling using spatial scaling in the network header feature pyramid module, is not breaking
The fusion between different resolution feature is carried out under the premise of bad feature, while greatly reducing the calculation amount of network header.
Spatial scaling feature pyramid module, as shown in Figure 6:
Wherein I is characterized figure, and before LR is spatial scaling, after SR is spatial scaling, r be the step-length of conversion, x, y, and c divides
Not Wei column coordinate, row coordinate, depth coordinate.
Note: all convolution are all the variable segment decoupling convolution used in the present invention, are able to solve the detection from zero training
The identification slow problem of network convergence and common convolution are difficult to distinguish the problem of difference and class inherited in class, use decoupling convolution
Amplitude and angle, using difference in amplitude measure class, angle measures class inherited, direct study amplitude and angle to replace
It is fitted.
Step 6: obtaining training image, and carry out scale amplification to it and hide at random to obtain { In, n=1,2 ...,
N }, as shown in Figure 7；
Step 7: the target detection of model training identifies: using { In, n=1,2 ..., N } to what is be oriented to based on attention
It alternately updates intensive connection to be trained from zero training network to obtain training pattern, using trained model to target image
Carry out detection identification.
Detection, identification and statistics result is shown in Fig. 8, the interior fish to be identified of rectangle frame, and carries out above rectangle frame
The display of target category and confidence level result, verified detection recognition result is essentially identical with legitimate reading, that is, illustrates this hair
It is bright that the feasibility and high accuracy of detection method are provided.
Claims (7)
1. a kind of object detection method for training network from zero based on the intensive connection of alternately update, which is characterized in that including following
Step:
(1) it collects target image and makes data set, and be labeled；
(2) intensive connection alternating update module；
(3) bounded can deformation convolution module；
(4) based on the attention power module combined from the transverse and longitudinal of attention and channel attention；
(5) spatial scaling feature pyramid module；
(6) variable segment decouples convolution；
(7) training image is obtained, data set { I is obtainedn, n=1,2 ..., N }, obtain training pattern；
(8) the target detection identification of model training: { I is usedn, n=1,2 ..., N and training pattern to target image carry out
Detection identification.
2. object detection method as described in claim 1, which is characterized in that the intensive connection alternating in the step (2) is more
New module:
X expression characteristic pattern, k represent the number alternately updated, and i represents i-th layer in certain alternating update, and W represents to be trained
Weight, * represent convolution operation, and g represents nonlinear function, m, and l is the index of summation sign.
3. object detection method as described in claim 1, which is characterized in that the bounded in the step (3) can deformation convolution
Module:
X indicates that characteristic pattern, p indicate the p of characteristic pattern X0Position and p=p after the offset of position receptive field0+ Δ p, wherein Δ p be
The variable that can learn, q indicate the position for four integral points being located at around p, and G (q, p) is the interpolation for asking p and q, passes through interpolation
Mode finds out the characteristic pattern after offset.
4. object detection method as described in claim 1, which is characterized in that in the step (4) based on from attention and
The attention power module that the transverse and longitudinal of channel attention combines, wherein lateral attention:
F (x)=Wfx
G (x)=Wgx
yi=γ oi
Wherein, x is the characteristic pattern of input, and W is weight to be trained, and f, g, h is respectively three kinds of features for laterally paying attention to power module
Extracting mode, γ are the significance level of lateral attention, are a trainable variable, yiLaterally to pay attention to the defeated of power module
Out；
Longitudinal attention are as follows:
S=σ (W2δ(W1Z))
Wherein, Z is the feature in global average pond, and H, W are characterized the ranks size of figure, and U is the feature for inputting longitudinal attention
Figure, σ, δ are nonlinear function, and W is weight to be trained, and s is longitudinal output for paying attention to power module；
Finally, lateral attention and longitudinal attention are merged:
Y=(X+yi)*(1+s)
Wherein, X is the input that transverse and longitudinal pays attention to power module, and Y is the output that transverse and longitudinal pays attention to power module.
5. object detection method as described in claim 1, which is characterized in that the spatial scaling feature gold in the step (5)
Word tower module:
Wherein I is characterized figure, and before LR is spatial scaling, after SR is spatial scaling, r be the step-length converted, x, y, and c is respectively
Column coordinate, row coordinate, depth coordinate.
6. object detection method as described in claim 1, which is characterized in that the variable segment in the step (6) decouples volume
Product:
Wherein, x is the characteristic pattern of input, and w is weight to be trained, and β, ρ are that can train variable, and E is to seek desired value.
7. object detection method as described in claim 1, which is characterized in that the multiple dimensioned training method in the step (7)
7~10 times of size of the characteristic pattern finally obtained using network are as network inputs.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810951609.5A CN109376576A (en) | 2018-08-21 | 2018-08-21 | The object detection method for training network from zero based on the intensive connection of alternately update |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810951609.5A CN109376576A (en) | 2018-08-21 | 2018-08-21 | The object detection method for training network from zero based on the intensive connection of alternately update |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109376576A true CN109376576A (en) | 2019-02-22 |
Family
ID=65403775
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810951609.5A Pending CN109376576A (en) | 2018-08-21 | 2018-08-21 | The object detection method for training network from zero based on the intensive connection of alternately update |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376576A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110020658A (en) * | 2019-03-28 | 2019-07-16 | 大连理工大学 | A kind of well-marked target detection method based on multitask deep learning |
| CN110210571A (en) * | 2019-06-10 | 2019-09-06 | 腾讯科技（深圳）有限公司 | Image-recognizing method, device, computer equipment and computer readable storage medium |
| CN110232316A (en) * | 2019-05-05 | 2019-09-13 | 杭州电子科技大学 | A kind of vehicle detection and recognition method based on improved DSOD model |
| CN110516670A (en) * | 2019-08-26 | 2019-11-29 | 广西师范大学 | Suggested based on scene grade and region from the object detection method for paying attention to module |
| CN110619369A (en) * | 2019-09-23 | 2019-12-27 | 常熟理工学院 | Fine-grained image classification method based on feature pyramid and global average pooling |
| CN111027512A (en) * | 2019-12-24 | 2020-04-17 | 北方工业大学 | Remote sensing image shore-approaching ship detection and positioning method and device |
| CN111079604A (en) * | 2019-12-06 | 2020-04-28 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Method for quickly detecting tiny target facing large-scale remote sensing image |
| CN111144364A (en) * | 2019-12-31 | 2020-05-12 | 北京理工大学重庆创新中心 | Twin network target tracking method based on channel attention updating mechanism |
| CN111210443A (en) * | 2020-01-03 | 2020-05-29 | 吉林大学 | Deformable convolution mixing task cascading semantic segmentation method based on embedding balance |
| CN111582225A (en) * | 2020-05-19 | 2020-08-25 | 长沙理工大学 | Remote sensing image scene classification method and device |
| CN111723829A (en) * | 2019-03-18 | 2020-09-29 | 四川大学 | Full-convolution target detection method based on attention mask fusion |
| CN111738045A (en) * | 2020-01-19 | 2020-10-02 | 中国科学院上海微系统与信息技术研究所 | Image detection method and device, electronic equipment and storage medium |
| CN111860619A (en) * | 2020-07-02 | 2020-10-30 | 苏州富鑫林光电科技有限公司 | Industrial detection AI intelligent model for deep learning |
| CN113239784A (en) * | 2021-05-11 | 2021-08-10 | 广西科学院 | Pedestrian re-identification system and method based on space sequence feature learning |
| CN113449756A (en) * | 2020-03-26 | 2021-09-28 | 太原理工大学 | Improved DenseNet-based multi-scale image identification method and device |
| CN117636078A (en) * | 2024-01-25 | 2024-03-01 | 华南理工大学 | Target detection method, target detection system, computer equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101241593A (en) * | 2007-02-06 | 2008-08-13 | 英特维数位科技股份有限公司 | Picture layer image processing unit and its method |
| CN105447864A (en) * | 2015-11-20 | 2016-03-30 | 小米科技有限责任公司 | Image processing method, device and terminal |
| US20160310043A1 (en) * | 2015-04-26 | 2016-10-27 | Endochoice, Inc. | Endoscopic Polyp Measurement Tool and Method for Using the Same |
| CN108038872A (en) * | 2017-12-22 | 2018-05-15 | 中国海洋大学 | One kind perceives follow method based on sound state target detection and Real Time Compression |
-
2018
- 2018-08-21 CN CN201810951609.5A patent/CN109376576A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101241593A (en) * | 2007-02-06 | 2008-08-13 | 英特维数位科技股份有限公司 | Picture layer image processing unit and its method |
| US20160310043A1 (en) * | 2015-04-26 | 2016-10-27 | Endochoice, Inc. | Endoscopic Polyp Measurement Tool and Method for Using the Same |
| CN105447864A (en) * | 2015-11-20 | 2016-03-30 | 小米科技有限责任公司 | Image processing method, device and terminal |
| CN108038872A (en) * | 2017-12-22 | 2018-05-15 | 中国海洋大学 | One kind perceives follow method based on sound state target detection and Real Time Compression |
Non-Patent Citations (8)
| Title |
|---|
| HAN ZHANG ET AL: ""Self-Attention Generative Adversarial Networks"", 《ARXIV》 * |
| JIE HU ET AL: ""Squeeze-and-Excitation Networks"", 《ARXIV》 * |
| JIFENG DAI ET AL: ""Deformable Convolutional Networks"", 《ARXIV》 * |
| PENG ZHOU ET AL: ""Scale-Transferrable Object Detection"", 《CVPR》 * |
| WEIYANG LIU ET AL: ""Decoupled Networks"", 《ARXIV》 * |
| YIBO YANG ET AL: ""Convolutional Neural Networks with Alternately Updated Clique"", 《ARXIV》 * |
| ZHIQIANG SHEN ET AL: "DSOD: Learning Deeply Supervised Object Detectors from Scratch", 《ARXIV》 * |
| 赵欣欣等: ""基于卷积神经网络的铁路桥梁高强螺栓缺失图像识别方法"", 《中国铁道科学》 * |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111723829B (en) * | 2019-03-18 | 2022-05-06 | 四川大学 | Full-convolution target detection method based on attention mask fusion |
| CN111723829A (en) * | 2019-03-18 | 2020-09-29 | 四川大学 | Full-convolution target detection method based on attention mask fusion |
| CN110020658A (en) * | 2019-03-28 | 2019-07-16 | 大连理工大学 | A kind of well-marked target detection method based on multitask deep learning |
| CN110232316A (en) * | 2019-05-05 | 2019-09-13 | 杭州电子科技大学 | A kind of vehicle detection and recognition method based on improved DSOD model |
| CN110210571A (en) * | 2019-06-10 | 2019-09-06 | 腾讯科技（深圳）有限公司 | Image-recognizing method, device, computer equipment and computer readable storage medium |
| CN110348543A (en) * | 2019-06-10 | 2019-10-18 | 腾讯医疗健康(深圳)有限公司 | Eye fundus image recognition methods, device, computer equipment and storage medium |
| CN110348543B (en) * | 2019-06-10 | 2023-01-06 | 腾讯医疗健康(深圳)有限公司 | Fundus image recognition method and device, computer equipment and storage medium |
| CN110210571B (en) * | 2019-06-10 | 2023-01-06 | 腾讯医疗健康(深圳)有限公司 | Image recognition method and device, computer equipment and computer readable storage medium |
| CN110516670A (en) * | 2019-08-26 | 2019-11-29 | 广西师范大学 | Suggested based on scene grade and region from the object detection method for paying attention to module |
| CN110516670B (en) * | 2019-08-26 | 2022-04-22 | 广西师范大学 | Target detection method based on scene level and area suggestion self-attention module |
| CN110619369B (en) * | 2019-09-23 | 2020-12-11 | 常熟理工学院 | Fine-grained image classification method based on feature pyramid and global average pooling |
| CN110619369A (en) * | 2019-09-23 | 2019-12-27 | 常熟理工学院 | Fine-grained image classification method based on feature pyramid and global average pooling |
| CN111079604A (en) * | 2019-12-06 | 2020-04-28 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Method for quickly detecting tiny target facing large-scale remote sensing image |
| CN111027512A (en) * | 2019-12-24 | 2020-04-17 | 北方工业大学 | Remote sensing image shore-approaching ship detection and positioning method and device |
| CN111027512B (en) * | 2019-12-24 | 2023-04-18 | 北方工业大学 | Remote sensing image quayside ship detection and positioning method and device |
| CN111144364B (en) * | 2019-12-31 | 2022-07-26 | 北京理工大学重庆创新中心 | Twin network target tracking method based on channel attention updating mechanism |
| CN111144364A (en) * | 2019-12-31 | 2020-05-12 | 北京理工大学重庆创新中心 | Twin network target tracking method based on channel attention updating mechanism |
| CN111210443A (en) * | 2020-01-03 | 2020-05-29 | 吉林大学 | Deformable convolution mixing task cascading semantic segmentation method based on embedding balance |
| CN111210443B (en) * | 2020-01-03 | 2022-09-13 | 吉林大学 | Deformable convolution mixing task cascading semantic segmentation method based on embedding balance |
| CN111738045A (en) * | 2020-01-19 | 2020-10-02 | 中国科学院上海微系统与信息技术研究所 | Image detection method and device, electronic equipment and storage medium |
| CN111738045B (en) * | 2020-01-19 | 2024-04-19 | 中国科学院上海微系统与信息技术研究所 | Image detection method and device, electronic equipment and storage medium |
| CN113449756A (en) * | 2020-03-26 | 2021-09-28 | 太原理工大学 | Improved DenseNet-based multi-scale image identification method and device |
| CN113449756B (en) * | 2020-03-26 | 2022-08-16 | 太原理工大学 | Improved DenseNet-based multi-scale image identification method and device |
| CN111582225A (en) * | 2020-05-19 | 2020-08-25 | 长沙理工大学 | Remote sensing image scene classification method and device |
| CN111582225B (en) * | 2020-05-19 | 2023-06-20 | 长沙理工大学 | Remote sensing image scene classification method and device |
| CN111860619A (en) * | 2020-07-02 | 2020-10-30 | 苏州富鑫林光电科技有限公司 | Industrial detection AI intelligent model for deep learning |
| CN113239784A (en) * | 2021-05-11 | 2021-08-10 | 广西科学院 | Pedestrian re-identification system and method based on space sequence feature learning |
| CN117636078A (en) * | 2024-01-25 | 2024-03-01 | 华南理工大学 | Target detection method, target detection system, computer equipment and storage medium |
| CN117636078B (en) * | 2024-01-25 | 2024-04-19 | 华南理工大学 | Target detection method, target detection system, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109376576A (en) | The object detection method for training network from zero based on the intensive connection of alternately update | |
| Huang et al. | Development and validation of a deep learning algorithm for the recognition of plant disease | |
| CN109299274A (en) | A kind of natural scene Method for text detection based on full convolutional neural networks | |
| CN108805070A (en) | A kind of deep learning pedestrian detection method based on built-in terminal | |
| CN108509976A (en) | The identification device and method of animal | |
| Li et al. | Semisupervised semantic segmentation of remote sensing images with consistency self-training | |
| CN109117877A (en) | A kind of Pelteobagrus fulvidraco and its intercropping kind recognition methods generating confrontation network based on depth convolution | |
| CN110334656A (en) | Multi-source Remote Sensing Images Clean water withdraw method and device based on information source probability weight | |
| Liu et al. | Two-stage underwater object detection network using swin transformer | |
| Zhang et al. | Adaptive anchor networks for multi-scale object detection in remote sensing images | |
| Liu et al. | Density-aware and background-aware network for crowd counting via multi-task learning | |
| Ye et al. | Recognition of terminal buds of densely-planted Chinese fir seedlings using improved YOLOv5 by integrating attention mechanism | |
| Wang et al. | Accurate real-time ship target detection using Yolov4 | |
| Wang et al. | TBC-YOLOv7: a refined YOLOv7-based algorithm for tea bud grading detection | |
| Peng et al. | An adaptive coarse-fine semantic segmentation method for the attachment recognition on marine current turbines | |
| Sun et al. | Prediction model for the number of crucian carp hypoxia based on the fusion of fish behavior and water environment factors | |
| Jia et al. | Polar-Net: Green fruit instance segmentation in complex orchard environment | |
| Xu et al. | Multiscale edge-guided network for accurate cultivated land parcel boundary extraction from remote sensing images | |
| Huang et al. | A Stepwise Refining Image-Level Weakly Supervised Semantic Segmentation Method for Detecting Exposed Surface for Buildings (ESB) From Very High-Resolution Remote Sensing Images | |
| CN118097709A (en) | Pig posture estimation method and device | |
| CN108230322A (en) | A kind of eyeground feature detection device based on weak sample labeling | |
| Zhao et al. | Ocean ship detection and recognition algorithm based on aerial image | |
| CN116758421A (en) | Remote sensing image directed target detection method based on weak supervised learning | |
| Mei et al. | A Method Based on Knowledge Distillation for Fish School Stress State Recognition in Intensive Aquaculture. | |
| CN115205215A (en) | Corneal nerve image segmentation method and system based on Transformer |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190222 |