CN110321944A - A kind of construction method of the deep neural network model based on contact net image quality evaluation - Google Patents
A kind of construction method of the deep neural network model based on contact net image quality evaluation Download PDFInfo
- Publication number
- CN110321944A CN110321944A CN201910557704.1A CN201910557704A CN110321944A CN 110321944 A CN110321944 A CN 110321944A CN 201910557704 A CN201910557704 A CN 201910557704A CN 110321944 A CN110321944 A CN 110321944A
- Authority
- CN
- China
- Prior art keywords
- image
- network model
- contact net
- images
- quality
- 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
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of construction methods of deep neural network model based on contact net image quality evaluation, comprising: 1. choose m images from contact net image set；2. screening training sample and test sample in each quality scale image by same ratio after m images are divided by quality category；3. testing the n network models trained respectively using test sample, the highest network model of measuring accuracy is filtered out as target network model；4. judging whether the measuring accuracy of target network model is not less than aimed at precision, if so, target network model is the deep neural network model for contact net image quality evaluation；Otherwise, step 5 is gone to；5. testing using training sample target network model, training sample is updated；6. going to step 4 to target network model training using training sample after updating；The present invention learns to contact the feature of net image automatically, has stronger test robustness and adaptability.
Description
Technical field
The invention belongs to the image quality measure of image procossing and contact net safety testing fields, more particularly, to one
Construction method based on the deep neural network model based on contact net image quality evaluation.
Background technique
Evaluation is carried out to the quality of image or photo and is widely used in field of image processing, and to contact net shooting figure as
Assessment is very necessary to defect problem present on subsequent detection contact net, if objective can judge which image is to belong to
In shooting failure, it can be greatly improved the speed and precision of the subsequent defects detection to different components, and railway contact line defect
Early find and repair importance it is even more self-evident.
In practical railway application scenarios, there is new image data daily, comprising a plurality of different rail tracks, every
Rail track data include the image of different elements of contacting net, parts of images may expose it is incorrect, cause all light be saturated or
The completely black shooting failure of person；Parts of images there is building, steel tower, number etc. manually or natural background is at components rear；Part figure
As may not photograph the elements of contacting net for needing to shoot because of a variety of causes.It is colored that image, which is the 24 of resolution ratio 2448 × 2050,
Image, usually 5,000,000 pixel industrial cameras are shot, and storage format JPG, compression ratio is about 20 times or so；It can also be used
Bar camera shoots 24 color images that resolution ratio is 1280 × 960；Therefore, image data amount is larger, artificial judgment work
Amount is very big；Based on current above situation, the labor intensity of people is not only greatly alleviated to the assessment of contact net picture quality, and
And can be according to contact net image quality measure as a result, carrying out different degrees of rejecting to the image of shooting failure, reduction needs to sentence
Disconnected amount of images is conducive to subsequent further according to the defect feelings of the different components of contact net image docking net-fault taken
Condition carries out detection judgement.
Traditional image quality measure method is largely by the way of picking identification feature, effective extraction of characteristics of image
Classification results have vital effect, and conventional method needs some features by extracting image, and the feature of image determines
The final performance of system；And long-term priori knowledge and design experiences are needed to the judgement of feature, in real system exploitation very
Hardly possible designs the optimal feature with discrimination.Therefore, the mode of conventionally employed picking identification feature claps railway facilities network
Take the photograph image assessment be primarily present two aspect deficiency: contact net image category is more, easily lead to different images feature for
The image adaptability of different scenes classification is poor, is difficult some unified feature and divides image well；Contact net image resolution
Rate is big, is taken a long time using conventional method.
Summary of the invention
In view of the drawbacks of the prior art, the purpose of the present invention is to provide a kind of, and the depth based on contact net image quality evaluation is refreshing
Construction method through network model, it is intended to which solve existing image quality measure method causes not because contact net image category is more
With the problem that characteristics of image is poor to image adaptability.
To achieve the above object, the present invention provides a kind of deep neural network models based on contact net image quality evaluation
Construction method, comprising:
(1) m images are chosen from contact net image set by the first preset ratio；
(2) after dividing m images by quality category, training is screened in each quality scale image by the second preset ratio
Sample, and residual image constitutes test sample；
(3) the n network models trained are tested using test sample respectively, filter out the highest network mould of measuring accuracy
Type is as target network model；
(4) judge whether the measuring accuracy of target network model is more than or equal to aimed at precision, if so, target network model
For the deep neural network model for contact net image quality evaluation；Otherwise, step (5) are gone to；
(5) target network model is tested using training sample, the image constructions of quality category test errors is new
Training sample；
(6) step (4) are gone to target network model training using new training sample.
Preferably, step (1) includes:
K images are randomly selected out from contact net image set by the first preset ratio and carry out gaussian filtering；
Mirror image processing, vertical reversion and 180 degree rotation are carried out to the k images by gaussian filtering, obtain m figures
Picture；
Wherein, the convolution kernel size of gaussian filtering is 3*3；M=8k；
Preferably, the training of n network model in step (3), comprising:
A. parts of images is taken out from training sample using the super ginseng parallel training of different learning rates by third preset ratio
Deep neural network model obtains the initial parameter value of n network model；
B., identical learning rate is set, is updated using parameter of the training sample to n network model, completes n network mould
The training of type.
Preferably, the identical learning rate setting are as follows: 0.001.
Preferably, the first preset ratio is 10% of all images in contact net image set；
Second preset ratio is that screening 90% is used as training sample, the image of residue 10% from the image of each quality scale
Constitute test sample.
Preferably, the quality category of step (2) are as follows: completely black, over-exposed, complete white and normal；
Wherein, completely black to account for the ratio of image area less than 10% for useful information pixel；Over-exposed is white pixel point
Account for the 30%~70% of image area；The ratio that Quan Baiwei white pixel point accounts for image area is more than or equal to 90%；Remaining image
It is divided into normal picture.
Preferably, step (2) specifically includes:
(2.1) ratio of image area is accounted for according to useful information pixel and white pixel point, sets quality category；
(2.2) m images are divided by quality category, and data balancing processing is done to the image in quality category；
(2.3) training sample is screened in by data balancing treated each quality scale image by the second preset ratio
This, and residual image constitutes test sample.
Contemplated above technical scheme through the invention, compared with prior art, can obtain it is following the utility model has the advantages that
1, the present invention chooses several images from contact net image set first and divides to its quality category, secondly, using
Multiple deep neural network models are tested in training sample and the test sample training that quality divides, and it is highest to filter out measurement accuracy
Target network model finally improves its measuring accuracy again；It can be to depth nerve according to the demand of Surveying Actual Precision
Network model is trained, and carries out abnormal point to contact net shooting figure picture using the higher deep neural network model of measurement accuracy
Class does not need manual extraction characteristics of image, but automatic study contacts the feature of net image, greatly strengthens the robust of measurement
Property, therefore even if even if contact net pattern class is more, but remain to distinguish different classes of figure well in practical application scene
Picture then does image quality evaluation to contact net shooting figure picture, therefore, the depth mind provided by the invention based on contact net image quality evaluation
There is extremely strong adaptability through network model.
2, the present invention carries out anomaly classification to shooting net image using deep neural network model, can frequency in neural network
It is numerous to use the operation such as convolution, down-sampling, the cracking reduction characteristic image resolution ratio of energy, while image main feature is saved, therefore phase
Compared with traditional images processing method, the high iron catenary image quality anomaly assessment method time-consuming based on deep neural network model is more
Few, speed is faster.
Detailed description of the invention
Fig. 1 is the building schematic diagram of deep neural network model provided by the invention；
Fig. 2 is the building schematic diagram for the deep neural network model that embodiment provides；
Fig. 3 (a) is a kind of contact net image schematic diagram provided by the invention for belonging to completely black classification；
Fig. 3 (b) is another contact net image schematic diagram provided by the invention for belonging to completely black classification；
Fig. 3 (c) is the contact net image schematic diagram provided by the invention for belonging to over-exposed classification；
Fig. 3 (d) is provided by the invention to belong to complete white contact net image schematic diagram；
Fig. 3 (e) is provided by the invention to belong to the normal contact net image schematic diagram of shooting；
Fig. 4 is the structural schematic diagram of deep neural network model provided by the invention；
Fig. 5 (a) is the test result schematic diagram provided by the invention for belonging to completely black classification；
Fig. 5 (b) is provided by the invention to belong to over-exposed test result schematic diagram；
Fig. 5 (c) is the test result schematic diagram provided by the invention for belonging to complete white classification；
Fig. 5 (d) is provided by the invention to belong to the normal test result schematic diagram of shooting.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
In practical application scene, acquisition contact net image is relatively difficult, therefore, the image quality measure method of no reference
It is the research topic of most practicability and challenge, and the image quality measure method of traditional no reference has very high meter
Complexity and time complexity are calculated, while the consistency between the objective quality assessed and subjective perception is poor.
As shown in Figure 1, the problem of being based on above-mentioned presentation, the present invention provides a kind of depth based on contact net image quality evaluation
The construction method of neural network model, comprising:
(1) m images are chosen from contact net image set by the first preset ratio；
(2) after dividing m images by quality category, training is screened in each quality scale image by the second preset ratio
Sample, and residual image constitutes test sample；
(3) the n network models trained are tested using test sample respectively, filter out the highest network mould of measuring accuracy
Type is as target network model；
(4) judge whether the measuring accuracy of target network model is more than or equal to aimed at precision, if so, target network model
For the deep neural network model for contact net image quality evaluation；Otherwise, step (5) are gone to；
(5) target network model is tested using training sample, the figure of quality scale classification error in test result
As constituting new training sample；
(6) step (4) are gone to target network model training using new training sample.
Embodiment
As shown in Fig. 2, choosing m images from contact net image set by the first preset ratio；
Specifically, n figures (1.1) image preprocessing: are randomly selected out from contact net image set by the first preset ratio
As carrying out gaussian filtering；
Further, it in order to avoid particularity, is first concentrated from source images and screens one in every ten images, selected at random altogether
Three thousand sheets images, since contact net image is shot at night more, image resolution ratio is high and there are more noises, to enhance contact net
Image quality measure and the precision of subsequent rejecting, therefore noise reduction first is filtered to the image picked out, to three thousand sheets images into
The convolution kernel size of row gaussian filtering, gaussian filtering is 3*3, and the Gaussian convolution core of 3*3 is as follows:
0.0585 0.0965 0.0585
0.0965 0.1529 0.0965
0.0585 0.0965 0.0585
(1.2) mirror image processing, vertical reversion and 180 degree are carried out to the k images by gaussian filtering to rotate, obtains k
Open image；
Wherein, the convolution kernel size of gaussian filtering is 3*3；M=8k；
Specifically, in order to increase trained high iron catenary image amount of training data, progress data augmentation: mirror image processing,
Vertical reversion and 180 degree rotation, original data volume are expanded octuple；
(2) after dividing m images by quality category, training is screened in each quality scale image by the second preset ratio
Sample, and residual image constitutes test sample；
Specifically, step (2) specifically includes:
(2.1) ratio of image area is accounted for according to useful information pixel and white pixel point, sets quality category；
Preferably, as shown in figure 3, for the assessment that contact net shoots picture quality, using the contact net figure of four quasi-representatives
As schematic diagram；By analyzing practical high iron catenary image data, by image clustering, matter is carried out to whole high iron catenary image
Amount calibration, is divided into four quality categories, the quality category of step (2) for the high iron catenary image after above-mentioned augmentation are as follows: complete
It is black, it is over-exposed, it is complete white and normal；
Wherein, as shown in figure (a) and Fig. 3 (b), the completely black ratio for accounting for image area for useful information pixel is less than
10%；It is over-exposed to account for the 30%~70% of image area for white pixel point as shown in Fig. 3 (c)；As shown in Fig. 3 (d), Quan Bai
The ratio for accounting for image area for white pixel point is more than or equal to 90%；As shown in Fig. 3 (e), remaining image is divided into normal picture；
(2.2) m images are divided by quality category, and data balancing processing is done to the image in quality category；
Specifically, m images are divided into four classes by quality category, that is, completely black, over-exposed, complete white and normal；It divides
After complete, it is predicted as the big quality category of data volume accounting in order not to mislead network for contact net image data, counts each quality
Amount of images in classification, and to the processing of its data balancing, update the image of each quality scale；
(2.3) training sample is screened in by data balancing treated each quality scale image by the second preset ratio
This, and residual image constitutes test sample；
90% will be screened from the image of each quality scale is used as training sample, the image construction test specimens of residue 10%
This；
(3) the n network models trained are tested using test sample respectively, filter out the highest network mould of measuring accuracy
Type is as target network model；
The training of n network model in step (3), comprising:
A. parts of images is taken out from training sample using the super ginseng parallel training of different learning rates by third preset ratio
Deep neural network model obtains the initial parameter value of n network model；
In the present embodiment, training is completed on ImageNet, MS COCO, Pascal Voc in online downloading
Initial weight of the caffemodel as deep neural network model, takes out parts of images from training sample, with different
The super ginseng parallel training deep neural network model of habit rate, obtains the initial parameter value of n network model；
B., identical learning rate is set, is updated using parameter of the training sample to n network model, completes n network mould
The training of type；
As shown in figure 4, it is 0.001 that identical learning rate is arranged in the present embodiment, it will be under last two layers of n network model
The learning rate of sampling expands 10 times, is updated using parameter of the training sample to n network model, completes the instruction of n network model
Practice；
(4) judge whether the measuring accuracy of target network model is more than or equal to aimed at precision, if so, target network model
For the deep neural network model for contact net image quality evaluation；Otherwise, step (5) are gone to；
(5) target network model is tested using training sample, the figure of quality scale classification error in test result
As constituting new training sample；
As shown in Fig. 5 (a), Fig. 5 (b), Fig. 5 (c) and Fig. 5 (d), in the present embodiment, test script is write, to training sample
In the image of each quality category test respectively, and test result is saved in corresponding differing document text, is taken out point
The data of class mistake will constitute new training sample after the image integration of classification error；
(6) step (4) are gone to target network model training using new training sample.
In practical application, the deep neural network model for contact net image quality evaluation obtained using step (4), to defeated
Enter after image carries out total quality monitoring and exports quality category.
Its quality category is divided in conclusion the present invention chooses several images from contact net image set first, secondly,
Multiple deep neural network models are tested in training sample and the test sample training divided using quality, filter out measurement accuracy
Highest target network model, finally improves its measuring accuracy again；It can be to depth according to the demand of Surveying Actual Precision
Degree neural network model is trained, and is carried out using the higher deep neural network model of measurement accuracy to contact net shooting figure picture
Anomaly classification does not need manual extraction characteristics of image, but automatic study contacts the feature of net image, greatly strengthens measurement
Robustness, therefore even if even if contact net pattern class is more, but remain to distinguish inhomogeneity well in practical application scene
Other image then does image quality evaluation to contact net shooting figure picture, therefore, provided by the invention based on contact net image quality evaluation
Deep neural network model has extremely strong adaptability.
The present invention carries out anomaly classification to shooting net image using deep neural network model, can be frequent in neural network
The operation such as convolution, down-sampling, the cracking reduction characteristic image resolution ratio of energy are used, while saving image main feature, therefore is compared
In traditional images processing method, the high iron catenary image quality anomaly assessment method based on deep neural network model is time-consuming less,
Speed is faster.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (7)
1. a kind of construction method of the deep neural network model based on contact net image quality evaluation characterized by comprising
(1) m images are chosen from contact net image set by the first preset ratio；
(2) after dividing m images by quality category, training sample is screened in each quality scale image by the second preset ratio
This, and residual image constitutes test sample；
(3) the n network models trained are tested using test sample respectively, the highest network model of measuring accuracy is filtered out and makees
For target network model；
(4) judge whether the measuring accuracy of target network model is more than or equal to aimed at precision, if so, target network model is to use
In the deep neural network model of contact net image quality evaluation；Otherwise, step (5) are gone to；
(5) target network model is tested using training sample, the image structure of quality scale classification error in test result
The training sample of Cheng Xin；
(6) step (4) are gone to target network model training using new training sample.
2. construction method as described in claim 1, which is characterized in that step (1) includes:
K images are randomly selected out from contact net image set by the first preset ratio and carry out gaussian filtering；
Mirror image processing, vertical reversion and 180 degree rotation are carried out to the k images by gaussian filtering, obtain m images；
Wherein, the convolution kernel size of gaussian filtering is 3*3；M=8k.
3. construction method as claimed in claim 1 or 2, which is characterized in that step (2) includes:
(2.1) ratio of image area is accounted for according to useful information pixel and white pixel point, sets quality category；
(2.2) m images are divided by quality category, and data balancing processing is done to the image in quality category；
(2.3) training sample is screened in by data balancing treated each quality scale image by the second preset ratio, and
Residual image constitutes test sample.
4. construction method as claimed in claim 3, which is characterized in that the quality category of step (2) are as follows: completely black, over-exposed,
It is complete white and normal；
Wherein, completely black to account for the ratio of image area less than 10% for useful information pixel；It is over-exposed to account for figure for white pixel point
The 30%~70% of image planes product；The ratio that Quan Baiwei white pixel point accounts for image area is more than or equal to 90%；Remaining image divides
For normal picture.
5. the construction method as described in Claims 1-4 is any, which is characterized in that the instruction of n network model in step (3)
Practice, comprising:
A. parts of images is taken out from training sample using the super ginseng parallel training depth of different learning rates by third preset ratio
Neural network model obtains the initial parameter value of n network model；
B., identical learning rate is set, is updated using parameter of the training sample to n network model, completes n network model
Training.
6. construction method as claimed in claim 5, which is characterized in that the identical learning rate setting are as follows: 0.001.
7. construction method as described in claim 1, which is characterized in that first preset ratio is complete in contact net image set
The 10% of portion's image；
Second preset ratio is that screening 90% is used as training sample, the image of residue 10% from the image of each quality scale
Constitute test sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910557704.1A CN110321944A (en) | 2019-06-26 | 2019-06-26 | A kind of construction method of the deep neural network model based on contact net image quality evaluation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910557704.1A CN110321944A (en) | 2019-06-26 | 2019-06-26 | A kind of construction method of the deep neural network model based on contact net image quality evaluation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110321944A true CN110321944A (en) | 2019-10-11 |
Family
ID=68120296
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910557704.1A Pending CN110321944A (en) | 2019-06-26 | 2019-06-26 | A kind of construction method of the deep neural network model based on contact net image quality evaluation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110321944A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111652395A (en) * | 2020-06-12 | 2020-09-11 | 成都国铁电气设备有限公司 | Health assessment method for high-speed railway contact network equipment |
| CN112381767A (en) * | 2020-10-27 | 2021-02-19 | 深圳大学 | Cornea reflection image screening method and device, intelligent terminal and storage medium |
| CN112784165A (en) * | 2021-01-29 | 2021-05-11 | 北京百度网讯科技有限公司 | Training method of incidence relation estimation model and method for estimating file popularity |
| CN114761145A (en) * | 2019-12-03 | 2022-07-15 | 克朗斯股份公司 | Method and device for identifying fallen and/or damaged containers in a material flow in a container |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104834898A (en) * | 2015-04-09 | 2015-08-12 | 华南理工大学 | Quality classification method for portrait photography image |
| CN107330455A (en) * | 2017-06-23 | 2017-11-07 | 云南大学 | Image evaluation method |
| CN107766929A (en) * | 2017-05-05 | 2018-03-06 | 平安科技（深圳）有限公司 | model analysis method and device |
| CN109558892A (en) * | 2018-10-30 | 2019-04-02 | 银河水滴科技（北京）有限公司 | A kind of target identification method neural network based and system |
| CN109859157A (en) * | 2018-11-16 | 2019-06-07 | 天津大学 | The full reference image quality appraisement method of view-based access control model attention characteristics |
-
2019
- 2019-06-26 CN CN201910557704.1A patent/CN110321944A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104834898A (en) * | 2015-04-09 | 2015-08-12 | 华南理工大学 | Quality classification method for portrait photography image |
| CN107766929A (en) * | 2017-05-05 | 2018-03-06 | 平安科技（深圳）有限公司 | model analysis method and device |
| CN107330455A (en) * | 2017-06-23 | 2017-11-07 | 云南大学 | Image evaluation method |
| CN109558892A (en) * | 2018-10-30 | 2019-04-02 | 银河水滴科技（北京）有限公司 | A kind of target identification method neural network based and system |
| CN109859157A (en) * | 2018-11-16 | 2019-06-07 | 天津大学 | The full reference image quality appraisement method of view-based access control model attention characteristics |
Non-Patent Citations (2)
| Title |
|---|
| 特伦斯·谢诺夫斯基: "深度学习", 《深度学习》 * |
| 王理同等: "基于循环神经网络的股指价格预测研究", 《浙江工业大学学报》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114761145A (en) * | 2019-12-03 | 2022-07-15 | 克朗斯股份公司 | Method and device for identifying fallen and/or damaged containers in a material flow in a container |
| CN111652395A (en) * | 2020-06-12 | 2020-09-11 | 成都国铁电气设备有限公司 | Health assessment method for high-speed railway contact network equipment |
| CN112381767A (en) * | 2020-10-27 | 2021-02-19 | 深圳大学 | Cornea reflection image screening method and device, intelligent terminal and storage medium |
| CN112381767B (en) * | 2020-10-27 | 2023-09-01 | 深圳大学 | Cornea reflection image screening method and device, intelligent terminal and storage medium |
| CN112784165A (en) * | 2021-01-29 | 2021-05-11 | 北京百度网讯科技有限公司 | Training method of incidence relation estimation model and method for estimating file popularity |
| CN112784165B (en) * | 2021-01-29 | 2024-07-19 | 北京百度网讯科技有限公司 | Training method of association relation prediction model and method for predicting file heat |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110321944A (en) | A kind of construction method of the deep neural network model based on contact net image quality evaluation | |
| CN108961217B (en) | Surface defect detection method based on regular training | |
| CN110047073B (en) | X-ray weld image defect grading method and system | |
| CN106875373B (en) | Mobile phone screen MURA defect detection method based on convolutional neural network pruning algorithm | |
| CN107123111B (en) | Deep residual error network construction method for mobile phone screen defect detection | |
| CN110473173A (en) | A kind of defect inspection method based on deep learning semantic segmentation | |
| CN109977790A (en) | A kind of video smoke detection and recognition methods based on transfer learning | |
| CN107607554A (en) | A kind of Defect Detection and sorting technique of the zinc-plated stamping parts based on full convolutional neural networks | |
| CN108021938A (en) | A kind of Cold-strip Steel Surface defect online detection method and detecting system | |
| CN106991666B (en) | A kind of disease geo-radar image recognition methods suitable for more size pictorial informations | |
| CN110378232B (en) | Improved test room examinee position rapid detection method of SSD dual-network | |
| CN109544555A (en) | Fine cracks dividing method based on production confrontation network | |
| CN110569730B (en) | Road surface crack automatic identification method based on U-net neural network model | |
| CN113469953B (en) | Transmission line insulator defect detection method based on improved YOLOv4 algorithm | |
| CN109191421B (en) | Visual detection method for pits on circumferential surface of cylindrical lithium battery | |
| CN109615604A (en) | Accessory appearance flaw detection method based on image reconstruction convolutional neural networks | |
| CN111047655A (en) | High-definition camera cloth defect detection method based on convolutional neural network | |
| CN106874929B (en) | Pearl classification method based on deep learning | |
| CN108564077A (en) | It is a kind of based on deep learning to detection and recognition methods digital in video or picture | |
| CN109583295B (en) | Automatic detection method for switch machine notch based on convolutional neural network | |
| CN108985337A (en) | A kind of product surface scratch detection method based on picture depth study | |
| CN108960413A (en) | A kind of depth convolutional neural networks method applied to screw surface defects detection | |
| CN116468666A (en) | Inspection image defect detection method special for operation and maintenance of power transmission line | |
| CN106951863A (en) | A kind of substation equipment infrared image change detecting method based on random forest | |
| CN109472790A (en) | A kind of machine components defect inspection method and system |
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: 20191011 |