CN109583367A - Image text row detection method and device, storage medium and electronic equipment - Google Patents
Image text row detection method and device, storage medium and electronic equipment Download PDFInfo
- Publication number
- CN109583367A CN109583367A CN201811435156.7A CN201811435156A CN109583367A CN 109583367 A CN109583367 A CN 109583367A CN 201811435156 A CN201811435156 A CN 201811435156A CN 109583367 A CN109583367 A CN 109583367A
- Authority
- CN
- China
- Prior art keywords
- image
- text
- training
- model
- confidence map
- 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
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/40—Document-oriented image-based pattern recognition
- G06V30/41—Analysis of document content
- G06V30/413—Classification of content, e.g. text, photographs or tables
-
- 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)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Biology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Embodiments of the present invention are related to field of computer technology, more specifically, embodiments of the present invention are related to image text row detection method and device, storage medium and electronic equipment.The described method includes: obtaining image to be detected；The text probability confidence map of described image to be detected is obtained using the identification model trained；The text probability confidence map is post-processed to obtain the line of text testing result of described image to be detected.The disclosure can be realized the accurate detection to line of text in image；And since a model being used only in detection process, to effectively improve detection speed；And testing process can be shortened, improve detection accuracy.
Description
Technical field
Embodiments of the present invention are related to field of computer technology, more specifically, embodiments of the present invention are related to image
Text line detection method and device, storage medium and electronic equipment.
Background technique
This part intends to provides background or context for the embodiments of the present invention stated in claim, retouching herein
It states and recognizes it is the prior art not because not being included in this section.
Natural scene image refers to by various capture apparatus (for example, camera, the mobile phone with shooting function, network are taken the photograph
As head etc.), image under conditions of no specific limitation, directly to the scene capture of necessary being in life.Due to nature
The factors such as differences and mixed and disorderly background such as the font, color of text, format in scene image, and in natural scene image
Later period artificially adds the situations such as text, so that detecting text in natural scene image is a more complex job.
Currently, in the Method for text detection having already appeared, it usually needs utilize multiple and different identification models, or repeat
Same model is used for multiple times to screen the character zone or text block in image, realizes the identification to text in image.
Summary of the invention
But in some technologies, on the one hand, image detection process is complicated, and step is more, need artificial setting in advance compared with
More rules influences to detect speed；On the other hand, the cascade of more step is easy to cause error accumulative, influences detection accuracy；Separately
On the one hand, most of method is only capable of detecting the horizontal text in image, can not effectively identify to multidirectional text.
Thus, it is also very desirable to a kind of improved image text row detection method and device, storage medium and electronic equipment, with
Optimize image text row testing process, promotes detection accuracy and speed.
In the present context, embodiments of the present invention are intended to provide a kind of image text row detection method and device, deposit
Storage media and electronic equipment.
According to one aspect of the disclosure, a kind of image text row detection method is provided, comprising:
Obtain image to be detected；
The text probability confidence map of described image to be detected is obtained using the identification model trained；
The text probability confidence map is post-processed to obtain the line of text testing result of described image to be detected.
In a kind of exemplary embodiment of the disclosure, the method also includes:
The identification model is trained in advance, comprising:
Obtain training image collection；
Full convolutional neural networks FCN model is trained to obtain the identification model according to the training image collection；
Wherein, the FCN model has feature pyramid network FPN structure.
It is described that FCN model is trained according to the training image collection in a kind of exemplary embodiment of the disclosure
Include:
Using the FCN model classification task and recurrence task are executed to the training image collection respectively, and is utilized respectively
Classification Loss function and recurrence loss function optimize the FCN model.
In a kind of exemplary embodiment of the disclosure, classification task is executed to the training image collection using FCN model,
FCN model is optimized using Classification Loss function and includes:
Classification task is executed to obtain corresponding text probability confidence map to training image using the FCN model；
It is lost using the error that Classification Loss function calculates the probability confidence map, and optimization institute is lost according to the error
State FCN model.
In a kind of exemplary embodiment of the disclosure, recurrence task is executed to the training image collection using FCN model,
Include: using returning loss function and being optimized to FCN model
Executing recurrence task to training image using the FCN model to obtain corresponding includes multiple line of text detection blocks
Text box detection image；
It is lost using the error that the recurrence loss function calculates the text box detection image, and is damaged according to the error
It loses and optimizes the FCN model.
In a kind of exemplary embodiment of the disclosure, the training identification model includes multiple cycles of training, institute
State method further include:
The weighted value for configuring Classification Loss function described in each cycle of training successively declines；And
The weighted value for configuring recurrence loss function described in each cycle of training is successively promoted.
In a kind of exemplary embodiment of the disclosure, the Classification Loss function are as follows:
Wherein, the xiIt is described for the predicted value of pixel on the probability confidence mapFor corresponding pixel points
The pixel value of Ground Truth.
In a kind of exemplary embodiment of the disclosure, the recurrence loss function are as follows:
Lloc=ω1*L1(r,r*)+ω2*L2(θ,θ*)+ω3*L3(r.w,r*.w)+ω4*L4(r.h,r*.h)
Wherein, ω1、ω2、ω3And ω4Respectively weight；r,r*Respectively indicate line of text detection block and Ground Truth
Detection block；θ,θ*Respectively indicate the angle of line of text detection block Yu Ground Truth frame；W is expressed as the width of detection block；H table
It is shown as the height of detection block.
In a kind of exemplary embodiment of the disclosure, classification task is executed to obtain to training image using FCN model
Corresponding text probability confidence map includes:
The training image is handled by FCN model to obtain each convolutional layer in the FCN model and export respectively
Characteristic pattern；
The text probability confidence map of the training image is predicted according to the characteristic pattern that each convolutional layer exports respectively.
In a kind of exemplary embodiment of the disclosure, the acquisition training image collection includes:
Obtain original image；
Random image processing is carried out to obtain training image to the original image, and generates training image collection.
In a kind of exemplary embodiment of the disclosure, the random image processing includes:
To the training image rotated, scaled and increased at random in random noise any one or it is any a variety of
Operation.
In a kind of exemplary embodiment of the disclosure, after the acquisition original image, the method also includes:
It identifies the original image type, and configures the ratio of all types of original images in preset rules.
It is described that the text probability confidence map is post-processed to obtain in a kind of exemplary embodiment of the disclosure
The line of text testing result of described image to be detected includes:
Binary conversion treatment is carried out to obtain bianry image to the probability confidence map；
Morphological scale-space is carried out by preset rules to the bianry image；
Connected component is carried out to the bianry image after Morphological scale-space to analyze to obtain line of text testing result.
It is described to obtain the mapping to be checked using the identification model trained in a kind of exemplary embodiment of the disclosure
The text probability confidence map of picture further include:
Size adjusting is carried out to obtain text identical with image to be detected size to the text probability confidence map
Probability confidence map.
According to one aspect of the disclosure, a kind of image text row detection device is provided, comprising:
Image collection module, for obtaining image to be detected；
Probability confidence map obtains module, and the text probability that the identification model for having trained obtains described image to be detected is set
Letter figure；
Testing result generation module, for obtaining the line of text of described image to be detected according to the text probability confidence map
Testing result.
In a kind of exemplary embodiment of the disclosure, described device further include:
Model training module, for training the identification model in advance；Include:
Obtain training image collection；
Full convolutional neural networks FCN model is trained to obtain the identification model according to the training image collection；
Wherein, the FCN model has feature pyramid network FPN structure.
In a kind of exemplary embodiment of the disclosure, the model training module includes:
Classification task module for executing classification task to training image collection using the FCN model, and utilizes classification damage
Function is lost to optimize the FCN model；
Task module is returned, for executing recurrence task to training image collection using the FCN model, and utilizes and returns damage
Function is lost to optimize the FCN model.
In a kind of exemplary embodiment of the disclosure, the classification task module includes:
Probability confidence map computing unit, for executing classification task to training image using the FCN model to obtain pair
The text probability confidence map answered；
Classification Loss computing unit, the error for being calculated the probability confidence map using Classification Loss function are lost, and
Optimize the FCN model according to error loss.
In a kind of exemplary embodiment of the disclosure, the recurrence task module includes:
Line of text detection block computing unit, for executing recurrence task to training image using the FCN model to obtain
The corresponding text box detection image including multiple line of text detection blocks；
Costing bio disturbance unit is returned, for calculating the error of the text box detection image using the recurrence loss function
Loss, and the FCN model is optimized according to error loss.
In a kind of exemplary embodiment of the disclosure, the model training module may include:
Cycle of training configuration unit, under configuring the weighted value of Classification Loss function described in each cycle of training successively
Drop；And
The weighted value for configuring recurrence loss function described in each cycle of training is successively promoted.
In a kind of exemplary embodiment of the disclosure, the model training module includes:
Original image acquiring unit, for obtaining original image；
Image processing unit for carrying out random image processing to the original image to obtain training image, and generates
Training image collection.
In a kind of exemplary embodiment of the disclosure, the model training module further include:
Image category recognition unit, the original image type for identification, and configure by preset rules all types of original
The ratio of image.
In a kind of exemplary embodiment of the disclosure, the testing result generation module includes:
Binary conversion treatment unit, for carrying out binary conversion treatment to the probability confidence map to obtain bianry image；
Morphological scale-space unit, for carrying out Morphological scale-space by preset rules to the bianry image；
Connected component analytical unit is analyzed for carrying out connected component to the bianry image after Morphological scale-space to obtain
Take line of text testing result.
In a kind of exemplary embodiment of the disclosure, the probability confidence map obtains module and includes:
Image scaling unit, for the text probability confidence map carry out size adjusting with obtain with it is described to be checked
The identical text probability confidence map of altimetric image size.
According to one aspect of the disclosure, a kind of storage medium is provided, computer program, the computer are stored thereon with
Above-mentioned image text row detection method when program is executed by processor.
According to one aspect of the disclosure, a kind of electronic equipment is provided, comprising:
Processor；And
Memory, for storing the executable instruction of the processor；
Wherein, the processor is configured to execute figure described in above-mentioned any one via the executable instruction is executed
As text line detection method.
The image text row detection method of embodiment according to the present invention, by utilizing the identification model trained to be checked
Altimetric image is handled, and the accurate text probability confidence map of image to be detected is obtained, to learn text box in image to be detected
Possible position.Again by being post-processed to probability confidence map, to obtain the accurate line of text position of image to be detected.
To realize the accurate detection to line of text in image.Since a model being used only in detection process, to effectively improve
Detect speed；And testing process can be shortened, improve detection accuracy.
Detailed description of the invention
The following detailed description is read with reference to the accompanying drawings, above-mentioned and other mesh of exemplary embodiment of the invention
, feature and advantage will become prone to understand.In the accompanying drawings, if showing by way of example rather than limitation of the invention
Dry embodiment, in which:
Fig. 1 schematically shows the flow chart of image text row detection method；
Fig. 2 schematically shows the flow charts for the method for obtaining training image collection；
Fig. 3 schematically shows the flow chart of the method post-processed to text probability confidence map；
Fig. 4 schematically shows the schematic diagrames an of image to be detected；
Fig. 5, which is schematically shown, shows the corresponding text probability confidence map of image to be detected in Fig. 4；
Fig. 6, which is schematically shown, shows the corresponding line of text testing result schematic diagram of image to be detected in Fig. 4；
Fig. 7 schematically shows the schematic diagram of a training image；
Fig. 8, which is schematically shown, executes each period schematic diagram of classification task to the training image shown in Fig. 7；
Fig. 9, which is schematically shown, executes each period schematic diagram of recurrence task to the training image shown in Fig. 7；
Figure 10 schematically shows the testing result schematic diagram to the training image shown in Fig. 7；
Figure 11 schematically shows the schematic diagram of another image to be detected；
Figure 12 schematically shows the schematic diagram of the text probability confidence map of image to be detected shown in Figure 11；
Figure 13 schematically shows the schematic diagram of the corresponding Morphological scale-space result of the confidence map of text probability shown in Figure 12；
Figure 14 schematically shows the signal of the corresponding connected component analysis result of Morphological scale-space result shown in Figure 13
Figure；
Figure 15 schematically shows the schematic diagram of the corresponding line of text testing result of image to be detected shown in Figure 11；
Figure 16 schematically shows the composition schematic diagram of image text row detection device；
Figure 17 shows the schematic diagrames of the storage medium of embodiment according to the present invention；And
Figure 18 diagrammatically illustrates the block diagram of the electronic equipment according to invention embodiment.
In the accompanying drawings, identical or corresponding label indicates identical or corresponding part.
Specific embodiment
The principle and spirit of the invention are described below with reference to several illustrative embodiments.It should be appreciated that providing this
A little embodiments are used for the purpose of making those skilled in the art can better understand that realizing the present invention in turn, and be not with any
Mode limits the scope of the invention.On the contrary, these embodiments are provided so that this disclosure will be more thorough and complete, and energy
It is enough that the scope of the present disclosure is completely communicated to those skilled in the art.
One skilled in the art will appreciate that embodiments of the present invention can be implemented as a kind of system, device, equipment, method
Or computer program product.Therefore, the present disclosure may be embodied in the following forms, it may be assumed that complete hardware, complete software
The form that (including firmware, resident software, microcode etc.) or hardware and software combine.
Embodiment according to the present invention provides a kind of image text row detection method, image text row detection device, deposits
Storage media and electronic equipment.
Herein, any number of elements in attached drawing is used to example rather than limitation and any name are only used for
It distinguishes, without any restrictions meaning.
Below with reference to several representative embodiments of the invention, the principle and spirit of the invention are elaborated.
Summary of the invention
The inventors discovered that the detection of text, which generally requires, in the prior art, in image is used for multiple times analysis model, or
Using multiple and different models, so that text detection process steps are more, process is complicated, is unfavorable for the training and entirety to model
Optimization, also limits detection speed；In addition, more step is easier to cause the accumulation of error, detection accuracy is influenced.
In view of above content, basic thought of the invention is: the image text row detection of embodiment according to the present invention
Method and image text row detection device, are handled image to be detected using the identification model trained, and are obtained to be detected
The accurate text probability confidence map of image, to learn the possible position of text box in image to be detected.Again by being set to probability
Letter figure is post-processed, to obtain the accurate line of text position of image to be detected.And then available literal line detection knot
Content of text in fruit.To realize the accurate detection to line of text in image and corresponding text.Due to only making in detection process
With a model, speed is detected to effectively improve；And testing process can be shortened, improve detection accuracy.
After introduced the basic principles of the present invention, lower mask body introduces various non-limiting embodiment party of the invention
Formula.
Illustrative methods
The image text row detection method of illustrative embodiments according to the present invention is described below with reference to Fig. 1.
With reference to Fig. 1, described image text line detection method be may comprise steps of:
S1, image to be detected is obtained；
S2, the text probability confidence map that described image to be detected is obtained using the identification model trained；
S3, the text probability confidence map is post-processed to obtain the line of text of described image to be detected detection knot
Fruit.
In the image text row detection method of embodiment of the present invention, on the one hand, by utilizing identification trained in advance
Model handles image to be detected, the available accurate text probability confidence map of image to be detected, to learn to be checked
The possible position of text box in altimetric image.By being post-processed to probability confidence map, to obtain the accurate of image to be detected
Line of text position, and then realize line of text each in image to be detected is accurately identified.Further, mapping to be checked is being obtained
After line of text as in, corresponding content of text in literal line testing result can also be obtained, and then realize to text in image
The accurate detection of capable and corresponding text.On the other hand, due to only being needed during line of text detects using an identification mould
Type, the execution process of the text line detection method effectively shortened, avoids the deviation accumulation between multiple steps, effectively mentions
The high detection accuracy of line of text.In addition, can also effectively improve the detection speed of line of text by method for reducing process.
In step sl, image to be detected is obtained.
In the illustrative embodiments of the disclosure, for server end, the to be checked of user terminal transmission can receive
Altimetric image.For example, image to be detected can be the net from types such as forum, social platform, game or shopping websites
The image obtained in the network page or application program, or the image using photographing device shooting.Original image may include nature
The image of the types such as scene image, or the composograph that is added the texts such as number, text.For example, with reference to Fig. 4, Fig. 7 or
Shown in Figure 11, image to be detected is the image for including text information.
In step s 2, the identification model that utilization has been trained obtains the text probability confidence map of described image to be detected.
In the illustrative embodiments of the disclosure, the identification model can be trained in advance.Using image to be detected as
Input, handles image to be detected using the identification model, obtains corresponding text probability confidence map.
Specifically, the above-mentioned identification model of training may include:
In the step s 21, training image collection is obtained.
For above-mentioned training image collection, including multiple training images, can be obtained by being handled original image
It takes.Specifically, refering to what is shown in Fig. 2, above-mentioned step S21 may include:
Step S211 obtains original image；
Step S212 identifies the original image type, and the ratio of all types of original images is configured in preset rules；
Step S213 carries out random image processing to the original image to obtain training image, and generates training image
Collection.
For example, original image may include natural scene image, or be added the conjunction of the texts such as number, text
At the image of the types such as image.When generating training dataset using original image, can classify first to original image.
And after the classification for identifying each image, different classes of original image is chosen according to a certain percentage to generate original image set.It is real
The ratio of all types of images and the reasonable distribution of quantity are now concentrated to training image, guarantee the diversity of training image, effectively
The training image that control does not include text is less than certain threshold value, and then lift scheme training in the ratio that entire training image is concentrated
Efficiency.
In addition, in some illustrative embodiments of the disclosure, can according in image whether comprising text to original
Image is divided into the classification of the classification comprising text and pure image.Alternatively, being divided into according to image background content complexity
Complex background image and simple background image.It certainly, can also be according to image in other exemplary embodiments of the disclosure
Other attributes or feature, or classify simultaneously according to several different attributes and feature, the disclosure is to specific image point
Rule-like does not do particular determination.
For example, if the ratio of pure image and the image comprising text that setting original image is concentrated is 1:1, in original graph
When image set setting total number of images amount is 100, then it can choose simple image 50 and open, and the image 50 comprising text is opened, with life
At original image set.
In addition, random image processing can also be carried out to original image therein after obtaining original image set, such as
Random to original image is rotated, scaled or is increased one or more processing such as random noise.And using at random image
Original image after reason generates training image set, realizes the amplification to training data.
Certainly, in other exemplary embodiments of the disclosure, can not also limit original image of all categories quantity or
Ratio directly generates training image collection using original image；Alternatively, generating training after carrying out random image processing to original image
Image set.
In step S22, full convolutional neural networks FCN model is trained to obtain according to the training image collection
State identification model；Wherein, the FCN model have feature pyramid network (Feature Pyramid Networks, referred to as
FPN) structure.
In the illustrative embodiments of the disclosure, the FCN with feature pyramid network (FPN) structure can use
(Fully Convolutional Networks, full convolutional neural networks) model executes classification to training image collection respectively simultaneously
Task and recurrence task utilize using the probability confidence map of classification task learning text and return tasking learning text detection frame
Return information.And it is excellent to the above-mentioned FCN model progress with FPN structure to be utilized respectively Classification Loss function and recurrence loss function
Change.According to the accurate text probability confidence of the probability confidence map of training image and the available training image of text detection frame
Figure.Feature pyramid network FPN structure, bottom-layer network high resolution used in embodiment of the present invention can learn more
More minutias；And its upper layer network resolution ratio is low, can learn more semantic features.Therefore, embodiment of the present invention
Bottom and high-rise feature can be effectively merged, can not only retain strong semantic information, and high-resolution prediction can be obtained
As a result.In addition, being more favorable to detect Small object.
Specifically, can use the classification that the FCN model executes Pixel-level to training image for classification task
Task is to obtain corresponding text probability confidence map；Obtain training image in each pixel whether be text probability.It is specific next
It says, above-mentioned training image can be handled by FCN model to obtain the feature that each convolutional layer exports respectively in FCN model
Figure；The text probability confidence map of above-mentioned training image is predicted according to the characteristic pattern that each convolutional layer exports respectively.
Classification Loss function is recycled to calculate the error loss of above-mentioned probability confidence map, and should according to error loss optimization
FCN model.Wherein, Classification Loss function are as follows:
Wherein, the xiFor the predicted value of pixel on the probability confidence map；The xi *For corresponding pixel points
The pixel value of Ground Truth, wherein pixel is that text is then 1, and non-legible value is 0.
For recurrence task, it can use above-mentioned FCN model and recurrence task executed to obtain correspondence to training image
The text box detection image including multiple line of text detection blocks, such as the recurrence task of Pixel-level can be executed, so that training
Each pixel all has corresponding detection block in image, and is determined whether to retain the corresponding detection of each pixel according to confidence score
Frame.Then, it recycles and returns the error loss that loss function calculates each text box detection image, and optimized according to error loss
The FCN model.Wherein, loss function is returned are as follows:
Lloc=ω1*L1(r,r*)+ω2*L2(θ,θ*)+ω3*L3(r.w,r*.w)+ω4*L4(r.h,r*.h)
Wherein, ω1、ω2、ω3And ω4Respectively weight；r,r*Respectively indicate line of text detection block and Ground Truth
Detection block；Line of text detection block r and Ground Truth detection block r*Area hand over and compare loss be denoted as L1(r,r*)；θ,θ*Point
Not Biao Shi line of text detection block and Ground Truth detection block angle, the angle, θ and Ground of line of text detection block
The angle, θ of Truth detection block*Complementary chord angle loss be denoted as L2(θ,θ*)；R.w is expressed as the width of line of text detection block, r*.w
It is expressed as the width of Ground Truth detection block, square of line of text detection block and the width difference of Ground Truth detection block
Canonical loss is denoted as L3(r.w,r*.w)；R.h is expressed as the height of line of text detection block, and r*.h is expressed as Ground Truth inspection
The height of frame is surveyed, a square canonical loss for line of text detection block and Ground Truth detection block difference in height is denoted as L4(r.h,r*
.h)。
Since the corresponding text box of text in training image is most of in rectangular configuration, and the width and height of text box
There may be biggish differences, and the regular terms loss by increasing detection block height and width can effectively promote detection block
Recall rate is conducive to the accurate detection of line of text.By adding the cosine for line of text detection block in Classification Loss function
Angle loss, refering to what is shown in Fig. 6, can realize accurate detection to text non-horizontal in image to be detected.Further, since rectangle
There may be biggish differences to lose different regular terms by configuring different weight coefficients for the width of text and height
With different weights, the difference can be effectively balanced, to be more advantageous to the detection of line of text.
By being trained to the FCN model with FPN structure, since the FCN model in present embodiment does not include entirely
Articulamentum, when using FCN model as basic frame (such as resnet-34 etc.), can make input picture is arbitrary size
Image, in the training process and to image carry out line of text detection during can be not required to adjust the size of image
It is whole.
Preferably, in other illustrative embodiments of the disclosure, in the above-mentioned identification model of training, in order to improve mould
The recognition effect of type, due to having stronger correlation between classification task and recurrence task, it is therefore possible to use multitask
The mode of habit is by classification task and returns task cooperative training, and configuration training pattern further includes multiple cycles of training, can be simultaneously
Improve the accuracy of identification of two tasks.Specifically, above-mentioned method can also include:
The weighted value for configuring Classification Loss function in each cycle of training successively declines；And
The weighted value for configuring recurrence loss function in each cycle of training is successively promoted.
For example, Classification Loss function can be distributed by following formula and return the weight and ratio of loss function, comprising:
Ldet=α1Lconf+α2*Lloc
Wherein, LconfFor Classification Loss, LlocTo return loss；α1、α2For coefficient.
Training image as shown in Figure 7 can configure the coefficient of loss function not in the different cycles of training pattern
Together, so that each stage has different training effects.For example, three instructions can will be arranged cycle of training with reference to shown in Fig. 8, Fig. 9
Practice the period, setting period 1 Classification Loss function and return loss function ratio be 10:1, the stage retraining classification
Task；It is 1:1 that second round Classification Loss function, which is arranged, and returns the ratio of loss function, which trains classification task simultaneously
With the task of recurrence；Be arranged period 3 Classification Loss function and return loss function ratio be 1:10, the stage retraining
Recurrence task.Further, the corresponding text detection row testing result of training image shown in Fig. 7 is as shown in Figure 10.
Certainly, in other illustrative embodiments of the disclosure, the cycle of training of other quantity also can be set, such as
Two, five or seven cycles of training are set, and be arranged in each period Classification Loss function with return loss function with different
Ratio.It can specifically be lost according to the ratio selection cycle quantity and configuration of the quantity of training image, type and all types of images
The ratio of function.The training identification model in the way of multi-task learning, cycle of training is configured, identification model can be made
The testing process of training process and identification model is more in line with the characteristic of line of text, and then can effectively improve FCN model
Robustness.
In addition, obtaining the corresponding probability confidence map of image to be detected in other illustrative embodiments of the disclosure
Afterwards, size adjusting can also be carried out to it, so as to obtain text probability confidence map identical with image to be detected size, made
Probability confidence map keeps size consistent with original image to be detected, convenient for the subsequent processing to probability confidence map.
In step s3, the text probability confidence map is post-processed to obtain the line of text of described image to be detected
Testing result.
In the illustrative embodiments of the disclosure, refering to what is shown in Fig. 3, above-mentioned step S3 can specifically include:
Step S31 carries out binary conversion treatment to the probability confidence map to obtain bianry image；
Step S32 carries out Morphological scale-space by preset rules to the bianry image；
Step S33 carries out connected component analysis to the bianry image after Morphological scale-space to obtain line of text detection
As a result.
After acquisition probability confidence map, binary conversion treatment can be carried out to it, according to preset rules by probability confidence map
In each pixel gray value be set as 0 or 255, obtain corresponding bianry image.Then morphology can be carried out to bianry image
The operations such as processing, such as burn into expansion, opening operation and closed operation.Specifically, can be selected according to image shape in bianry image
Take corresponding Morphological scale-space mode.For example, text has width and the apparent feature of difference in height in bianry image, then it can be with
The filtering core of strip is set, expansion process is carried out to it.For example, image to be detected as shown in figure 11, at identified model
It is as shown in figure 12 that corresponding text probability confidence map is obtained after reason.With reference to shown in Figure 13, to probability confidence map shown in Figure 12 into
Row binary conversion treatment, and carry out expansion process, the bianry image after obtaining corresponding Morphological scale-space.
After carrying out Morphological scale-space to bianry image, it can be analyzed using connected component, obtain mapping to be checked
As corresponding line of text testing result.For example, utilizing strip kernel to the bianry image after Morphological scale-space shown in Figure 13
Connected component analysis, obtain connected component analysis result as shown in figure 14.Finally, image to be detected shown in Figure 11 is obtained
Corresponding line of text testing result finally detects the multiple line of text for including in image to be detected as shown in figure 15.
In conclusion by the text probability confidence map for obtaining image to be detected using identification model, then by probability
Confidence map is post-processed, to accurately obtain the line of text position of image to be detected.Due to only being needed in line of text detection process
Using an identification model, shorten testing process, to effectively improve detection speed and detection accuracy.
Exemplary means
After describing the image text row detection method of exemplary embodiment of the invention, next, with reference to Figure 16
The image text row detection device of exemplary embodiment of the invention is described.
With reference to shown in Figure 16, the image text row detection device 100 of exemplary embodiment of the invention may include: image
Obtain module 1001, probability confidence map obtains module 1002 and testing result generation module 1003.Wherein:
Described image, which obtains module 1001, can be used for obtaining image to be detected.
The probability confidence map obtains the identification model that module 1002 can be used for having trained and obtains described image to be detected
Text probability confidence map.
The testing result generation module 1003 can be used for being obtained according to the text probability confidence map described to be detected
The line of text testing result of image.
According to an exemplary embodiment of the present disclosure, described image line of text detection device 100 can also include: model training
Module.
The model training module can be used for training the identification model in advance, comprising:
Obtain training image collection；
Full convolutional neural networks FCN model is trained to obtain the identification model according to the training image collection；
Wherein, the FCN model has feature pyramid network FPN structure.
According to an exemplary embodiment of the present disclosure, the model training module may include: classification task module and recurrence
Task module.Wherein:
The classification task module can be used for executing classification task to the training image collection using the FCN model,
And the FCN model is optimized using Classification Loss function.
The recurrence task module can be used for executing recurrence task to the training image collection using the FCN model,
And the FCN model is optimized using loss function is returned.
According to an exemplary embodiment of the present disclosure, the classification task module may include: probability confidence map computing unit
With Classification Loss computing unit.Wherein:
The probability confidence map computing unit can be used for using the FCN model to training image execute classification task with
Obtain corresponding text probability confidence map.
The Classification Loss computing unit can be used for calculating the error of the probability confidence map using Classification Loss function
Loss, and the FCN model is optimized according to error loss.
According to an exemplary embodiment of the present disclosure, the recurrence task module may include: that line of text detection block calculates list
Member and recurrence costing bio disturbance unit.Wherein:
The line of text detection block computing unit can be used for executing recurrence task to training image using the FCN model
To obtain the corresponding text box detection image including multiple line of text detection blocks.
The recurrence costing bio disturbance unit can be used for calculating the text box detection figure using the recurrence loss function
The error of picture is lost, and optimizes the FCN model according to error loss.
According to an exemplary embodiment of the present disclosure, the model training module may include: configuration unit cycle of training.
Configuration module cycle of training can be used for configuring the weighted value of Classification Loss function described in each cycle of training
Successively decline；And the weighted value for configuring recurrence loss function described in each cycle of training is successively promoted.
According to an exemplary embodiment of the present disclosure, the Classification Loss function are as follows:
Wherein, the xiIt is described for the predicted value of pixel on the probability confidence mapFor corresponding pixel points
The pixel value of Ground Truth.
According to an exemplary embodiment of the present disclosure, the recurrence loss function are as follows:
Lloc=ω1*L1(r,r*)+ω2*L2(θ,θ*)+ω3*L3(r.w,r*.w)+ω4*L4(r.h,r*.h)
Wherein, ω1、ω2、ω3And ω4Respectively weight；r,r*Respectively indicate line of text detection block and Ground Truth
Detection block；θ,θ*Respectively indicate the angle of line of text detection block Yu Ground Truth detection block；W is expressed as the width of detection block
Degree；H is expressed as the height of detection block.
According to an exemplary embodiment of the present disclosure, the probability confidence map computing unit may include: that characteristic pattern calculates list
Member, characteristic pattern predicting unit.Wherein:
The characteristic pattern computing unit can be used for handling the training image by FCN model with described in obtaining
The characteristic pattern that each convolutional layer exports respectively in FCN model.
The characteristic pattern that the characteristic pattern predicting unit can be used for being exported respectively according to each convolutional layer predicts the instruction
Practice the text probability confidence map of image.
According to an exemplary embodiment of the present disclosure, the model training module may include: original image acquiring unit, figure
As processing unit.Wherein:
The original image acquiring unit can be used for obtaining original image.
Described image processing unit can be used for carrying out the original image random image and handle to obtain training image,
And generate training image collection.
According to an exemplary embodiment of the present disclosure, the random image processing includes: to carry out at random to the training image
Any one in rotation, scaling and increase random noise or any a variety of operations.
According to an exemplary embodiment of the present disclosure, the model training module can also include: image category recognition unit.
Described image classification recognition unit can be used for identifying the original image type, and all kinds of by preset rules configuration
The ratio of type original image.
According to an exemplary embodiment of the present disclosure, the testing result generation module 1003 may include: binary conversion treatment
Unit, Morphological scale-space unit and connected component analytical unit.Wherein:
The binary conversion treatment unit can be used for carrying out the probability confidence map binary conversion treatment to obtain binary map
Picture.
The Morphological scale-space unit can be used for carrying out Morphological scale-space by preset rules to the bianry image.
The connected component analytical unit can be used for carrying out connected component to the bianry image after Morphological scale-space
Analysis is to obtain line of text testing result.
According to an exemplary embodiment of the present disclosure, it can also include: image ruler that the probability confidence map, which obtains module 1002,
Very little adjustment unit.
Described image size adjusting unit can be used for the text probability confidence map carry out size adjusting with obtain with
The identical text probability confidence map of image to be detected size.
Each functional module and above-mentioned image text row due to the image text row detection device of embodiment of the present invention
It is identical in detection method invention embodiment, therefore details are not described herein.
Exemplary storage medium
After the image text row detection method and device for describing exemplary embodiment of the invention, next, ginseng
Figure 17 is examined to be illustrated the storage medium of exemplary embodiment of the invention.
With reference to shown in Figure 17, the program product for realizing the above method of embodiment according to the present invention is described
700, can using portable compact disc read only memory (CD-ROM) and including program code, and can in equipment, such as
It is run on PC.However, program product of the invention is without being limited thereto, in this document, readable storage medium storing program for executing, which can be, appoints
What include or the tangible medium of storage program that the program can be commanded execution system, device or device use or and its
It is used in combination.
Described program product can be using any combination of one or more readable mediums.Readable medium can be readable letter
Number medium or readable storage medium storing program for executing.Readable storage medium storing program for executing for example can be but be not limited to electricity, magnetic, optical, electromagnetic, infrared ray or
System, device or the device of semiconductor, or any above combination.The more specific example of readable storage medium storing program for executing is (non exhaustive
List) include: electrical connection with one or more conducting wires, portable disc, hard disk, random access memory (RAM), read-only
Memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable compact disc read only memory
(CD-ROM), light storage device, magnetic memory device or above-mentioned any appropriate combination.
Computer-readable signal media may include in a base band or as carrier wave a part propagate data-signal,
In carry readable program code.The data-signal of this propagation can take various forms, including but not limited to electromagnetic signal,
Optical signal or above-mentioned any appropriate combination.Readable signal medium can also be any readable Jie other than readable storage medium storing program for executing
Matter, the readable medium can send, propagate or transmit for by instruction execution system, device or device use or and its
The program of combined use.
The program code for including on readable medium can transmit with any suitable medium, including but not limited to wirelessly, have
Line, optical cable, RF etc. or above-mentioned any appropriate combination.
The program for executing operation of the present invention can be write with any combination of one or more programming languages
Code, described program design language include object oriented program language-Java, C++ etc., further include conventional mistake
Formula programming language-such as " C " language or similar programming language.Program code can be calculated fully in user
It is executed in equipment, part executes on a remote computing or sets completely in remote computation on the user computing device for part
It is executed on standby or server.In the situation for being related to remote computing device, remote computing device can pass through the net of any kind
Network, including local area network (LAN) or wide area network (WAN), are connected to user calculating equipment, or, it may be connected to outside calculates and sets
Standby (such as being connected using ISP by internet).
Example electronic device
After describing the storage medium of exemplary embodiment of the invention, next, with reference to Fig. 8 to example of the present invention
The electronic equipment of property embodiment is illustrated.
The electronic equipment 800 that Figure 18 is shown is only an example, function to the embodiment of the present invention and should not use model
Shroud carrys out any restrictions.
As shown in figure 18, electronic equipment 800 is showed in the form of universal computing device.The component of electronic equipment 800 can be with
Including but not limited to: at least one above-mentioned processing unit 810, at least one above-mentioned storage unit 820, the different system components of connection
The bus 830 of (including storage unit 820 and processing unit 810), display unit 840.
Wherein, the storage unit is stored with program code, and said program code can be held by the processing unit 810
Row, so that various according to the present invention described in the execution of the processing unit 810 above-mentioned " illustrative methods " part of this specification
The step of illustrative embodiments.For example, the processing unit 810 can execute step S1 to step S3 as shown in fig. 1.
Storage unit 820 may include volatile memory cell, for example, Random Access Storage Unit (RAM) 8201 and/or
Cache memory unit 8202 can further include read-only memory unit (ROM) 8203.
Storage unit 820 can also include program/utility with one group of (at least one) program module 8205
8204, such program module 8205 includes but is not limited to: operating system, one or more application program, other program moulds
It may include the realization of network environment in block and program data, each of these examples or certain combination.
Bus 830 may include data/address bus, address bus and control bus.
Electronic equipment 800 can also be with one or more external equipments 900 (such as keyboard, sensing equipment, bluetooth equipment
Deng) communicate, this communication can be carried out by input/output (I/O) interface 850.Electronic equipment 800 further includes display unit
840, it is connected to input/output (I/O) interface 850, for being shown.Also, electronic equipment 800 can also pass through network
Adapter 860 and one or more network (such as local area network (LAN), wide area network (WAN) and/or public network, such as because of spy
Net) communication.As shown, network adapter 860 is communicated by bus 830 with other modules of electronic equipment 800.It should be bright
It is white, although not shown in the drawings, other hardware and/or software module can be used in conjunction with electronic equipment 800, including but not limited to:
Microcode, device driver, redundant processing unit, external disk drive array, RAID system, tape drive and data are standby
Part storage system etc..
It should be noted that although being referred to several modules or submodule of image text row detection device in the above detailed description
Block, but it is this division be only exemplary it is not enforceable.In fact, embodiment according to the present invention, is retouched above
The feature and function for two or more units/modules stated can embody in a units/modules.Conversely, above description
A units/modules feature and function can with further division be embodied by multiple units/modules.
In addition, although describing the operation of the method for the present invention in the accompanying drawings with particular order, this do not require that or
Hint must execute these operations in this particular order, or have to carry out shown in whole operation be just able to achieve it is desired
As a result.Additionally or alternatively, it is convenient to omit multiple steps are merged into a step and executed by certain steps, and/or by one
Step is decomposed into execution of multiple steps.Although detailed description of the preferred embodimentsthe spirit and principles of the present invention are described by reference to several,
However, it should be understood that the invention is not limited to the specific embodiments disclosed, these are not meant that yet to the division of various aspects
Feature in aspect cannot combine it is benefited to carry out, it is this divide merely to statement convenience.It is appended the present invention is directed to cover
Included various modifications and equivalent arrangements in spirit and scope of the claims.
Claims (10)
1. a kind of image text row detection method characterized by comprising
Obtain image to be detected；
The text probability confidence map of described image to be detected is obtained using the identification model trained；
The text probability confidence map is post-processed to obtain the line of text testing result of described image to be detected.
2. image text row detection method according to claim 1, which is characterized in that the method also includes:
The identification model is trained in advance, comprising:
Obtain training image collection；
Full convolutional neural networks FCN model is trained to obtain the identification model according to the training image collection；Wherein,
The FCN model has feature pyramid network FPN structure.
3. image text row detection method according to claim 2, which is characterized in that described according to the training image collection
FCN model is trained and includes:
Using the FCN model classification task and recurrence task are executed to the training image collection respectively, and is utilized respectively classification
Loss function and recurrence loss function optimize the FCN model.
4. image text row detection method according to claim 3, which is characterized in that using FCN model to the training
Image set executes classification task, is optimized using Classification Loss function to FCN model and includes:
Classification task is executed to obtain corresponding text probability confidence map to training image using the FCN model；
It is lost using the error that Classification Loss function calculates the probability confidence map, and according to error loss optimization
FCN model.
5. image text row detection method according to claim 3, which is characterized in that using FCN model to the training
Image set executes recurrence task, includes: using returning loss function and optimizing to FCN model
Recurrence task is executed to obtain the corresponding text including multiple line of text detection blocks to training image using the FCN model
This frame detection image；
It is lost using the error that the recurrence loss function calculates the text box detection image, and excellent according to error loss
Change the FCN model.
6. image text row detection method according to claim 3, which is characterized in that the training identification model packet
Multiple cycles of training are included, the method also includes:
The weighted value for configuring Classification Loss function described in each cycle of training successively declines；And
The weighted value for configuring recurrence loss function described in each cycle of training is successively promoted.
7. image text row detection method according to claim 3, which is characterized in that the Classification Loss function are as follows:
Wherein, the xiIt is described for the predicted value of pixel on the probability confidence mapFor the Ground of corresponding pixel points
The pixel value of Truth.
8. a kind of image text row detection device characterized by comprising
Image collection module, for obtaining image to be detected；
Probability confidence map obtains module, and the identification model for having trained obtains the text probability confidence of described image to be detected
Figure；
Testing result generation module, the line of text for obtaining described image to be detected according to the text probability confidence map detect
As a result.
9. a kind of storage medium, is stored thereon with computer program, which is characterized in that the computer program is executed by processor
Shi Shixian image text row detection method according to any one of claims 1 to 7.
10. a kind of electronic equipment characterized by comprising
Processor；And
Memory, for storing the executable instruction of the processor；
Wherein, the processor is configured to come described in any one of perform claim requirement 1~7 via the execution executable instruction
Image text row detection method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811435156.7A CN109583367A (en) | 2018-11-28 | 2018-11-28 | Image text row detection method and device, storage medium and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811435156.7A CN109583367A (en) | 2018-11-28 | 2018-11-28 | Image text row detection method and device, storage medium and electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109583367A true CN109583367A (en) | 2019-04-05 |
Family
ID=65925244
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811435156.7A Pending CN109583367A (en) | 2018-11-28 | 2018-11-28 | Image text row detection method and device, storage medium and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109583367A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111274985A (en) * | 2020-02-06 | 2020-06-12 | 咪咕文化科技有限公司 | Video text recognition network model, video text recognition device and electronic equipment |
| CN112135108A (en) * | 2020-09-27 | 2020-12-25 | 苏州科达科技股份有限公司 | Video stream subtitle detection method, system, device and storage medium |
| CN113033240A (en) * | 2019-12-09 | 2021-06-25 | 上海高德威智能交通系统有限公司 | Multi-line text recognition method, model training method, device, equipment and medium |
| CN113469931A (en) * | 2020-03-11 | 2021-10-01 | 北京沃东天骏信息技术有限公司 | Image detection model training, modification detection method, device and storage medium |
| CN113554026A (en) * | 2021-07-28 | 2021-10-26 | 广东电网有限责任公司 | Power equipment nameplate identification method and device and electronic equipment |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106897732A (en) * | 2017-01-06 | 2017-06-27 | 华中科技大学 | Multi-direction Method for text detection in a kind of natural picture based on connection word section |
| CN106980858A (en) * | 2017-02-28 | 2017-07-25 | 中国科学院信息工程研究所 | The language text detection of a kind of language text detection with alignment system and the application system and localization method |
| CN107316001A (en) * | 2017-05-31 | 2017-11-03 | 天津大学 | Small and intensive method for traffic sign detection in a kind of automatic Pilot scene |
| CN107609549A (en) * | 2017-09-20 | 2018-01-19 | 北京工业大学 | The Method for text detection of certificate image under a kind of natural scene |
| CN108108731A (en) * | 2016-11-25 | 2018-06-01 | 中移（杭州）信息技术有限公司 | Method for text detection and device based on generated data |
| CN108805131A (en) * | 2018-05-22 | 2018-11-13 | 北京旷视科技有限公司 | Text line detection method, apparatus and system |
-
2018
- 2018-11-28 CN CN201811435156.7A patent/CN109583367A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108108731A (en) * | 2016-11-25 | 2018-06-01 | 中移（杭州）信息技术有限公司 | Method for text detection and device based on generated data |
| CN106897732A (en) * | 2017-01-06 | 2017-06-27 | 华中科技大学 | Multi-direction Method for text detection in a kind of natural picture based on connection word section |
| CN106980858A (en) * | 2017-02-28 | 2017-07-25 | 中国科学院信息工程研究所 | The language text detection of a kind of language text detection with alignment system and the application system and localization method |
| CN107316001A (en) * | 2017-05-31 | 2017-11-03 | 天津大学 | Small and intensive method for traffic sign detection in a kind of automatic Pilot scene |
| CN107609549A (en) * | 2017-09-20 | 2018-01-19 | 北京工业大学 | The Method for text detection of certificate image under a kind of natural scene |
| CN108805131A (en) * | 2018-05-22 | 2018-11-13 | 北京旷视科技有限公司 | Text line detection method, apparatus and system |
Non-Patent Citations (1)
| Title |
|---|
| WENHAO HE等: "Deep Direct Regression for Multi-Oriented Scene Text Detection", 《ARXIV》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113033240A (en) * | 2019-12-09 | 2021-06-25 | 上海高德威智能交通系统有限公司 | Multi-line text recognition method, model training method, device, equipment and medium |
| CN113033240B (en) * | 2019-12-09 | 2023-05-02 | 上海高德威智能交通系统有限公司 | Multi-line text recognition method, model training method, device, equipment and medium |
| CN111274985A (en) * | 2020-02-06 | 2020-06-12 | 咪咕文化科技有限公司 | Video text recognition network model, video text recognition device and electronic equipment |
| CN111274985B (en) * | 2020-02-06 | 2024-03-26 | 咪咕文化科技有限公司 | Video text recognition system, video text recognition device and electronic equipment |
| CN113469931A (en) * | 2020-03-11 | 2021-10-01 | 北京沃东天骏信息技术有限公司 | Image detection model training, modification detection method, device and storage medium |
| CN112135108A (en) * | 2020-09-27 | 2020-12-25 | 苏州科达科技股份有限公司 | Video stream subtitle detection method, system, device and storage medium |
| CN113554026A (en) * | 2021-07-28 | 2021-10-26 | 广东电网有限责任公司 | Power equipment nameplate identification method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11734851B2 (en) | Face key point detection method and apparatus, storage medium, and electronic device | |
| CN111797893B (en) | Neural network training method, image classification system and related equipment | |
| CN108345890B (en) | Image processing method, device and related equipment | |
| CN109583367A (en) | Image text row detection method and device, storage medium and electronic equipment | |
| EP3968179A1 (en) | Place recognition method and apparatus, model training method and apparatus for place recognition, and electronic device | |
| WO2022068623A1 (en) | Model training method and related device | |
| CN111401516A (en) | Neural network channel parameter searching method and related equipment | |
| CN108229341A (en) | Sorting technique and device, electronic equipment, computer storage media, program | |
| CN110084313A (en) | A method of generating object detection model | |
| CN109934173A (en) | Expression recognition method, device and electronic equipment | |
| CN113705769A (en) | Neural network training method and device | |
| CN109902716B (en) | Training method for alignment classification model and image classification method | |
| CN111414915B (en) | Character recognition method and related equipment | |
| CN110163205A (en) | Image processing method, device, medium and calculating equipment | |
| WO2022111387A1 (en) | Data processing method and related apparatus | |
| CN115376518B (en) | Voiceprint recognition method, system, equipment and medium for real-time noise big data | |
| CN115510795A (en) | Data processing method and related device | |
| WO2024083121A1 (en) | Data processing method and apparatus | |
| CN114445684A (en) | Method, device and equipment for training lane line segmentation model and storage medium | |
| CN113554653A (en) | Semantic segmentation method for long-tail distribution of point cloud data based on mutual information calibration | |
| CN117218300A (en) | Three-dimensional model construction method, three-dimensional model construction training method and device | |
| CN111353577B (en) | Multi-task-based cascade combination model optimization method and device and terminal equipment | |
| CN110826726B (en) | Target processing method, target processing device, target processing apparatus, and medium | |
| CN113705404A (en) | Face detection method facing embedded hardware | |
| Loong et al. | Image‐based structural analysis for education purposes: A proof‐of‐concept study |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20190704 Address after: 311215 Room 102, 6 Blocks, C District, Qianjiang Century Park, Xiaoshan District, Hangzhou City, Zhejiang Province Applicant after: Hangzhou Yixian Advanced Technology Co., Ltd. Address before: 310052 Building No. 599, Changhe Street Network Business Road, Binjiang District, Hangzhou City, Zhejiang Province, 4, 7 stories Applicant before: NetEase (Hangzhou) Network Co., Ltd. |
|
| TA01 | Transfer of patent application right | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190405 |
|
| RJ01 | Rejection of invention patent application after publication |