CN111178447B - Model compression method, image processing method and related device - Google Patents
Model compression method, image processing method and related device Download PDFInfo
- Publication number
- CN111178447B CN111178447B CN201911421331.1A CN201911421331A CN111178447B CN 111178447 B CN111178447 B CN 111178447B CN 201911421331 A CN201911421331 A CN 201911421331A CN 111178447 B CN111178447 B CN 111178447B
- Authority
- CN
- China
- Prior art keywords
- training model
- model
- data
- parameter
- trained
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 89
- 230000006835 compression Effects 0.000 title claims abstract description 46
- 238000007906 compression Methods 0.000 title claims abstract description 46
- 238000003672 processing method Methods 0.000 title claims description 10
- 238000012549 training Methods 0.000 claims abstract description 309
- 238000012545 processing Methods 0.000 claims abstract description 43
- 238000004590 computer program Methods 0.000 claims description 15
- 238000012821 model calculation Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 description 7
- 230000008569 process Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000005516 engineering process Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 230000007613 environmental effect Effects 0.000 description 5
- 238000013473 artificial intelligence Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 230000004927 fusion Effects 0.000 description 3
- 238000002372 labelling Methods 0.000 description 3
- 230000008447 perception Effects 0.000 description 3
- 230000003190 augmentative effect Effects 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000013434 data augmentation Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000002452 interceptive effect Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 239000004576 sand Substances 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 230000001629 suppression Effects 0.000 description 2
- 230000003321 amplification Effects 0.000 description 1
- 230000003416 augmentation Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000007306 turnover Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
-
- 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
- G06N3/08—Learning methods
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)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the application provides a method and a related device for model compression and image processing, wherein the method for model compression comprises the following steps: determining a first parameter set according to characteristic parameters of data to be trained; adjusting an initial training model according to the first parameter set to obtain a reference training model, wherein the operand of the reference training model is smaller than that of the initial training model; acquiring a second parameter set; and compressing the reference training model according to the second parameter set to obtain a target training model, wherein the precision of the target training model and the precision of the reference training model belong to a preset precision range, and the applicability of the network model in application can be improved.
Description
Technical Field
The present disclosure relates to the field of data processing technologies, and in particular, to a model compression method, an image processing method, and a related device.
Background
With the development of artificial intelligence technology and the great breakthrough of the artificial intelligence technology in computer vision, many industries apply deep learning technology to perform perception detection and identification on data acquired by a camera. In specific application scenarios, such as the field of automatic driving, the sensing algorithm is particularly important, and as an "eye" of a vehicle, the sensing algorithm needs to accurately and efficiently sense surrounding environmental objects and then transmit the surrounding environmental objects to a subsequent decision-making module to perform related actions. However, for the application of artificial intelligence technology on automobiles, the application of real carts for experiments is obviously dangerous and impractical, so that the development of automatic driving carts based on embedded platforms is required for various experiments, the cost is low and the safety is high, and the application and practice of the artificial intelligence technology in the automatic driving field are high.
However, when the automatic driving trolley perception algorithm based on the embedded platform is developed and deployed, the applicability of the existing network model in application is poor due to the small calculation capability and storage space of the automatic driving trolley perception algorithm.
Disclosure of Invention
The embodiment of the application provides a model compression method, an image processing method and a related device.
A first aspect of an embodiment of the present application provides a method for compressing a model, the method including:
determining a first parameter set according to characteristic parameters of data to be trained;
adjusting an initial training model according to the first parameter set to obtain a reference training model, wherein the operand of the reference training model is smaller than that of the initial training model;
acquiring a second parameter set;
and compressing the reference training model according to the second parameter set to obtain a target training model, wherein the precision of the target training model and the precision of the reference training model belong to a preset precision range.
In the embodiment of the disclosure, the initial training model is adjusted through the first parameter set to obtain the reference training model, the reference training model is compressed according to the acquired second parameter set to obtain the target training model, the reference training model can be compressed through the parameter set, the size of the model is reduced, and therefore the training model can be applied to equipment with smaller resources, and the practicability of the training model in application is improved.
With reference to the first aspect, in one possible implementation manner, the first parameter set includes a first sub-parameter and a second sub-parameter, the feature parameter includes an aspect ratio distribution and an area distribution, and the determining the first parameter set according to the feature parameter of the data to be trained includes:
determining the first sub-parameter according to the area distribution;
and determining the second sub-parameter according to the aspect ratio distribution.
In the embodiment of the disclosure, the first sub-parameter and the second sub-parameter are determined through the area distribution and the aspect ratio distribution, which can accurately reflect the characteristics of the data, and can improve the accuracy in determining the first parameter set.
With reference to the first aspect, in a possible implementation manner, the determining the first sub-parameter according to the area distribution includes:
dividing the area distribution into M sections, wherein M is a positive integer;
acquiring the average value of each of the M sections;
and determining the first sub-parameter according to the average value.
With reference to the first aspect, in a possible implementation manner, the determining the second sub-parameter according to the aspect ratio distribution includes:
Dividing the aspect ratio distribution into N sections, wherein N is a positive integer;
acquiring the average value of each of the N sections;
and determining the mean value as the second sub-parameter.
With reference to the first aspect, in a possible implementation manner, the acquiring the second parameter set includes:
and determining the second parameter set according to the preset model precision and model calculation amount.
In the embodiment of the disclosure, the second parameter set is determined through the preset model precision and model calculation amount, so that the model compressed through the second parameter set can meet the preset model precision and model calculation amount, and the accuracy of the model in compression is improved.
With reference to the first aspect, in one possible implementation manner, the second parameter set includes a convolution kernel size, a channel number, and an accuracy parameter.
With reference to the first aspect, in one possible implementation manner, the method further includes:
training the target training model through the first data set to be trained to obtain a first training model;
and training the first training model through the second data set to be trained to obtain a second training model.
With reference to the first aspect, in one possible implementation manner, before the training, by the first to-be-trained data set, the training the target training model, to obtain a first training model, the method further includes:
Obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained.
In the embodiment of the disclosure, the method for amplifying the data adopts a mode of fusing the object to be detected and the image in the data set, and compared with the prior art, the method for amplifying the image generally adopts a mode of adding gaussian noise, and the image is subjected to overturn, rotation, miscut and the like, so that the accuracy of the image during the amplification can be improved.
With reference to the first aspect, in one possible implementation manner, before the training the first training model by using the second data set to be trained to obtain a second training model, the method includes:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained.
With reference to the first aspect, in a possible implementation manner, the training the first training model by using the second training data set to obtain a second training model includes:
Acquiring the weight of each type of data in the second data set to be trained;
training the first training model according to each type of data in the second data set to be trained and the weight of each type of data to obtain a second training model; and/or the number of the groups of groups,
the training of the target training model through the first data set to be trained to obtain a first training model comprises the following steps:
acquiring a weight of each type of data in a first data set to be trained;
and training the target training model according to each type of data and the weight of each type of data in the first data set to be trained to obtain a first training model.
In the embodiment of the disclosure, each type of data in the first data set to be trained and/or the second data set to be trained is classified by weight, so that the data can be balanced, and the accuracy of model training is improved.
A second aspect of embodiments of the present application provides an image processing method, including:
receiving an image to be processed;
inputting the image to be processed into an image processing model for processing to obtain target data of the target image;
wherein the image processing model is a target training model obtained by the model compression method described in the first aspect, or the image processing model is a second training model obtained by the model compression method described in the first aspect.
A third aspect of embodiments of the present application provides a model compression apparatus, the apparatus including:
the determining unit is used for determining a first parameter set according to the characteristic parameters of the data to be trained;
the adjusting unit is used for adjusting the initial training model according to the first parameter set to obtain a reference training model, and the operation amount of the reference training model is smaller than that of the initial training model;
an acquisition unit configured to acquire a second parameter set;
the compression unit is used for compressing the reference training model according to the second parameter set to obtain a target training model, and the precision of the target training model and the precision of the reference training model belong to a preset precision range.
With reference to the third aspect, in a possible implementation manner, the first parameter set includes a first sub-parameter and a second sub-parameter, the feature parameter includes an aspect ratio distribution and an area distribution, and the determining unit is configured to:
determining the first sub-parameter according to the area distribution;
and determining the second sub-parameter according to the aspect ratio distribution.
With reference to the third aspect, in a possible implementation manner, in the determining the first sub-parameter according to the area distribution, the determining unit is configured to:
Dividing the area distribution into M sections, wherein M is a positive integer;
acquiring the average value of each of the M sections;
and determining the first sub-parameter according to the average value.
With reference to the third aspect, in a possible implementation manner, in the determining the second sub-parameter according to the aspect ratio distribution, the determining unit is configured to:
dividing the aspect ratio distribution into N sections, wherein N is a positive integer;
acquiring the average value of each of the N sections;
and determining the mean value as the second sub-parameter.
With reference to the third aspect, in one possible implementation manner, the acquiring unit is configured to:
and determining the second parameter set according to the preset model precision and model calculation amount.
With reference to the third aspect, in one possible implementation manner, the second parameter set includes a convolution kernel size, a channel number, and an accuracy parameter.
With reference to the third aspect, in one possible implementation manner, the apparatus is further configured to:
training the target training model through the first data set to be trained to obtain a first training model;
and training the first training model through the second data set to be trained to obtain a second training model.
With reference to the third aspect, in one possible implementation manner, the apparatus is further configured to:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained.
With reference to the third aspect, in one possible implementation manner, the apparatus is further configured to:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained.
With reference to the third aspect, in a possible implementation manner, in the training the first training model by using the second training data set to obtain a second training model, the apparatus is further configured to:
acquiring the weight of each type of data in the second data set to be trained;
training the first training model according to each type of data in the second data set to be trained and the weight of each type of data to obtain a second training model; and/or the number of the groups of groups,
in the aspect of training the target training model through the first data set to be trained to obtain a first training model, the apparatus is further configured to:
Acquiring a weight of each type of data in a first data set to be trained;
and training the target training model according to each type of data and the weight of each type of data in the first data set to be trained to obtain a first training model.
A fourth aspect of the embodiments of the present application provides an image processing apparatus, including:
a receiving unit for receiving an image to be processed;
the processing unit is used for inputting the image to be processed into an image processing model for processing to obtain target data of the target image;
wherein the image processing model is a target training model obtained by the model compression method described in the first aspect, or the image processing model is a second training model obtained by the model compression method described in the first aspect.
A fifth aspect of the embodiments of the present application provides a terminal comprising a processor and a memory, the processor and the memory being interconnected, wherein the memory is adapted to store a computer program comprising program instructions, the processor being configured to invoke the program instructions to execute the step instructions as in the first and/or second aspects of the embodiments of the present application.
A sixth aspect of the embodiments of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program causes a computer to perform some or all of the steps as described in the first aspect and/or the second aspect of the embodiments of the present application.
A seventh aspect of embodiments of the present application provides a computer program product, wherein the computer program product comprises a computer readable storage medium storing a computer program operable to cause a computer to perform part or all of the steps as described in the first and/or second aspects of embodiments of the present application. The computer program product may be a software installation package.
These and other aspects of the invention will be more readily apparent from the following description of the embodiments.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic flow chart of a model compression method according to an embodiment of the present application;
FIG. 2 is a flow chart of another method for compressing models according to an embodiment of the present application;
FIG. 3 is a flow chart of another method for compressing models according to an embodiment of the present application;
fig. 4 is a schematic structural diagram of a terminal according to an embodiment of the present application;
fig. 5 is a schematic structural diagram of a model compression device according to an embodiment of the present application;
fig. 6 is a schematic structural diagram of an image processing apparatus according to an embodiment of the present application.
Detailed Description
The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure.
The terms first, second and the like in the description and in the claims of the present application and in the above-described figures, are used for distinguishing between different objects and not for describing a particular sequential order. Furthermore, the terms "comprise" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those listed steps or elements but may include other steps or elements not listed or inherent to such process, method, article, or apparatus.
Reference in the specification to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearances of such phrases in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of skill in the art will explicitly and implicitly understand that the embodiments described herein may be combined with other embodiments.
The embodiment of the application can be applied to hardware terminals, including terminals, servers, automatic driving trolleys and the like, for example, the automatic driving trolleys based on embedded type or automatic driving trolleys based on other mechanisms, and can also be realized by executing executable codes through a processor. For convenience of description, the above-mentioned apparatuses are collectively referred to as an electronic device.
The first parameter set in this embodiment of the present application includes a first sub-parameter and a second sub-parameter, where the first sub-parameter may be an anchor frame area scaling array Anchor_scale, and the second sub-parameter may be an anchor frame aspect ratio array anchors_ratios.
The aspect ratio distribution and the area distribution described in the embodiments of the present application are the width-height distribution and the area distribution of the annotation frame of the data to be trained.
Referring to fig. 1, fig. 1 is a schematic flow chart of a model compression method according to an embodiment of the present application. As shown in fig. 1, the model compression method includes steps 101-104, specifically as follows:
101. and determining a first parameter set according to the characteristic parameters of the data to be trained.
The data to be trained may comprise data for training a training model, for example an environmental image of the autonomous driving trolley, which may be, for example, an image of a sand table, which may be understood as the environment in which the autonomous driving trolley is driving. The automatic driving trolley can run in a sand table.
The characteristic parameters can comprise the quantity ratio, the length-width ratio distribution, the area distribution and the like of various annotation frames of the environment image. The labeling frame may be understood as a labeling frame for labeling the content of interest in the environmental image, for example, a standard frame for an object to be detected in the environmental image, where the object to be detected may be, for example, an interfering object (an object similar to a traffic sign, etc.), or may be a non-interfering object (a traffic sign), etc.
The first parameter set is a parameter characterizing an initial training model operand, etc., for example, in one possible implementation, the first parameter set includes a first sub-parameter, which may be an anchor frame area scaling array anchor_scale, and a second sub-parameter, which may be an anchor frame aspect ratio array anchors_ratios.
102. And adjusting the initial training model according to the first parameter set to obtain a reference training model.
The initial training model may be understood as a training model obtained by training the data to be trained. For example, a model obtained by training an environment image of an automated guided vehicle. Of course, in other embodiments, the initial training model may also be an untrained model.
The adjusting of the initial training model may include replacing corresponding parameters in the initial training model with parameters in the first set of parameters to complete the adjusting of the initial training model. The operand of the reference training model is smaller than that of the initial training model.
103. A second set of parameters is obtained.
The second parameter set can be obtained according to the preset model precision and model calculation amount, so that the model is compressed through the second parameter set, and the precision and calculation amount of the compressed model can be ensured.
The second set of parameters is parameters characterizing the compression amplitude of the reference model. In one possible implementation, the second parameter set includes pre_nms (the number of target candidate frames selected before the non-maximum suppression processing) and post_nms (the number of target candidate frames selected after the non-maximum suppression processing), where pre_nms and post_nms are parameters that characterize the model accuracy, and where pre_nms and post_nms increase, the model accuracy also increases.
104. And compressing the reference training model according to the second parameter set to obtain the target training model.
The accuracy of the target training model and the accuracy of the reference training model belong to a preset accuracy range, and can be specifically understood as: the accuracy of the target training model is very small in difference from the accuracy of the reference training model, and the preset accuracy range is set through experience values or historical data. When the reference training model is compressed according to the second parameter set, the second parameter set may be determined as a corresponding parameter set in the reference training model, so as to complete the compression of the reference training model.
In the embodiment of the disclosure, the initial training model is adjusted through the first parameter set to obtain the reference training model, the reference training model is compressed according to the acquired second parameter set to obtain the target training model, the reference training model can be compressed through the parameter set, the size of the model is reduced, and therefore the training model can be applied to equipment with smaller resources, and the practicability of the training model in application is improved.
In a possible embodiment, the first parameter set comprises a first sub-parameter and a second sub-parameter, the feature parameters comprise an aspect ratio distribution and an area distribution, and one possible method for determining the first parameter set according to the feature parameters of the data to be trained comprises the steps A1-A2, in particular as follows:
A1, determining a first sub-parameter according to the area distribution;
a2, determining a second sub-parameter according to the aspect ratio distribution.
The aspect ratio distribution and the area distribution are the wide-height distribution and the area distribution of the annotation frame of the data to be trained.
In one possible implementation, the first sub-parameter is determined from the mean of the area distribution and the second sub-parameter is determined from the mean of the aspect ratio distribution.
The first sub-parameter and the second sub-parameter are determined through the average value, so that accuracy in determining the first sub-parameter and the second sub-parameter can be improved.
In one possible implementation manner, the first sub-parameter may be further determined according to a mean value of the main concentrated area of the area distribution, where the first sub-parameter may specifically be: and removing the maximum value and the minimum value in the area distribution, or removing a small amount of area values close to the maximum value and the minimum value at the same time, determining the average value of the rest area values as a first sub-parameter, wherein the small amount of area values can be understood as values of which the difference value between the small amount of area values and the maximum value or the minimum value is in a preset range, and the preset range is set through experience values or historical data. In determining the second sub-parameter, reference may also be made to the method for determining the first sub-parameter described above.
In one possible embodiment, one possible method for determining the first sub-parameter from the area distribution comprises steps B1-B3, in particular as follows:
b1, dividing the area distribution into M sections, wherein M is a positive integer;
b2, obtaining the average value of each section in the M sections;
and B3, determining a first sub-parameter according to the average value.
The lengths of the M segments may be equal or unequal.
The method for determining the first subparameter according to the average value of each segment can be as follows: for example, assuming that the mean value a is any one of the mean values in the M segments, the first sub-parameter may be expressed as:
c=b/d, equation (1)
Where b is the positive square root of a, c is the first sub-parameter, and d is 2 times the sampling multiple of the initial training model.
In the embodiment of the disclosure, the accuracy of determining the first sub-parameter can be improved by dividing the area distribution into M segments and determining the first sub-parameter according to the average value of each segment.
In one possible embodiment, a possible method for determining the second sub-parameter according to the aspect ratio distribution comprises the steps C1-C3, in particular as follows:
c1, dividing the aspect ratio distribution into N sections, wherein N is a positive integer;
c2, acquiring the average value of each of the N sections;
And C3, determining the average value as a second sub-parameter.
The lengths of the N segments may be equal or unequal.
In the embodiment of the disclosure, the accuracy of determining the second sub-parameter can be improved by dividing the area distribution into N segments and determining the average value of each segment as the second sub-parameter.
One specific example may be: the area and the aspect ratio of all groups of data in the statistical training set are divided into 5 sections and 2 sections respectively, when the area distribution is divided into 5 sections, the length of each section can be equal or unequal, and when the aspect ratio distribution is divided into 2 sections, the length of each section can be the same or different, and the average value of each section is obtained, and the Anchor_scale and the anchor_ratio can be obtained according to the average values. The specific method comprises the following steps: assuming that one of the average values of the areas is a, b=a's evolution, c=b/32 (dividing by 32 instead of 16 because the input width and height of the initial training model are half of the original image), c is the anchor_scale corresponding to the area a; and each average of the aspect ratios is the desired anchor_ratio.
In one possible embodiment, the second parameter set includes a convolution kernel size, a number of channels, and an accuracy parameter, and one possible method for determining the second parameter set according to a preset model accuracy and a model calculation amount includes:
The method comprises the steps of changing the size, the number of channels and precision parameters of a convolution kernel through a reference training model to obtain a plurality of compressed models, calculating an image to be processed, and determining the compressed model with calculated amount smaller than the reference training model and the difference value between model precision and model precision of the reference training model smaller than a preset difference value as a target training model. The preset difference value may be set by an empirical value or historical data. For example, a plurality of pre-selected combinations (i.e., a second parameter set) of the convolution kernel size, the number of channels and the precision parameters may be predetermined according to experience or historical training data, a plurality of reference training models corresponding to different pre-selected combinations may be obtained by varying the convolution kernel size, the number of channels and the precision parameters a plurality of times, the plurality of reference training models may be trained, the minimum calculation amount may be selected according to the calculated amount and the precision of the trained models, and/or a model with a precision that is smaller than a preset difference and the minimum model precision of the reference training models may be selected as the compressed model.
The precision parameters may include pre_nms and post_nms, a comparative example of a reference training model and a target training model is as follows: pre_nms=4000 and post_nms=300 with reference to the training model before compression; after compression, the target training model, pre_nms=400 and post_nms=30, is compressed by 10 times, but the difference between the model accuracy is smaller than the preset difference, so that the calculated amount of the training model can be greatly reduced.
In the case of compressing the model, only a specific layer is compressed, but not all layers, and the specific layer may be understood as a layer where redundancy parameters occur, for example, a layer having an anchor frame area scaling factor.
In one possible embodiment, the training data may also be augmented, and one possible method of augmentation includes steps D1-D2, as follows:
d1, obtaining a mask image of an object to be detected in data to be trained;
and D2, fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained.
The second data set may also be augmented, as follows:
e1, acquiring a mask image of an object to be detected in data to be trained;
and E2, fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained.
When the mask image of the object to be detected is fused with the image in the first data set and the image in the second data set, the fusion can be performed in a random pasting mode, and other fusion modes can be adopted.
The first data set may include a network general training image set such as COCO image data, and the second data set may be a terminal view angle actually applied, for example, a trolley view angle, a large number of collected background images, and as many objects easy to be detected by mistake appear in the background, for example, the objects easy to be detected by mistake may be interference objects, and have higher similarity with the objects to be detected.
In one possible embodiment, a possible method for fusing a mask map of an object to be detected with an image in a first data set to obtain the first data set to be trained includes:
and randomly pasting the mask image of the object to be detected to the image in the first data set to obtain a first data set to be trained.
When the mask patterns are randomly pasted, the number and the positions of the pasting can be random, namely, the number of the mask patterns of the object to be detected in different images can be different, the mask patterns can be the same, and the pasting positions can be the same or different.
Of course, in the case of data augmentation, conventional data augmentation methods, such as image rotation, folding, noise increase, and the like, may be combined.
In the embodiment of the disclosure, the first data set and the second data set are subjected to data fusion to obtain the corresponding first data set to be trained and the second data set to be trained fused with the terminal visual angle image which is actually applied, and compared with the training data set which is obtained by adopting methods of image rotation, turnover, noise increase and the like in the existing scheme, the first data set to be trained and/or the second data set to be trained are used for model training, so that the problems of easy fitting and low generalization can be effectively relieved, and the situations of false detection and omission detection are reduced.
In one possible embodiment, the present solution may further include the following method:
e1, training a target training model through a first data set to be trained to obtain a first training model;
and E2, training the first training model through a second data set to be trained to obtain a second training model.
The first training model may be trained with the data set to be trained while the first training model is trained with the second data set to be trained.
In the embodiment of the disclosure, through two times of training, the accuracy of the image recognition of the second training model can be improved.
In one possible embodiment, a possible method for training a first training model by using a second training data set to obtain a second training model includes steps F1-F2, specifically as follows:
f1, acquiring a weight of each type of data in a second data set to be trained;
and F2, training the first training model according to each type of data and the weight of each type of data in the second data set to be trained to obtain a second training model.
The weight of each type of data in the second data set to be trained can be obtained through the corresponding relation between the preset type and the weight. The weight is a weight determined for the number of data of each class. When training the first training model according to each type of data in the second data set to be trained and the weight of each type of data, counting the number of each type of data in the second data set to be trained, determining the weight of each type of data according to the number of each type of data, and when calculating the model loss in the training process, weighting each type according to the corresponding weight. The method comprises the steps of setting a larger weight for a few categories, so that the loss of the training model is focused on the data with few categories, and the model with the parameters adjusted can be used for successfully identifying sparse data more easily.
In one possible embodiment, a possible method for training a target training model by using a first training data set to obtain a first training model includes steps F3-F4, specifically as follows:
f3, obtaining the weight of each type of data in the first data set to be trained;
and F4, training the target training model according to each type of data and the weight of each type of data in the first data set to be trained to obtain a first training model.
The weight of each type of data in the first data set to be trained can be obtained through the corresponding relation between the preset type and the weight. The specific weight determining process and training process may refer to the above steps F1-F2, and will not be described herein. By setting the corresponding weight for each type of data, the problems that some types of objects are easy to misdetect and miss-detect due to unbalanced types can be reduced, and the accuracy in training is improved.
The embodiment of the application also provides a model compression method of the trolley, which comprises the following steps: training the original model through the data to be trained to obtain an initial training model, wherein the original training model does not have related data such as weight.
Because the method is applied to the trolley, the operation capability of the trolley is limited, in order to meet the requirement of deployment on the trolley, an initial training model is required to be compressed, when the initial training model is compressed, a first parameter set is determined according to characteristic parameters of data to be trained, the first parameter set comprises a first subparameter and a second subparameter, the first subparameter can be an anchor frame area scaling array anchor_scale, the second subparameter can be an anchor frame aspect ratio array anchors_ratio, and the characteristic parameters can be aspect ratio distribution and area distribution of a marking frame in the data to be trained; and determining the first parameter set as a corresponding parameter in the initial training model, and carrying out parameter replacement to obtain a reference training model, wherein the operation amount of the reference training model is smaller than that of the initial training model.
In order to compress the model further, the model is deployed on the trolley better, the model is compressed again, a second parameter set is obtained during compression, the second parameter set comprises convolution kernel size, channel number and precision parameters, the reference training model is compressed according to the second parameter set, the target training model is obtained, and the precision of the target training model and the precision of the reference training model belong to a preset precision range, so that the initial training model is compressed. After the target training model is obtained, the target training model is deployed into a trolley, and then the target training model is trained through a first data set to be trained and a second data set to be trained, so that a second training model is finally obtained. A brief description of a scenario taking a trolley as an example is provided herein, where a specific implementation may refer to an implementation described in the foregoing embodiment, which is not described herein again.
The embodiment of the application also provides an image processing method, which comprises the following steps:
receiving an image to be processed;
inputting the image to be processed into an image processing model for processing to obtain target data of the target image;
The image processing model is a target training model obtained by the model compression method described in the above embodiment, or the image processing model is a second training model obtained by the model compression method described in the above embodiment.
The image to be processed may be an image to be processed acquired by a trolley, for example, specifically may be: the image to be processed needs to be subjected to lane line segmentation, and the target data can be lane line data after the lane line segmentation, and the like. This is merely illustrative and of course, an image of a cart for processing is also possible.
Referring to fig. 2, fig. 2 is a flow chart of another model compression method according to an embodiment of the present application. As shown in fig. 2, the first parameter set includes a first sub-parameter and a second sub-parameter, the feature parameters include an aspect ratio distribution and an area distribution, and the model compression method includes steps 201 to 205, which are specifically as follows:
201. determining a first sub-parameter according to the area distribution;
202. determining a second sub-parameter according to the aspect ratio distribution;
203. adjusting the initial training model according to the first sub-parameter and the second sub-parameter to obtain a reference training model;
204. acquiring a second parameter set;
205. And compressing the reference training model according to the second parameter set to obtain the target training model.
In the embodiment of the disclosure, the first sub-parameter and the second sub-parameter are determined through the area distribution and the aspect ratio distribution, and the area distribution and the aspect ratio distribution can accurately reflect the characteristics of the data, so that the accuracy in determining the first parameter set can be improved, and the accuracy in determining the target training model is further improved.
Referring to fig. 3, fig. 3 is a flow chart of another model compression method according to an embodiment of the present application. As shown in fig. 3, the model compression method includes steps 301-309, which are specifically as follows:
301. determining a first parameter set according to characteristic parameters of data to be trained;
wherein the first set of parameters comprises a first sub-parameter and a second sub-parameter, and the feature parameters comprise an aspect ratio distribution and an area distribution.
302. Adjusting the initial training model according to the first parameter set to obtain a reference training model;
303. determining a second parameter set according to the preset model precision and model calculation amount;
the second set of parameters includes convolution kernel size, number of channels, and accuracy parameters.
304. Compressing the reference training model according to the second parameter set to obtain a target training model;
305. Obtaining a mask image of an object to be detected in data to be trained;
306. fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained;
307. fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained;
308. training a target training model through a first data set to be trained to obtain a first training model;
309. and training the first training model through the second data set to be trained to obtain a second training model.
In the embodiment of the disclosure, the data is processed through the weight corresponding to each type of data, so that the problem that some types of objects are easy to misdetect and miss-detect due to unbalanced types can be reduced, and the accuracy of training to obtain the second training model is improved.
Referring to fig. 4, fig. 4 is a schematic structural diagram of a terminal provided in an embodiment of the present application, where, as shown in fig. 4, the terminal includes a processor and a memory, and may further include an input device, an output device, where the processor, the input device, the output device, and the memory are connected to each other, where the memory is configured to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions, and the program includes instructions for executing the following steps;
Determining a first parameter set according to characteristic parameters of data to be trained;
adjusting an initial training model according to the first parameter set to obtain a reference training model, wherein the operand of the reference training model is smaller than that of the initial training model;
acquiring a second parameter set;
and compressing the reference training model according to the second parameter set to obtain a target training model, wherein the precision of the target training model and the precision of the reference training model belong to a preset precision range.
In one possible embodiment, the first parameter set includes a first sub-parameter and a second sub-parameter, the feature parameters include an aspect ratio distribution and an area distribution, and the determining the first parameter set according to the feature parameters of the data to be trained includes:
determining the first sub-parameter according to the area distribution;
and determining the second sub-parameter according to the aspect ratio distribution.
In one possible embodiment, the determining the first sub-parameter according to the area distribution includes:
dividing the area distribution into M sections, wherein M is a positive integer;
acquiring the average value of each of the M sections;
And determining the first sub-parameter according to the average value.
In a possible implementation manner, the determining the second sub-parameter according to the aspect ratio distribution includes:
dividing the aspect ratio distribution into N sections, wherein N is a positive integer;
acquiring the average value of each of the N sections;
and determining the mean value as the second sub-parameter.
In one possible implementation, the acquiring the second parameter set includes:
and determining the second parameter set according to the preset model precision and model calculation amount.
In one possible embodiment, the second set of parameters includes a convolution kernel size, a number of channels, and a precision parameter.
In one possible embodiment, the method further comprises:
training the target training model through the first data set to be trained to obtain a first training model;
and training the first training model through the second data set to be trained to obtain a second training model.
In one possible implementation manner, before the training the target training model by using the first data set to be trained to obtain a first training model, the method further includes:
Obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained.
In one possible implementation manner, before the training the first training model through the second data set to be trained to obtain a second training model, the training method includes:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained.
In one possible implementation manner, the training the first training model through the second training data set to obtain a second training model includes:
acquiring the weight of each type of data in the second data set to be trained;
training the first training model according to each type of data in the second data set to be trained and the weight of each type of data to obtain a second training model; and/or the number of the groups of groups,
the training of the target training model through the first data set to be trained to obtain a first training model comprises the following steps:
Acquiring a weight of each type of data in a first data set to be trained;
and training the target training model according to each type of data and the weight of each type of data in the first data set to be trained to obtain a first training model.
The above further program includes instructions for performing the steps of:
receiving an image to be processed;
inputting the image to be processed into an image processing model for processing to obtain target data of the target image;
the image processing model is a target training model obtained by the model compression method in the above embodiment, or the image processing model is a second training model obtained by the model compression method in the above embodiment.
The specific execution process of the above steps may refer to the execution process of the model compression method or the image processing method provided in the embodiments of the present disclosure, which is not described herein again.
The foregoing description of the embodiments of the present application has been presented primarily in terms of a method-side implementation. It will be appreciated that, in order to achieve the above-mentioned functions, the terminal includes corresponding hardware structures and/or software modules for performing the respective functions. Those of skill in the art will readily appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein may be embodied as hardware or a combination of hardware and computer software. Whether a function is implemented as hardware or computer software driven hardware depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
The embodiment of the application may divide the functional units of the terminal according to the above method example, for example, each functional unit may be divided corresponding to each function, or two or more functions may be integrated in one processing unit. The integrated units may be implemented in hardware or in software functional units. It should be noted that, in the embodiment of the present application, the division of the units is schematic, which is merely a logic function division, and other division manners may be implemented in actual practice.
In accordance with the foregoing, referring to fig. 5, fig. 5 is a schematic structural diagram of a model compressing apparatus according to an embodiment of the present application. As shown in fig. 5, the apparatus includes:
a determining unit 501, configured to determine a first parameter set according to a feature parameter of data to be trained;
the adjusting unit 502 is configured to adjust the initial training model according to the first parameter set to obtain a reference training model, where an operand of the reference training model is smaller than that of the initial training model;
an obtaining unit 503, configured to obtain a second parameter set;
the compression unit 504 is configured to compress the reference training model according to the second parameter set to obtain a target training model, where the accuracy of the target training model and the accuracy of the reference training model belong to a preset accuracy range.
In a possible embodiment, the first parameter set comprises a first sub-parameter and a second sub-parameter, the feature parameters comprise an aspect ratio distribution and an area distribution, and the determining unit 501 is configured to:
determining a first sub-parameter according to the area distribution;
and determining a second sub-parameter according to the aspect ratio distribution.
In a possible embodiment, the determining unit 501 is configured to, in determining the first sub-parameter from the area distribution:
dividing the area distribution into M sections, wherein M is a positive integer;
acquiring the average value of each of the M sections;
and determining a first sub-parameter according to the average value.
In a possible implementation, the determining unit 501 is configured to, in determining the second sub-parameter according to the aspect ratio distribution:
dividing the aspect ratio distribution into N sections, wherein N is a positive integer;
acquiring the average value of each of the N sections;
the mean value is determined as the second sub-parameter.
In one possible implementation, the obtaining unit 503 is configured to:
and determining a second parameter set according to the preset model precision and model calculation amount.
In one possible implementation, the second set of parameters includes a convolution kernel size, a number of channels, and a precision parameter.
In one possible embodiment, the apparatus is further for:
Training the target training model through the first data set to be trained to obtain a first training model;
and training the first training model through the second data set to be trained to obtain a second training model.
In one possible embodiment, the apparatus is further for:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained.
In one possible embodiment, the apparatus is further for:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained.
In one possible implementation manner, in the aspect of fusing the mask map of the object to be detected with the image in the first data set to obtain the first data set to be trained, the apparatus is further configured to:
and randomly pasting the mask map of the object to be detected to the image in the first data set to obtain the first data set to be trained.
In a possible implementation manner, in the aspect that the first training model is trained through the second training data set to obtain a second training model, the apparatus is further configured to:
Acquiring the weight of each type of data in the second data set to be trained;
training the first training model according to each type of data in the second data set to be trained and the weight of each type of data to obtain a second training model; and/or the number of the groups of groups,
the training of the target training model through the first data set to be trained to obtain a first training model comprises the following steps:
acquiring a weight of each type of data in a first data set to be trained;
and training the target training model according to each type of data and the weight of each type of data in the first data set to be trained to obtain a first training model.
The embodiment of the application also provides a structural schematic diagram of the image processing device. As shown in fig. 6, the image processing apparatus includes:
a receiving unit 601, configured to receive an image to be processed;
the processing unit 602 is configured to input the image to be processed into an image processing model for processing, so as to obtain target data of the target image;
the image processing model is a target training model obtained by the model compression method described in the above embodiment, or the image processing model is a second training model obtained by the model compression method described in the above embodiment.
The present application also provides a computer storage medium storing a computer program that causes a computer to execute part or all of the steps of any one of the model compression method and/or the image processing method described in the above method embodiments.
Embodiments of the present application also provide a computer program product comprising a computer-readable storage medium storing a computer program that causes a computer to perform part or all of the steps of any one of the model compression methods and/or the image processing methods described in the method embodiments above.
It should be noted that, for simplicity of description, the foregoing method embodiments are all expressed as a series of action combinations, but it should be understood by those skilled in the art that the present application is not limited by the order of actions described, as some steps may be performed in other order or simultaneously in accordance with the present application. Further, those skilled in the art will also appreciate that the embodiments described in the specification are all preferred embodiments, and that the acts and modules referred to are not necessarily required in the present application.
In the foregoing embodiments, the descriptions of the embodiments are emphasized, and for parts of one embodiment that are not described in detail, reference may be made to related descriptions of other embodiments.
In the several embodiments provided in this application, it should be understood that the disclosed apparatus may be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative, such as the division of the units, merely a logical function division, and there may be additional manners of dividing the actual implementation, such as multiple units or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, or may be in electrical or other forms.
The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
In addition, each functional unit in each embodiment of the present invention may be integrated in one processing unit, each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units described above may be implemented either in hardware or in software program modules.
The integrated units, if implemented in the form of software program modules, may be stored in a computer-readable memory for sale or use as a stand-alone product. Based on such understanding, the technical solution of the present application may be embodied in essence or a part contributing to the prior art or all or part of the technical solution in the form of a software product stored in a memory, including several instructions for causing a computer device (which may be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in the embodiments of the present application. And the aforementioned memory includes: a U-disk, a read-only memory (ROM), a random access memory (random access memory, RAM), a removable hard disk, a magnetic disk, or an optical disk, or other various media capable of storing program codes.
Those of ordinary skill in the art will appreciate that all or a portion of the steps in the various methods of the above embodiments may be implemented by a program that instructs associated hardware, and the program may be stored in a computer readable memory, which may include: flash disk, read-only memory, random access memory, magnetic or optical disk, etc.
The foregoing has outlined rather broadly the more detailed description of embodiments of the present application, wherein specific examples are provided herein to illustrate the principles and embodiments of the present application, the above examples being provided solely to assist in the understanding of the methods of the present application and the core ideas thereof; meanwhile, as those skilled in the art will have modifications in the specific embodiments and application scope in accordance with the ideas of the present application, the present description should not be construed as limiting the present application in view of the above.
Claims (13)
1. A method of model compression, the method comprising:
determining a first parameter set according to characteristic parameters of data to be trained;
adjusting an initial training model according to the first parameter set to obtain a reference training model, wherein the operand of the reference training model is smaller than that of the initial training model;
Acquiring a second parameter set; the second set of parameters is a parameter characterizing the compression amplitude of the reference training model;
compressing the reference training model according to the second parameter set to obtain a target training model, wherein the precision of the target training model and the precision of the reference training model belong to a preset precision range;
the first parameter set comprises a first sub-parameter and a second sub-parameter, the characteristic parameters comprise aspect ratio distribution and area distribution, and the aspect ratio distribution and the area distribution are wide-height distribution and area distribution of a label frame of data to be trained; the determining the first parameter set according to the characteristic parameters of the data to be trained includes:
determining the first sub-parameter according to the area distribution;
determining the second sub-parameter according to the aspect ratio distribution;
the obtaining the second parameter set includes:
and determining the second parameter set according to the preset model precision and model calculation amount.
2. The method of claim 1, wherein said determining said first sub-parameter from said area distribution comprises:
dividing the area distribution into M sections, wherein M is a positive integer;
Acquiring the average value of each of the M sections;
and determining the first sub-parameter according to the average value.
3. The method according to claim 1 or 2, wherein said determining said second sub-parameter from said aspect ratio distribution comprises:
dividing the aspect ratio distribution into N sections, wherein N is a positive integer;
acquiring the average value of each of the N sections;
and determining the mean value as the second sub-parameter.
4. A method according to any one of claims 1 to 3, wherein the second set of parameters comprises a convolution kernel size, a number of channels and a precision parameter.
5. The method according to any one of claims 1 to 4, further comprising:
training the target training model through a first data set to be trained to obtain a first training model;
and training the first training model through a second data set to be trained to obtain a second training model.
6. The method of claim 5, wherein the training the target training model with the first set of data to be trained further comprises, prior to obtaining a first training model:
Obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the first data set to obtain a first data set to be trained.
7. The method according to claim 5 or 6, wherein before training the first training model by the second data set to be trained to obtain a second training model, the method comprises:
obtaining a mask image of an object to be detected in the data to be trained;
and fusing the mask image of the object to be detected with the image in the second data set to obtain a second data set to be trained.
8. The method of claim 5, wherein training the first training model with the second set of data to be trained to obtain a second training model comprises:
acquiring the weight of each type of data in the second data set to be trained;
training the first training model according to each type of data in the second data set to be trained and the weight of each type of data to obtain a second training model;
and/or the number of the groups of groups,
the training of the target training model through the first data set to be trained to obtain a first training model comprises the following steps:
Acquiring a weight of each type of data in the first data set to be trained;
and training the target training model according to each type of data and the weight of each type of data in the first data set to be trained to obtain a first training model.
9. An image processing method, the method comprising:
receiving an image to be processed;
inputting the image to be processed into an image processing model for processing to obtain target data of a target image;
wherein the image processing model is a target training model obtained by the model compression method according to any one of claims 1 to 4, or the image processing model is a second training model obtained by the model compression method according to any one of claims 5 to 8.
10. A model compression apparatus, the apparatus comprising:
the determining unit is used for determining a first parameter set according to the characteristic parameters of the data to be trained;
the adjusting unit is used for adjusting the initial training model according to the first parameter set to obtain a reference training model, and the operation amount of the reference training model is smaller than that of the initial training model;
an acquisition unit configured to acquire a second parameter set; the second set of parameters is a parameter characterizing the compression amplitude of the reference training model;
The compression unit is used for compressing the reference training model according to the second parameter set to obtain a target training model, and the precision of the target training model and the precision of the reference training model belong to a preset precision range;
the first parameter set comprises a first sub-parameter and a second sub-parameter, the characteristic parameters comprise aspect ratio distribution and area distribution, and the aspect ratio distribution and the area distribution are wide-height distribution and area distribution of a label frame of data to be trained; the determining unit is used for:
determining the first sub-parameter according to the area distribution;
determining the second sub-parameter according to the aspect ratio distribution;
the acquisition unit is used for:
and determining the second parameter set according to the preset model precision and model calculation amount.
11. An image processing apparatus, characterized in that the apparatus comprises:
a receiving unit for receiving an image to be processed;
the processing unit is used for inputting the image to be processed into an image processing model for processing to obtain target data of a target image;
wherein the image processing model is a target training model obtained by the model compression method according to any one of claims 1 to 4, or the image processing model is a second training model obtained by the model compression method according to any one of claims 5 to 8.
12. A terminal comprising a processor and a memory, the processor and the memory being interconnected, wherein the memory is adapted to store a computer program, the computer program comprising program instructions, the processor being configured to invoke the program instructions to perform the method of any of claims 1 to 9.
13. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program comprising program instructions which, when executed by a processor, cause the processor to perform the method of any of claims 1-9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911421331.1A CN111178447B (en) | 2019-12-31 | 2019-12-31 | Model compression method, image processing method and related device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911421331.1A CN111178447B (en) | 2019-12-31 | 2019-12-31 | Model compression method, image processing method and related device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111178447A CN111178447A (en) | 2020-05-19 |
| CN111178447B true CN111178447B (en) | 2024-03-08 |
Family
ID=70657656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911421331.1A Active CN111178447B (en) | 2019-12-31 | 2019-12-31 | Model compression method, image processing method and related device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111178447B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112288686B (en) * | 2020-07-29 | 2023-12-19 | 深圳市智影医疗科技有限公司 | Model training method and device, electronic equipment and storage medium |
| CN113780515A (en) * | 2021-01-05 | 2021-12-10 | 京东鲲鹏(江苏)科技有限公司 | Data processing method, image processing method, device, storage medium and equipment |
| CN114004772A (en) * | 2021-09-30 | 2022-02-01 | 阿里巴巴（中国）有限公司 | Image processing method, image synthesis model determining method, system and equipment |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015167525A1 (en) * | 2014-04-30 | 2015-11-05 | Hewlett-Packard Development Company, L.P. | Automatic page layout for text items and image items |
| CN107909147A (en) * | 2017-11-16 | 2018-04-13 | 深圳市华尊科技股份有限公司 | A kind of data processing method and device |
| CN108416440A (en) * | 2018-03-20 | 2018-08-17 | 上海未来伙伴机器人有限公司 | A kind of training method of neural network, object identification method and device |
| CN110222780A (en) * | 2019-06-12 | 2019-09-10 | 北京百度网讯科技有限公司 | Object detecting method, device, equipment and storage medium |
| CN110334571A (en) * | 2019-04-03 | 2019-10-15 | 复旦大学 | A kind of millimeter-wave image human body method for secret protection based on convolutional neural networks |
| WO2019233297A1 (en) * | 2018-06-08 | 2019-12-12 | Oppo广东移动通信有限公司 | Data set construction method, mobile terminal and readable storage medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10540574B2 (en) * | 2017-12-07 | 2020-01-21 | Shanghai Cambricon Information Technology Co., Ltd | Image compression method and related device |
-
2019
- 2019-12-31 CN CN201911421331.1A patent/CN111178447B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015167525A1 (en) * | 2014-04-30 | 2015-11-05 | Hewlett-Packard Development Company, L.P. | Automatic page layout for text items and image items |
| CN107909147A (en) * | 2017-11-16 | 2018-04-13 | 深圳市华尊科技股份有限公司 | A kind of data processing method and device |
| CN108416440A (en) * | 2018-03-20 | 2018-08-17 | 上海未来伙伴机器人有限公司 | A kind of training method of neural network, object identification method and device |
| WO2019233297A1 (en) * | 2018-06-08 | 2019-12-12 | Oppo广东移动通信有限公司 | Data set construction method, mobile terminal and readable storage medium |
| CN110334571A (en) * | 2019-04-03 | 2019-10-15 | 复旦大学 | A kind of millimeter-wave image human body method for secret protection based on convolutional neural networks |
| CN110222780A (en) * | 2019-06-12 | 2019-09-10 | 北京百度网讯科技有限公司 | Object detecting method, device, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111178447A (en) | 2020-05-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111178447B (en) | Model compression method, image processing method and related device | |
| CN113056769B (en) | Semantic segmentation with soft cross entropy loss | |
| CN108090508B (en) | classification training method, device and storage medium | |
| EP2905722B1 (en) | Method and apparatus for detecting salient region of image | |
| US10262190B2 (en) | Method, system, and computer program product for recognizing face | |
| CN110059728B (en) | RGB-D image visual saliency detection method based on attention model | |
| CN113128481A (en) | Face living body detection method, device, equipment and storage medium | |
| CN111597932A (en) | Road crack image identification method, device and system based on convolutional neural network | |
| CN114419570A (en) | Point cloud data identification method and device, electronic equipment and storage medium | |
| CN111652181B (en) | Target tracking method and device and electronic equipment | |
| CN111067522A (en) | Brain addiction structural map assessment method and device | |
| CN113724128A (en) | Method for expanding training sample | |
| CN107066980A (en) | A kind of anamorphose detection method and device | |
| CN114170438A (en) | Neural network training method, electronic device and computer storage medium | |
| CN112347936A (en) | Rapid target detection method based on depth separable convolution | |
| CN115063672A (en) | Small target detection method based on improved YOLOX model | |
| CN113420871B (en) | Image quality evaluation method, image quality evaluation device, storage medium, and electronic device | |
| WO2022222036A1 (en) | Method and apparatus for determining parking space | |
| CN111178181B (en) | Traffic scene segmentation method and related device | |
| CN111339973A (en) | Object identification method, device, equipment and storage medium | |
| CN112651351B (en) | Data processing method and device | |
| CN115019057A (en) | Image feature extraction model determining method and device and image identification method and device | |
| CN113298102B (en) | Training method and device for target classification model | |
| CN113902933A (en) | Ground segmentation network model training method, device, equipment and medium | |
| CN113822373A (en) | Image classification model training method based on integration and knowledge distillation |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |