CN110349146A - The building method of fabric defect identifying system based on lightweight convolutional neural networks - Google Patents
The building method of fabric defect identifying system based on lightweight convolutional neural networks Download PDFInfo
- Publication number
- CN110349146A CN110349146A CN201910623819.6A CN201910623819A CN110349146A CN 110349146 A CN110349146 A CN 110349146A CN 201910623819 A CN201910623819 A CN 201910623819A CN 110349146 A CN110349146 A CN 110349146A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural networks
- lzfnet
- convolutional
- convolutional neural
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Quality & Reliability (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention proposes a kind of building methods of fabric defect identifying system based on lightweight convolutional neural networks, it the steps include: first, configure the running environment of fabric defect identifying system, according to can factorization convolution obtain lightweight convolutional neural networks, then, acquire textile image sample data, and textile image sample data is standardized, textile image sample data after standardization is divided into training image collection and test chart image set, it recycles the Training strategy of asynchronous gradient decline to input training image collection in lightweight convolutional neural networks to be trained, obtain LZFNet-Fast model, finally, test chart image set input LZFNet-Fast model is tested, verify the performance of LZFNet-Fast model.The present invention using can factorization convolutional coding structure alternate standard convolutional layer, effectively identify texture complexity colored fabric, reduce the number of parameters and calculation amount of model, greatly improve recognition efficiency.
Description
Technical field
The present invention relates to textile images identification technology field, particularly relate to a kind of based on lightweight convolutional neural networks
The building method of fabric defect identifying system.
Background technique
The production cost of weaving finished product is often influenced by original grey at present, and most of quality is asked in apparel industry
Topic is all related with fabric defect, this is one of the main problem that textile industry faces.Fabric defect, frequently referred to fabric defects, refer to
Defect in cloth weaving process, in the product appearance as caused by various unfavorable factors.From fibrous raw material to finished textile product, generally
It need to be possible to generate defect in each processing link by multiple working procedures such as spinning, weaving, printing and dyeing.And artificial detection fabric
Defect cost is excessively high, and is easy to produce missing inspection.Therefore textile production enterprise needs to guarantee using automatic defect identifying system
The quality of fabric.
Existing fabric defect detection method mainly has property of the histogram to analyze both at home and abroad at present, local contrast enhances,
Fourier transformation, wavelet transformation, dictionary learning and histograms of oriented gradients etc..However, not with entire society's textile industry
The raising of disconnected development and consumer's appreciation level, existing fabric have complicated a texture and pattern mostly, and these tradition views
Feeling recognizer, often adaptivity is poor, it is difficult to extract the visual signature for being conducive to classifier identification fault.So traditional
Fabric defect recognition methods is only applicable to monochromatic fabric or the relatively regular fabric of texture, and for texture and more complex fabric of printing and dyeing
Defect recognition, at present be most suitable for completed using depth learning technology.
Convolutional neural networks are rapidly developed, and as a kind of important identification model in certain necks in recent years
Domain achieves significant achievement.It there is also some problems at present however, convolutional neural networks are applied to fabric defect identification.Convolution
Neural model had not only had design feature computation-intensive but also with memory-intensive, this makes convolutional Neural model be difficult portion
Administration is on the limited field programmable gate array of hardware resource and embedded system.Current depth convolutional neural networks are in order to mention
The high-level semantics feature of target image is taken, often there is extremely complex structure, which greatly increases computing costs, are unfavorable for
Fabric defect identifies in real time.Secondly, the trend of convolutional neural networks development in recent years is to build deeper more complicated network structure,
To reach higher accuracy of identification.However, these new technologies for improving model recognition accuracy might not make system transport
It is more efficient in terms of scanning frequency degree and EMS memory occupation.
Summary of the invention
There is the technical problem that structure is complicated, computationally intensive for conventional depth convolutional neural networks, the invention proposes
A kind of fabric defect recognition methods based on lightweight convolutional neural networks constructs the convolution of a fabric identification domain-specific
Module, the module merged it is advanced can factorization convolutional coding structure, first by textile image pass through Three dimensional convolution operation extension
Input as convolution module after to 32 dimensions, then using can factorization convolutional layer progress space filtering, subsequent characteristic pattern quilt
The overall situation at the top of network be averaged pondization lamination be reduced to lower dimensional space.The present invention can use artificial neuron in TensorFlow frame
Network library Keras and tensor flow pattern machine learning library TFslim are fast implemented, suitable for the fabric under computing resource confined condition
Defect recognition.
The technical scheme of the present invention is realized as follows:
A kind of building method of the fabric defect identifying system based on lightweight convolutional neural networks, its step are as follows:
S1, the running environment for configuring fabric defect identifying system；
S2, design can factorization convolutional coding structure, and using can factorization convolutional coding structure construct lightweight convolutional Neural
Network LZFNet-Fast；
S3, acquisition textile image sample data, and textile image sample data is standardized, the fabric after standardization
Image sample data is divided into training image collection and test chart image set；
S4, training image collection is inputted into lightweight convolutional neural networks using the Training strategy of asynchronous gradient decline
It is trained in LZFNet-Fast, obtains LZFNet-Fast model；
S5, LZFNet-Fast model obtained in test chart image set input step S4 is tested, verifies LZFNet-
The performance of Fast model.
The running environment of fabric defect identifying system in the step S1 includes hardware system and software systems, hardware system
The processor of system includes two CPU and two GPU, and the model of CPU is the model of Intel Xeon (R) e5-2650-v4, GPU
It is tall and handsome up to Quadro M5000；Software systems include operating system and convolution library, wherein operating system Windows10,
Convolution library is that convolutional neural networks accelerate library CUDNN7.0.
The construction method of lightweight convolutional neural networks LZFNet-Fast in the step S2 are as follows: define a part
The three-dimensional information fused layer of two-dimensional surface convolutional layer and 1 × 1 × n that receptive field is 3 × 3, then by two-dimensional surface convolutional layer and
Three-dimensional information fused layer be compiled as one can factorization convolutional coding structure, and using can factorization convolutional coding structure build lightweight
Convolutional neural networks LZFNet-Fast.
The lightweight convolutional neural networks LZFNet-Fast include a Standard convolution layer, nine can factorization volume
Product structure, 19 batch regularization layers, an overall situation are averaged pond layer, a full articulamentum and a Softmax classifier；
One can factorization convolutional coding structure include a two-dimensional surface convolutional layer and a three-dimensional information fused layer, a Standard convolution
Layer, nine two-dimensional surface convolutional layers and nine three-dimensional information fused layers totally ten nine convolutional layers, convolutional layer and batch regularization layer one
One is corresponding, the last one batch regularization layer is connected with the average pond layer of the overall situation, the average pond layer of the overall situation and full articulamentum phase
Connection, full articulamentum are connected with Softmax classifier.
The input feature vector figure for being F × F for size, the calculation amount C of Standard convolution layersAre as follows: Cs=F × F × K × K × m
× n, wherein m is the quantity of input channel, and n is the quantity of output channel, and K × K is the convolution kernel size of Standard convolution layer；
Can factorization convolutional coding structure be that convolutional layer is divided into roll lamination and information fused layer two parts, the convolution of roll lamination
Core size is K × K, and the convolution kernel size of information fused layer is 1 × 1, then can factorization convolutional coding structure calculation amount Cf
Are as follows: Cf=F × F × K × K × m+F × F × m × n；Can factorization convolutional coding structure and Standard convolution layer the ratio between calculation amount are as follows:
The Standard convolution layer and four two-dimensional surface convolutional layers are all made of the convolution that stride is 2 and substitute maximum pond, own
The standardization of convolutional layer be all made of batch standardize, activation primitive use codomain for 0~6 amendment linear unit, terminal use
Softmax classifier is as fabric defect decision device.
Textile image sample data is divided into two classes, respectively normal fabric image and defect textile image, normal fabric figure
The quantity of picture and the quantity of defect textile image are close；The quantity of the training image collection accounts for the 4/5 of total quantity, test chart image set
Quantity account for the 1/5 of total quantity.
Training image collection is inputted into lightweight convolutional Neural using the Training strategy of asynchronous gradient decline in the step S4
It is trained in network LZFNet-Fast, the method for obtaining LZFNet-Fast model are as follows:
S41, activation convolutional neural networks accelerate library CUDNN7.0, activate artificial neural network library Keras and tensor flow pattern
Machine learning library TFslim；
S42, normal fabric image and defect textile image that training image is concentrated are separately converted to tfrecord format
File, and two individual files are saved as respectively；
S43, the initial learning rate Θ of initialization, iteration momentum parameter vi, primary iteration number i=0, setting greatest iteration time
Number imax；
S44, training image collection is inputted into lightweight convolutional neural networks using the Training strategy of asynchronous gradient decline
It is trained in LZFNet-Fast:
Wherein, α is weight decaying, and L is loss function, DiThe quantity of training image when for i-th iteration, wiFor wait instruct
Experienced LZFNet-Fast model parameter；
Update LZFNet-Fast model parameter: wi+1=wi+vi+1；
S45, the number of iterations increase by 1, and circulation executes step S44 until reaching maximum number of iterations imax, end loop, life
At LZFNet-Fast model.
It is that the technical program can generate the utility model has the advantages that the present invention using can factorization convolutional coding structure establish LZFNet-
Fast model can match with the design requirement of field programmable gate array or embedded system, can identify texture complexity
Colored fabric, and under the premise of guaranteeing accuracy of identification, than the model parameter number that original neural network VGG16 reduces 98.4%
Amount and 97.6% calculation amount, considerably reduce the dependence to hardware computing capability and memory size, so that depth nerve net
Network is easier to run in industry spot.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is system block diagram of the invention.
Fig. 2 can factorization convolution algorithm figure for of the invention；(a) it is standard three-dimensional convolutional layer, (b) is two-dimensional surface convolution
Layer (c) is information fused layer.
Fig. 3 is the fabric defect figure used in specific example of the present invention；(a) it is lacked for latitude and longitude, (b) is scratch defects,
(c) it is twill defect, is (d) printing and dyeing fault.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained every other under that premise of not paying creative labor
Embodiment shall fall within the protection scope of the present invention.
As shown in Figure 1, a kind of building method of the fabric defect identifying system based on lightweight convolutional neural networks, step
It is rapid as follows:
S1, the running environment for configuring fabric defect identifying system；The running environment of fabric defect identifying system includes hardware
The processor of system and software systems, hardware system includes two CPU and two GPU, and the model of CPU is Intel Xeon
(R) model of e5-2650-v4, GPU are tall and handsome up to Quadro M5000, and software systems include operating system and convolution library,
In, operating system Windows10, convolution library is that convolutional neural networks accelerate library CUDNN7.0.
S2, design can factorization convolutional coding structure, and using can factorization convolutional coding structure construct lightweight convolutional Neural
Network LZFNet-Fast；Define the three-dimensional feature of two-dimensional surface convolutional layer and 1 × 1 × n that a local receptor field is 3 × 3
Figure information fused layer, then by two-dimensional surface convolutional layer and three-dimensional information fused layer be compiled as one can factorization convolutional coding structure,
And using can factorization convolutional coding structure build lightweight convolutional neural networks LZFNet-Fast.The lightweight convolutional Neural
Network LZFNet-Fast include a Standard convolution layer, nine can factorization convolutional coding structure, 19 batch regularization layers,
One overall situation is averaged pond layer, a full articulamentum and a Softmax classifier；One can factorization convolutional coding structure include
One two-dimensional surface convolutional layer and a three-dimensional information fused layer, a Standard convolution layer, nine two-dimensional surface convolutional layers and nine three
Information fused layer totally ten nine convolutional layers are tieed up, (textile image first passes around one for convolutional layer and batch regularization layer one-to-one correspondence
Standard convolution layer is then passed through a batch regularization layer, then passes through a two-dimensional surface convolutional layer, later using one batch
Regularization layer is measured, a three-dimensional information fused layer is then passed through, using a batch regularization layer, two-dimensional surface convolutional layer is criticized
Amount regularization layer and three-dimensional information fused layer are alternately distributed), the last one batch regularization layer is connected with the average pond layer of the overall situation
It connects, the average pond layer of the overall situation is connected with full articulamentum, and full articulamentum is connected with Softmax classifier.
The lightweight convolutional neural networks LZFNet-Fast as shown in Fig. 2 (a), is marked based on Standard convolution structure
It is that K × K × m × n convolution kernel carries out space volume to input feature vector matrix that the convolutional layer of quasi- convolutional coding structure, which is directly using scale,
Product filtering, the calculation amount C of Standard convolution layersAre as follows: Cs=F × F × K × K × m × n, wherein m is the quantity of input channel, and n is
The quantity of output channel, K × K are the convolution kernel size of Standard convolution layer, and F × F is the dimension of input feature vector figure；Such as Fig. 2 (b) and
Shown in Fig. 2 (c), lightweight convolutional neural networks LZFNet-Fast is that convolutional layer is divided into face by factorization convolutional coding structure
Convolutional layer and information fused layer two parts, roll lamination distinguish each input of convolution using single mutually independent convolution filter
Characteristic pattern matrix, then information fused layer carries out the multidimensional characteristic figure that convolutional layer exports by using simple 1 × 1 convolution kernel
Linear combination, can factorization convolutional coding structure calculation amount CfAre as follows: Cf=F × F × K × K × m+F × F × m × n；It is possible thereby to
Find out can the calculation amount of factorization convolutional coding structure be the sum of calculation amount and calculation amount of information fused layer of roll lamination, information
1 × 1 convolution of three-dimensional of fused layer accounts for biggish computation complexity.Assuming that the RAM capacity calculated in equipment is sufficiently large, it is sufficient to deposit
Function mapping and parameter are stored up, then internal storage access cost or internal memory operation number MAC are as follows: MAC=(m+n) × F2+m×n.In described
It deposits access cost or internal memory operation corresponds respectively to the internal storage access and kernel weight of input/output characteristic pattern, internal storage access behaviour
It counts in the presence of the lower limit provided by trigger, when the number for outputting and inputting channel is equal, under internal storage access operand reaches
Limit.But in fact, the RAM on many embedded devices is not big enough.Furthermore modern neuro network operation library generallys use complexity
Block policy is to make full use of caching mechanism, therefore really internal storage access operand may deviation theory value.
In the architecture of GoogLeNet, " multipath " structure is widely adopted in each convolution module.Many fragments
Change operator to be used, rather than several large operations accord with.Although this fragmentation operation structure helps to improve precision, due to
Have the equipment of powerful computation capability unfriendly GPU etc., it may reduce computational efficiency.And it also introduces additional
Expense, as RAM starting with it is synchronous.In lightweight convolutional neural networks, Element-Level operation is occupied for quite a long time, especially
It is on GPU, therefore, the operator being calculated as element includes ReLU, tensor addition, biasing addition etc., in particular, roll product is also
One Element-Level operator, because it also has the ratio between a very high internal storage access operand and floating-point operation rate per second.Passing through will
Standard convolution layer is decomposed into roll lamination and information fused layer, so as to constitute lightweight convolution module easily.It can factor
Decompose the ratio between the calculation amount of convolutional coding structure and Standard convolution layer are as follows:
The lightweight convolutional neural networks LZFNet-Fast is using LZFNet as baseline network, Standard convolution layer
Be substituted for can factorization convolutional coding structure, realize fast convolution operation, with reach compression neural network volume and reduce fabric lack
The purpose that identifying system calculates consumption is fallen into, new network configuration is as shown in table 1.
The network configuration table of 1 LZFNet-Fast of table
| Input dimension | Functional layer | Stride | Convolution kernel dimension |
| 224×224×3 | Three dimensional convolution | 2 | 3×3×3×32 |
| 112×112×32 | Roll lamination | 1 | 3×3×32 |
| 112×112×32 | Fused layer | 1 | 1×1×32×64 |
| 112×112×64 | Roll lamination | 2 | 3×3×64 |
| 56×56×64 | Fused layer | 1 | 1×1×64×128 |
| 56×56×128 | Roll lamination | 1 | 3×3×128 |
| 56×56×128 | Fused layer | 1 | 1×1×128×128 |
| 56×56×128 | Roll lamination | 2 | 3×3×128 |
| 28×28×128 | Fused layer | 1 | 1×1×128×256 |
| 28×28×256 | Roll lamination | 1 | 3×3×256 |
| 28×28×256 | Fused layer | 1 | 1×1×256×256 |
| 28×28×256 | Roll lamination | 2 | 3×3×256 |
| 14×14×256 | Fused layer | 1 | 1×1×256×512 |
| 14×14×512 | Roll lamination | 1 | 3×3×512 |
| 14×14×512 | Fused layer | 1 | 1×1×512×512 |
| 14×14×512 | Roll lamination | 2 | 3×3×512 |
| 7×7×512 | Fused layer | 1 | 1×1×512×1024 |
| 7×7×1024 | Roll lamination | 1 | 3×3×1024 |
| 7×7×1024 | Fused layer | 1 | 1×1×1024×1024 |
| 7×7×1024 | Average pond layer | - | 7×7 |
| 1×1×1024 | Full articulamentum | - | 1024×2 |
| 1×1×2 | Classifier | - | - |
The Standard convolution layer and four two-dimensional surface convolutional layers are all made of the convolution that stride is 2 and substitute under maximum pondization realization
The function of sampling, the last one in lightweight convolutional neural networks LZFNet-Fast can factorization convolutional coding structure output it is special
Levying figure dimension is 7 × 7 × 1024, and all characteristic patterns are then input to average pond layer and realize dimensionality reduction.And all convolutional layers
Standardization is all made of batch standardization (Batch Normalization), and activation primitive uses codomain linear for 0~6 amendment
Unit (ReLU6), terminal is using Softmax classifier as fabric defect decision device.
The specific construction method of lightweight convolutional neural networks LZFNet-Fast is as follows:
1) activate TensorFlow environment, and load " tf.contrib.slim ", " tf.ConfigProto ",
The advanced encapsulation library " functools ", " namedtuple ".
2) definition can factorization convolution operation, two-dimensional surface convolutional layer and information fused layer are merged into a function, and
Assignment is carried out to convolution kernel dimension kernel, stride stride, number of plies depth (line number of table 1) according to table 1.
3) lightweight convolutional neural networks LZFNet-Fast is built in the network structure configuration provided according to table 1, and by super ginseng
Number is set as following values:
The image resolution ratio inputs=224 of textile image,
Categorical measure num_classes=2,
Network depth multiplier depth_multiplier=1.0,
Weight decaying weight_decay=0.00004.
It 4) is " LZFNet-Fast.py " by lightweight convolutional neural networks LZFNet-Fast file designation.
Roll lamination includes 26208 parameters altogether in lightweight convolutional neural networks LZFNet-Fast, and information fused layer is total
Comprising 2091008 parameters, entire model only contains 2120128 weight parameters.Wherein the calculation amount of face convolution operation is
13773312 times, line convolution operation calculation amount is 333971456 times, and the total calculation amount of model is 358584832 times, only accounts for standard
The 10.1% of convolutional network calculation amount.
S3, acquisition textile image sample data, and textile image sample data is standardized, the fabric after standardization
Image sample data is divided into training image collection and test chart image set；The textile image sample data that the present invention uses is that have complexity
The color textile product of grain background, as shown in Figure 3.The standardized method of textile image are as follows: utilize the machine learning of tensor flow pattern
The textile image of .jpg format is converted to .tfrecord format by library TFslim, is conducive to handle large batch of textile image, be mentioned
High training speed.The textile image sample data is divided into two classes, respectively normal fabric image and defect textile image, normally
The quantity of textile image and the quantity of defect textile image are close, and the total quantity of textile image sample data is 3800 width；It is described
The quantity of training image collection accounts for the 4/5 of total quantity, is 3000 width, and the quantity of test chart image set accounts for the 1/5 of total quantity, is 800 width.
S4, training image collection is inputted into lightweight convolutional neural networks using the Training strategy of asynchronous gradient decline
It is trained in LZFNet-Fast, obtains LZFNet-Fast model, steps are as follows:
S41, activation convolutional neural networks accelerate library CUDNN7.0, activate artificial neural network library Keras and tensor flow pattern
Machine learning library TFslim；
S42, normal fabric image and defect textile image that training image is concentrated are separately converted to tfrecord format
File, image size are fixed as 224 × 224 RGB Three Channel Color textile image, and save as two respectively individually
File；
S43, steps for importing 4) " LZFNet-Fast.py " file, initialize initial learning rate Θ=0.01, it is initial repeatedly
For momentum parameter vi=0.9, primary iteration number i=0, setting maximum number of iterations are imax=2000；
S44, training image collection is inputted into lightweight convolutional neural networks using the Training strategy of asynchronous gradient decline
It is trained in LZFNet-Fast:
Wherein, α is weight decaying, and weight attenuation alpha=0.00004, L is loss function, DiTraining when for i-th iteration
The quantity of image, the quantity D of training imagei=64, wiFor LZFNet-Fast model parameter to be trained；I-th iteration is from
64 width images are taken to be used to train in one file, i+1 time iteration takes 64 width images to be used to train out of second file,
Image in file is that have that puts back to take；
Update LZFNet-Fast model parameter: wi+1=wi+vi+1；
S45, circulation execute step S44 until reaching maximum number of iterations imax, end loop, generation LZFNet-Fast mould
Type.
S5, LZFNet-Fast model obtained in test chart image set input step S4 is tested, verifies LZFNet-
The performance of Fast model.
For the performance of assessment models, the image comprising fault and normal picture are separately input to by training completely
In LZFNet-Fast model, and generate respective accuracy rate.In the data set comprising 400 defective textile images, common recognition
Not Chu 378 defect images and 22 normal pictures, omission factor 5.5%.In the data set comprising 400 width normal fabric images
In, 383 width normal pictures and 17 width defect images, false detection rate 4.2% are identified altogether.Generally, total correct of whole system
Discrimination is 95.1%, and the average recognition time of average every width input picture is 13.2 milliseconds.With extensive convolutional neural networks
Detailed performance it is more as shown in table 2.
The amount of images that the coloured fabrics image library that the present invention uses includes is limited, if it is possible to provide more colours and knit
Object image carries out model training, will obtain more preferably experimental result.
The performance comparison of table 2 LZFNet-Fast and extensive convolutional network
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (8)
1. a kind of building method of the fabric defect identifying system based on lightweight convolutional neural networks, which is characterized in that it is walked
It is rapid as follows:
S1, the running environment for configuring fabric defect identifying system；
S2, design can factorization convolutional coding structure, and using can factorization convolutional coding structure construct lightweight convolutional neural networks
LZFNet-Fast；
S3, acquisition textile image sample data, and textile image sample data is standardized, the textile image after standardization
Sample data is divided into training image collection and test chart image set；
S4, training image collection is inputted into lightweight convolutional neural networks LZFNet- using the Training strategy of asynchronous gradient decline
It is trained in Fast, obtains LZFNet-Fast model；
S5, LZFNet-Fast model obtained in test chart image set input step S4 is tested, verifies LZFNet-Fast
The performance of model.
2. the building method of the fabric defect identifying system according to claim 1 based on lightweight convolutional neural networks,
It is characterized in that, the running environment of the fabric defect identifying system in the step S1 includes hardware system and software systems, firmly
The processor of part system includes two CPU and two GPU, and the model of CPU is Intel Xeon (R) e5-2650-v4, GPU's
Model is tall and handsome up to Quadro M5000；Software systems include operating system and convolution library, wherein operating system is
Windows10, convolution library are that convolutional neural networks accelerate library CUDNN7.0.
3. the building method of the fabric defect identifying system according to claim 1 based on lightweight convolutional neural networks,
It is characterized in that, the construction method of the lightweight convolutional neural networks LZFNet-Fast in the step S2 are as follows: define an office
The three-dimensional information fused layer of two-dimensional surface convolutional layer and 1 × 1 × n that portion's receptive field is 3 × 3, then by two-dimensional surface convolutional layer
With three-dimensional information fused layer be compiled as one can factorization convolutional coding structure, and using can factorization convolutional coding structure build light weight
Grade convolutional neural networks LZFNet-Fast.
4. the side of building of the fabric defect identifying system according to claim 1 or 3 based on lightweight convolutional neural networks
Method, which is characterized in that the lightweight convolutional neural networks LZFNet-Fast include a Standard convolution layer, nine can factor
Decompose convolutional coding structure, 19 batch regularization layers, an overall situation are averaged pond layer, a full articulamentum and a Softmax
Classifier；One can factorization convolutional coding structure include a two-dimensional surface convolutional layer and a three-dimensional information fused layer, one mark
Quasi- convolutional layer, nine two-dimensional surface convolutional layers and nine three-dimensional information fused layers totally ten nine convolutional layers, convolutional layer and batch canonical
Change layer to correspond, the last one batch regularization layer is connected with the average pond layer of the overall situation, the average pond layer of the overall situation and Quan Lian
It connects layer to be connected, full articulamentum is connected with Softmax classifier.
5. the building method of the fabric defect identifying system according to claim 4 based on lightweight convolutional neural networks,
It is characterized in that, the input feature vector figure for being F × F for size, the calculation amount C of Standard convolution layersAre as follows: Cs=F × F × K × K × m
× n, wherein m is the quantity of input channel, and n is the quantity of output channel, and K × K is the convolution kernel size of Standard convolution layer；It can
Factorization convolutional coding structure is that convolutional layer is divided into roll lamination and information fused layer two parts, the convolution kernel size of roll lamination
For K × K, the convolution kernel size of information fused layer is 1 × 1, then can factorization convolutional coding structure calculation amount CfAre as follows: Cf=F × F
×K×K×m+F×F×m×n；Can factorization convolutional coding structure and Standard convolution layer the ratio between calculation amount are as follows:
6. the building method of the fabric defect identifying system according to claim 1 based on lightweight convolutional neural networks,
It is characterized in that, the Standard convolution layer and four two-dimensional surface convolutional layers, which are all made of the convolution that stride is 2, substitutes maximum pond, institute
There is the standardization of convolutional layer to be all made of batch to standardize, activation primitive uses codomain for 0~6 amendment linear unit, and terminal adopts
Use Softmax classifier as fabric defect decision device.
7. the building method of the fabric defect identifying system according to claim 1 based on lightweight convolutional neural networks,
It is characterized in that, textile image sample data is divided into two classes, respectively normal fabric image and defect textile image, normal fabric
The quantity of image and the quantity of defect textile image are close；The quantity of the training image collection accounts for the 4/5 of total quantity, test image
The quantity of collection accounts for the 1/5 of total quantity.
8. the building method of the fabric defect identifying system according to claim 1 based on lightweight convolutional neural networks,
It is characterized in that, using the Training strategy of asynchronous gradient decline that training image collection input lightweight convolution is refreshing in the step S4
Through being trained in network LZFNet-Fast, the method for obtaining LZFNet-Fast model are as follows:
S41, activation convolutional neural networks accelerate library CUDNN7.0, activate artificial neural network library Keras and tensor flow pattern machine
Learning database TFslim；
S42, normal fabric image and defect textile image that training image is concentrated are separately converted to tfrecord formatted file,
And two individual files are saved as respectively；
S43, the initial learning rate Θ of initialization, iteration momentum parameter vi, primary iteration number i=0, setting maximum number of iterations
imax；
S44, training image collection is inputted into lightweight convolutional neural networks LZFNet- using the Training strategy of asynchronous gradient decline
It is trained in Fast:
Wherein, α is weight decaying, and L is loss function, DiThe quantity of training image when for i-th iteration, wiWait train
LZFNet-Fast model parameter；
Update LZFNet-Fast model parameter: wi+1=wi+vi+1；
S45, the number of iterations increase by 1, and circulation executes step S44 until reaching maximum number of iterations imax, end loop, generation
LZFNet-Fast model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910623819.6A CN110349146B (en) | 2019-07-11 | 2019-07-11 | Method for constructing fabric defect identification system based on lightweight convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910623819.6A CN110349146B (en) | 2019-07-11 | 2019-07-11 | Method for constructing fabric defect identification system based on lightweight convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110349146A true CN110349146A (en) | 2019-10-18 |
| CN110349146B CN110349146B (en) | 2020-06-02 |
Family
ID=68175700
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910623819.6A Active CN110349146B (en) | 2019-07-11 | 2019-07-11 | Method for constructing fabric defect identification system based on lightweight convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110349146B (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110827263A (en) * | 2019-11-06 | 2020-02-21 | 创新奇智(南京)科技有限公司 | Magnetic shoe surface defect detection system and detection method based on visual identification technology |
| CN110929603A (en) * | 2019-11-09 | 2020-03-27 | 北京工业大学 | Weather image identification method based on lightweight convolutional neural network |
| CN110930387A (en) * | 2019-11-21 | 2020-03-27 | 中原工学院 | Fabric defect detection method based on depth separable convolutional neural network |
| CN111402203A (en) * | 2020-02-24 | 2020-07-10 | 杭州电子科技大学 | Fabric surface defect detection method based on convolutional neural network |
| CN111476138A (en) * | 2020-03-31 | 2020-07-31 | 万翼科技有限公司 | Construction method and identification method of building drawing component identification model and related equipment |
| CN111709429A (en) * | 2020-06-01 | 2020-09-25 | 江南大学 | Method for identifying structural parameters of woven fabric based on convolutional neural network |
| CN112115986A (en) * | 2020-08-31 | 2020-12-22 | 南京航空航天大学 | Power transmission line scene classification method based on lightweight neural network |
| CN112418397A (en) * | 2020-11-19 | 2021-02-26 | 重庆邮电大学 | Design method of lightweight convolutional neural network |
| CN112598657A (en) * | 2020-12-28 | 2021-04-02 | 锋睿领创(珠海)科技有限公司 | Defect detection method and device, model construction method and computer equipment |
| CN113158968A (en) * | 2021-05-10 | 2021-07-23 | 苏州大学 | Embedded object cognitive system based on image processing |
| CN113298751A (en) * | 2020-09-29 | 2021-08-24 | 湖南长天自控工程有限公司 | Detection method for auxiliary door blockage |
| CN113936001A (en) * | 2021-12-17 | 2022-01-14 | 杭州游画丝界文化艺术发展有限公司 | Textile surface flaw detection method based on image processing technology |
| CN115713533A (en) * | 2023-01-10 | 2023-02-24 | 佰聆数据股份有限公司 | Method and device for detecting surface defects of electrical equipment based on machine vision |
| GB2623140A (en) * | 2023-03-02 | 2024-04-10 | Imagination Tech Ltd | Methods and systems for performing a sparse submanifold convolution using an NNA |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20010012381A1 (en) * | 1997-09-26 | 2001-08-09 | Hamed Sari-Sarraf | Vision-based, on-loom fabric inspection system |
| CN101866427A (en) * | 2010-07-06 | 2010-10-20 | 西安电子科技大学 | Method for detecting and classifying fabric defects |
| CN104751472A (en) * | 2015-04-10 | 2015-07-01 | 浙江工业大学 | Fabric defect detection method based on B-spline wavelets and deep neural network |
| CN107833220A (en) * | 2017-11-28 | 2018-03-23 | 河海大学常州校区 | Fabric defect detection method based on depth convolutional neural networks and vision significance |
| CN108717568A (en) * | 2018-05-16 | 2018-10-30 | 陕西师范大学 | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network |
| US20180322623A1 (en) * | 2017-05-08 | 2018-11-08 | Aquifi, Inc. | Systems and methods for inspection and defect detection using 3-d scanning |
| CN109325940A (en) * | 2018-09-05 | 2019-02-12 | 深圳灵图慧视科技有限公司 | Textile detecting method and device, computer equipment and computer-readable medium |
| CN109613006A (en) * | 2018-12-22 | 2019-04-12 | 中原工学院 | A kind of fabric defect detection method based on end-to-end neural network |
-
2019
- 2019-07-11 CN CN201910623819.6A patent/CN110349146B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20010012381A1 (en) * | 1997-09-26 | 2001-08-09 | Hamed Sari-Sarraf | Vision-based, on-loom fabric inspection system |
| CN101866427A (en) * | 2010-07-06 | 2010-10-20 | 西安电子科技大学 | Method for detecting and classifying fabric defects |
| CN104751472A (en) * | 2015-04-10 | 2015-07-01 | 浙江工业大学 | Fabric defect detection method based on B-spline wavelets and deep neural network |
| US20180322623A1 (en) * | 2017-05-08 | 2018-11-08 | Aquifi, Inc. | Systems and methods for inspection and defect detection using 3-d scanning |
| CN107833220A (en) * | 2017-11-28 | 2018-03-23 | 河海大学常州校区 | Fabric defect detection method based on depth convolutional neural networks and vision significance |
| CN108717568A (en) * | 2018-05-16 | 2018-10-30 | 陕西师范大学 | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network |
| CN109325940A (en) * | 2018-09-05 | 2019-02-12 | 深圳灵图慧视科技有限公司 | Textile detecting method and device, computer equipment and computer-readable medium |
| CN109613006A (en) * | 2018-12-22 | 2019-04-12 | 中原工学院 | A kind of fabric defect detection method based on end-to-end neural network |
Non-Patent Citations (2)
| Title |
|---|
| LI C, GAO G, LIU Z, ET AL: "Fabric defect detection based on biological vision modeling", 《IEEE ACCESS, 2018》 * |
| LIU, ZHOUFENG, ET AL.: "Fabric Defect Detection Algorithm Based on Convolution Neural Network and Low-Rank Representation", 《2017 4TH IAPR ASIAN CONFERENCE ON PATTERN RECOGNITION (ACPR)》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110827263A (en) * | 2019-11-06 | 2020-02-21 | 创新奇智(南京)科技有限公司 | Magnetic shoe surface defect detection system and detection method based on visual identification technology |
| CN110827263B (en) * | 2019-11-06 | 2021-01-29 | 创新奇智(南京)科技有限公司 | Magnetic shoe surface defect detection system and detection method based on visual identification technology |
| CN110929603A (en) * | 2019-11-09 | 2020-03-27 | 北京工业大学 | Weather image identification method based on lightweight convolutional neural network |
| CN110929603B (en) * | 2019-11-09 | 2023-07-14 | 北京工业大学 | Weather image recognition method based on lightweight convolutional neural network |
| CN110930387A (en) * | 2019-11-21 | 2020-03-27 | 中原工学院 | Fabric defect detection method based on depth separable convolutional neural network |
| CN111402203B (en) * | 2020-02-24 | 2024-03-01 | 杭州电子科技大学 | Fabric surface defect detection method based on convolutional neural network |
| CN111402203A (en) * | 2020-02-24 | 2020-07-10 | 杭州电子科技大学 | Fabric surface defect detection method based on convolutional neural network |
| CN111476138A (en) * | 2020-03-31 | 2020-07-31 | 万翼科技有限公司 | Construction method and identification method of building drawing component identification model and related equipment |
| CN111476138B (en) * | 2020-03-31 | 2023-08-18 | 万翼科技有限公司 | Construction method, identification method and related equipment for building drawing component identification model |
| CN111709429B (en) * | 2020-06-01 | 2023-05-05 | 江南大学 | Woven fabric structural parameter identification method based on convolutional neural network |
| CN111709429A (en) * | 2020-06-01 | 2020-09-25 | 江南大学 | Method for identifying structural parameters of woven fabric based on convolutional neural network |
| CN112115986A (en) * | 2020-08-31 | 2020-12-22 | 南京航空航天大学 | Power transmission line scene classification method based on lightweight neural network |
| CN112115986B (en) * | 2020-08-31 | 2024-04-16 | 南京航空航天大学 | Lightweight neural network-based power transmission line scene classification method |
| CN113298751A (en) * | 2020-09-29 | 2021-08-24 | 湖南长天自控工程有限公司 | Detection method for auxiliary door blockage |
| CN112418397A (en) * | 2020-11-19 | 2021-02-26 | 重庆邮电大学 | Design method of lightweight convolutional neural network |
| CN112418397B (en) * | 2020-11-19 | 2021-10-26 | 重庆邮电大学 | Image classification method based on lightweight convolutional neural network |
| CN112598657B (en) * | 2020-12-28 | 2022-03-04 | 锋睿领创(珠海)科技有限公司 | Defect detection method and device, model construction method and computer equipment |
| CN112598657A (en) * | 2020-12-28 | 2021-04-02 | 锋睿领创(珠海)科技有限公司 | Defect detection method and device, model construction method and computer equipment |
| CN113158968A (en) * | 2021-05-10 | 2021-07-23 | 苏州大学 | Embedded object cognitive system based on image processing |
| WO2022237061A1 (en) * | 2021-05-10 | 2022-11-17 | 苏州大学 | Embedded object cognitive system based on image processing |
| CN113936001A (en) * | 2021-12-17 | 2022-01-14 | 杭州游画丝界文化艺术发展有限公司 | Textile surface flaw detection method based on image processing technology |
| CN115713533A (en) * | 2023-01-10 | 2023-02-24 | 佰聆数据股份有限公司 | Method and device for detecting surface defects of electrical equipment based on machine vision |
| GB2623140A (en) * | 2023-03-02 | 2024-04-10 | Imagination Tech Ltd | Methods and systems for performing a sparse submanifold convolution using an NNA |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110349146B (en) | 2020-06-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110349146A (en) | The building method of fabric defect identifying system based on lightweight convolutional neural networks | |
| Fu et al. | DSAGAN: A generative adversarial network based on dual-stream attention mechanism for anatomical and functional image fusion | |
| Zhang et al. | Fabric defect detection using salience metric for color dissimilarity and positional aggregation | |
| CN109636772A (en) | The defect inspection method on the irregular shape intermetallic composite coating surface based on deep learning | |
| CN109063728A (en) | A kind of fire image deep learning mode identification method | |
| CN109446992A (en) | Remote sensing image building extracting method and system, storage medium, electronic equipment based on deep learning | |
| CN107833220A (en) | Fabric defect detection method based on depth convolutional neural networks and vision significance | |
| CN110264484A (en) | A kind of improvement island water front segmenting system and dividing method towards remotely-sensed data | |
| CN109543688A (en) | A kind of novel meter reading detection and knowledge method for distinguishing based on multilayer convolutional neural networks | |
| CN105678788A (en) | Fabric defect detection method based on HOG and low-rank decomposition | |
| CN109272500A (en) | Fabric classification method based on adaptive convolutional neural networks | |
| CN109934154A (en) | A kind of remote sensing image variation detection method and detection device | |
| CN109376787A (en) | Manifold learning network and computer visual image collection classification method based on it | |
| CN109410195A (en) | A kind of magnetic resonance imaging brain partition method and system | |
| Zhang et al. | Transformer based conditional GAN for multimodal image fusion | |
| CN113989683A (en) | Ship detection method for synthesizing synchronous orbit sequence optical image space-time information | |
| Wang et al. | A fabric defect detection system based improved yolov5 detector | |
| Li et al. | Data augmentation and layered deformable mask R-CNN-based detection of wood defects | |
| An et al. | Fabric defect detection using deep learning: An Improved Faster R-approach | |
| Wang et al. | Automatic assessment of fabric smoothness appearance based on a compact convolutional neural network with label smoothing | |
| CN111709429B (en) | Woven fabric structural parameter identification method based on convolutional neural network | |
| CN109558803A (en) | SAR target discrimination method based on convolutional neural networks Yu NP criterion | |
| CN109344779A (en) | A kind of method for detecting human face under ring road scene based on convolutional neural networks | |
| CN108537266A (en) | A kind of cloth textured fault sorting technique of depth convolutional network | |
| Qiao | Application of Gabor image recognition technology in intelligent clothing design |
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 |