CN110097547A - A kind of automatic testing method that the weld seam egative film based on deep learning is faked - Google Patents

A kind of automatic testing method that the weld seam egative film based on deep learning is faked Download PDF

Info

Publication number
CN110097547A
CN110097547A CN201910370302.0A CN201910370302A CN110097547A CN 110097547 A CN110097547 A CN 110097547A CN 201910370302 A CN201910370302 A CN 201910370302A CN 110097547 A CN110097547 A CN 110097547A
Authority
CN
China
Prior art keywords
weld seam
egative film
image
character
weld
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
Application number
CN201910370302.0A
Other languages
Chinese (zh)
Other versions
CN110097547B (en
Inventor
张帆
张新红
张伯言
杜海顺
史潇婉
李珍珍
赵茹楠
侯婷婷
任方涛
刘茜
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Henan University
Original Assignee
Henan University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Henan University filed Critical Henan University
Priority to CN201910370302.0A priority Critical patent/CN110097547B/en
Publication of CN110097547A publication Critical patent/CN110097547A/en
Application granted granted Critical
Publication of CN110097547B publication Critical patent/CN110097547B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N23/00—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00
    • G01N23/02—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material
    • G01N23/04—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and forming images of the material
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/14—Image acquisition
    • G06V30/148—Segmentation of character regions
    • G06V30/153—Segmentation of character regions using recognition of characters or words
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10116—X-ray image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30152—Solder
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Analytical Chemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Chemical & Material Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Biochemistry (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)
  • Analysing Materials By The Use Of Radiation (AREA)

Abstract

For current the case where lacking weld seam egative film fraud automatic detection means, the present invention devises a kind of automatic testing method that the weld seam x-ray inspection egative film based on deep learning is faked, first by ray colloid film scanning at digitlization weld seam egative film image, it include the weld seam egative film image of all weld seam character marks by screening a part, and all character marks in weld seam egative film image are split to the character recognition model for being used to train recognizable character, go out the new character information submitted in weld seam egative film image by character recognition model extraction, the new weld information submitted in weld seam egative film image is extracted additionally by improved VGGNet-16 convolutional neural networks, the character information that will be extracted, weld information respectively with the character information that is saved in database, weld information is compared, quickly, it accurately detects and former Have that the character mark of weld seam egative film is identical or the identical weld seam egative film of weld seam, that is, the weld seam egative film faked guarantees construction quality.

Description

A kind of automatic testing method that the weld seam egative film based on deep learning is faked
Technical field
The invention belongs to pipelines to weld detection technique field, in particular to a kind of weld seam egative film based on deep learning is faked Automatic testing method.
Background technique
Welding defect is different according to its position in weld seam, is divided into External Defect (undercut, collapses overlap) and inside Defect (crackle, stomata, slag inclusion, lack of penetration etc.) can be achieved to distinguish to the External Defect for being located at face of weld by directly observation Not, and internal flaw need to be carried out through non-destructive testing technology detection realize distinguish.
Common Standards for Pipeline welding Defects detection technique mainly utilizes ultrasonic wave, vortex, infrared and ray detection etc..Ray inspection It surveys and is also known as radiographic inspection, so-called radiographic inspection needs to be the absorption energy using object different structure to ray using film as carrier Power is different, generates photosensitization difference to check that a kind of method of weld seam internal flaw, the basic principle of radiographic inspection are as follows: when When the beam transillumination of even intensity penetrates object, if object regional area existing defects or structure have differences, it will change Decaying of the object to ray, so that different parts transmitted ray intensity is different, in this way, using certain detector (for example, ray Film is used in photograph) detection transmitted ray intensity, so that it may defect and substance distribution inside judgment object etc..Radial energy is not Metal material is penetrated with degree, photosensitization is generated to photographic film, using this performance, when ray passes through examined weldering When seam, because weld defect is different to the absorbability of ray, ray is set to fall in the intensity on film different, light reaching the film degree Also different, thus can it is accurate, reliable, non-destructively the shape of display defect, position and size, technical staff pass through sight Examining the image information shown on film can reach the purpose of identification defect.Common ray has X-ray and two kinds of gamma-rays, wherein X-ray nondestructive inspection is one of the main method of industy pipe welding non-destructive testing, and testing result is as weld defect point The important judgment basis of analysis and quality evaluation.
According to x-ray inspection technological procedure, image quality indicator must be shown on every egative film, telltale mark (centralized positioning label, Overlap joint label) it is marked with identification.Identification label includes pipeline number, weld bond number, welder number, welding date etc..If reprocessing when retaking After piece number plus label is reprocessed, the footnotes such as R1, R2 are number of rewelding.Above-mentioned various label fonts are placed on specified position before shooting It sets, is imaged together with weld seam, therefore altogether include two parts, respectively weld image and for distinguishing in X-ray colloid egative film image The character picture of weld seam.
In engineering, there is the phenomenon that faking in weld seam x-ray inspection egative film, i.e., what the identification on egative film was marked and to be shot Weld seam is misfitted, in fact it could happen that is shot old weld seam with new font label or is shot the fraud side of new weld seam with old font label Formula.For example, in the actual acceptance of work, if engineering side thinks certain weld seam or certain weld bond, wherein one section of weldquality is not It is good, it can not be by checking, they may re-shoot a high-quality weld seam with new label, pretend to be the weld seam that quality is bad (shooting old weld seam with new font label), and submit the weld seam egative film of fraud.Weld seam egative film quantity in a usual engineering is huge Greatly, only it is difficult to find this fraud with eye-observation, if the weld seam of these existing defects passes through examination, it will directly affect welding The quality and service life of product, there are great security risks.
Check that the method that egative film is faked is weld seam backsheet landing seam, such as the weld seam bottom of a weld bond before and after manual review at present It is not overlapped before and after piece, there is the suspicion of fraud.But manual review is time-consuming and laborious, is easy missing inspection, and if entire weldering Whole weld seam egative films of mouth all fake and (change a weld bond shooting), and overlap joint, which checks, can not solve the problems, such as fraud, moreover read tablet people Member should be placed on energy welding defect inspection rather than prevent in fraud.
Summary of the invention
New weldering is shot in order to quickly and accurately check to be shot old weld seam with new font label or marked with old font The fraud egative film of seam, the automatic testing method for the weld seam egative film fraud based on deep learning that the present invention provides a kind of, can be certainly It is dynamic to detect that weld seam present in weld seam egative film identifies the identical egative film of identical or weld image, quality in engineering is found in time Underproof weld seam, ensure that construction quality, eliminate safe hidden trouble, and improve work efficiency.
The technical solution adopted by the present invention are as follows:
A kind of automatic testing method that the weld seam egative film based on deep learning is faked, includes the following steps:
S1: by X-ray colloid film scanning at digitlization weld seam egative film image；
S2: the weld seam egative film image of a batch 200 or more is screened, should include all to be detected in this batch of weld seam egative film image All characters being likely to occur in weld seam egative film；
S3: being split this batch of weld seam egative film image using Radon transform, obtains the small image of multiple characters, completes training The production of character set；
S4: training, the study of deep learning are carried out to training character set using LeNet-5 convolutional neural networks, is welded Stitch character recognition model；
S5: establishing database, for saving the new weld image feature and weld seam character string for submitting weld seam egative film image；
S6: the weld seam egative film image newly submitted is split using Radon transform, the weld seam negative map newly submitted The small image of multiple characters of picture；
S7: the small image of character in S6 is inputted into weld seam character recognition model, identification extracts new submission weld seam negative map The weld seam character string of picture；
S8: the weld seam character string extracted in S7 and the weld seam character string saved in database are compared, comparison processing As a result are as follows: if result mismatches, the new weld seam character string for submitting weld seam egative film image is saved in the database；If result Match, then alarm, detects to mark identical weld seam egative film；
S9: image segmentation, scaling are carried out to the weld seam egative film image newly submitted, obtain the small image of multiple weld seams；
S10: carrying out deep learning feature extraction to the small image of weld seam in S9 using improved VGGNet-16 network, will The characteristic sequence extracted from the small image of each weld seam merges, and obtains the feature of whole weld image, and be stored in Weld image feature in database is compared, comparison result processing are as follows: if result mismatches, will newly submit weld seam negative map Picture and the characteristics of weld seam extracted save in the database, if result matches, alarm, detect the identical weld seam bottom of weld seam Piece.
Specifically, the VGGNet-16 convolutional neural networks are made of five convolutional layer groups and three full articulamentums, described LeNet-5 convolutional neural networks are made of two convolutional layers, two pond layers, a full articulamentum and one softmax layers.
Specifically, specific steps weld seam negative map picture being split using Radon transform are as follows:
Step a. is to weld seam negative map picture degree of comparing stretch processing；
Step b. carries out binary conversion treatment using OTSU algorithm to weld seam negative map picture；
Step c. automatically splits all characters, label in weld seam egative film image using Radon transform, obtains multiple The small image of character.
In order to eliminate the small burr and small holes in the small image of character on character, carry out after the step c to splitting The small image of weld seam character handled using mathematical Morphology Algorithm.
Further, in order to improve the precision of character recognition model, need to expand trained character set, the specific steps are to place The small image rotation processing of weld seam character after reason, the small image of character before obtaining the small image of multiple characters, and rotation form together Training character set after expansion.
Specifically, the treatment process of the S7 are as follows:
The small image of character in S6 is inputted into character recognition model, the corresponding character of the small image of character is identified, identification Character out is combined classification in order, obtains a character string, i.e., the weld seam character string for the weld seam egative film image newly submitted.
Specifically, in order to protrude the feature of weld seam and facilitate segmentation, the S9 includes the following steps:
S91. weld seam egative film image degree of the comparing stretch processing to newly submitting；
S92. to treated, weld seam egative film image uses Radon transform to be divided into the small image of multiple weld seams
S93. image scaling, the size of the unified small image of weld seam are carried out to the small image of weld seam after segmentation；
S94. negative process is carried out to the small image of weld seam after scaling.
Beneficial effects of the present invention:
The present invention weld seam film scanning at digitlization weld seam egative film image, further by the word in digital negative image Symbol image segmentation comes out, and training is used for the deep learning model of weld seam character recognition, can be fast by weld seam character recognition model Speed identifies the new identification information submitted in weld seam egative film image, before saving the identification information extracted and in database It submits the identification information of weld seam egative film image to be compared, may recognize that the identical weld seam egative film image of mark；The present invention utilizes Improved VGGNet-16 convolutional neural networks can rapidly extracting newly submit the characteristics of weld seam of weld seam egative film image, and be stored in Characteristics of weld seam in database is compared, and quickly and accurately detects the identical weld seam egative film image of weld seam, and the present invention is effective It detects the weld seam egative film for the fraud submitted in the weld seam welding quality course of receiving, guarantees construction quality.
Detailed description of the invention
Fig. 1 is weld seam egative film image schematic diagram；
Fig. 2 is the small image schematic diagram of character being partitioned into from weld seam egative film image；
Fig. 3 is weld seam character recognition model training flow chart；
Fig. 4 is the new character recognizing process figure for submitting weld seam egative film image；
Fig. 5 is the new weld seam recognition flow chart for submitting weld seam egative film image；
Fig. 6 is the new small image schematic diagram of weld seam submitted after weld seam egative film image segmentation；
Fig. 7 is LeNet-5 convolutional neural networks structural schematic diagram of the present invention；
Fig. 8 is improved VGGNet-16 convolutional neural networks structural schematic diagram of the present invention；
Fig. 9 is flow chart of the invention.
Specific embodiment
In order to clearly illustrate the method for the invention, the present invention will be done specifically in conjunction with specific weld seam egative film image It is bright,
A kind of automatic testing method that the weld seam egative film based on deep learning is faked, as shown in figure 9, including the following steps:
S1: by X-ray colloid film scanning at digitlization weld seam egative film image, the weld seam egative film image of completion is scanned as schemed Shown in 1, image size is about 4200 × 880, and the size of every weld image is slightly different after scanning, but difference is little, figure As format is tif format；
S2: the weld seam egative film image of screening a batch 200 or more, because the character mark of weld seam is the arrangement of multiple characters Combination can be with so need all characters comprising being likely to occur in weld seam egative film image to be detected in this batch of weld seam egative film image The weld seam egative film image of other weld seams or weld bond, specific weld seam also can be selected in the weld seam egative film image for selecting weld seam to be detected Character mark be following 42 characters permutation and combination, 42 characters are as follows: 0~9 this 10 Arabic numerals, 26 English Text mother, " year, month, day " this 3 Chinese characters and 3 special characters, 3 special characters are 3 of fixation Character is respectively intended to mark center position and lap position；
S3: this batch of weld seam egative film image is split using Radon transform, obtains 4000 or more the small images of character, word The quantity for according with small image is determined according to the character quantity in the weld seam egative film image of screening, completes the production of training character set, tool Body proceeds as follows:
Step a: image enhancement processing is carried out by contrast stretching to weld seam negative map picture, contrast stretching belongs to gray scale Map function can be enhanced the contrast of image each section, that is, enhance the gray areas of weld image character marking part, after being convenient for Phase splits character marking；
Step b: carrying out binary conversion treatment using OTSU algorithm to enhancing treated weld seam egative film image, OTSU algorithm, That is maximum variance between clusters, sometimes referred to as Otsu algorithm are the algorithms of determination binarization threshold relatively good at present, specifically Principle are as follows: a threshold value is selected according to gray level, digital picture is divided into greater than threshold value and is less than two parts of threshold value, that is, Prospect and two parts of background, calculate the inter-class variance of the two parts, inter-class variance is bigger, just illustrates foreground and background two Partial gray-level difference finds out the maximum threshold value of inter-class variance, as optimal threshold away from bigger.In terms of respectively by 0~255 for threshold value Calculate totally 256 inter-class variances, the corresponding threshold value of maximum inter-class variance is exactly optimal threshold, will according to the optimal threshold found Weld seam egative film image binaryzation；
Step c: using Radon transform automatically all Character segmentations in the weld seam egative film image after binary conversion treatment Out, since digital picture can be expressed as the form of matrix, Radon transform is exactly by digital image matrix at a certain specified angle Projective transformation is done in degree directions of rays, which does drawing east for digitlization weld seam egative film image array first in the horizontal direction and become It changes, is partitioned into the region where character marking, then do Radon transform in vertical direction, is partitioned into each character, obtains multiple words Small image is accorded with, as shown in Fig. 2, zooming in and out operation to the small image of the character being partitioned into, obtains the word that multiple sizes are 32 × 32 Small image is accorded with, the small image of character after all segmentations together forms the training character for training weld seam character recognition model Collection.
Further, small to the character split in order to keep the boundary on the small image of the character split smoother Image is handled using mathematical Morphology Algorithm, that is, is done an opening operation, first corroded reflation, to eliminate on the small image of character Small burr and small holes.
Further, in order to improve the precision of character recognition model, training character set can be expanded, specifically Ground is rotated three times the small image of each character respectively clockwise and anticlockwise for step-length with 1 degree, obtains 6 postrotational characters The small image of character before small image, the small image of the character obtained after all rotations and rotation forms the training character after expanding together Collection, the small amount of images of character after expansion in training character set are 7 times originally；
S4: training, the study of deep learning are carried out to training character set using LeNet-5 convolutional neural networks, is welded Character recognition model is stitched, training process is as shown in Figure 3.Fig. 7 be the LeNet-5 convolutional neural networks structure chart, specifically by Two convolutional layers, two pond layers, a full articulamentum and a softmax layers of composition.It is realized first by convolutional layer to word Accord with the extraction of small characteristics of image, then by pond layer to characteristic image carry out down-sampling, specifically be maximum value down-sampling Method, pond layer can reduce the dimension of characteristic image, and reducing number of parameters prevents over-fitting, have characteristics of image relatively strong Robustness, full articulamentum by weighted sum to characteristics of image carry out nonlinear fitting, two dimensional image is become one-dimensional characteristic Vector carries out classification output finally by softmax layers, obtains deep learning character recognition model.Specifically, the first convolutional layer Input be the small image 32 × 32 of original character, convolution kernel size is 5 × 5, is filled without using full 0, step-length 1, this layer Output be 6 28 × 28 characteristic patterns (feature map).The input of first pond layer be first layer output, 6 28 × 28 Characteristic pattern (or being interpreted as the port number of image is 6), the window size of down-sampling is 2 × 2, and using maximum value pond, step-length is 2, so the output characteristic pattern size of this layer is 6 14 × 14.The input of second convolutional layer is 6 14 × 14 characteristic patterns, is used Convolution kernel size be 5 × 5, filled without using full 0, step-length 1, the output characteristic pattern size of this layer is 10 × 10, totally 16. The input of second pond layer is 16 10 × 10 characteristic patterns, and the window size of down-sampling is 2 × 2, using maximum value pond, step-length Be 2, output characteristic pattern size be 5 × 5, totally 16.The input of full articulamentum is 16 5 × 5 characteristic patterns, and exporting is 120 × 1 Feature vector, finally using (softmax layers) progress categorised decisions of output layer.Weld seam character recognition model output is 42 Input picture is divided into 42 classes by deep learning by class, respectively represent 10 Arabic numerals, 26 English alphabets, in 3 Literary Chinese character and 3 spcial characters；
S5: establishing database, for saving the new weld image feature and weld seam character string for submitting weld seam egative film image, Middle weld image feature is specially the feature vector of the weld image extracted based on convolutional neural networks, and weld seam character string is character The identified rear permutation and combination of multiple characters in image obtains, and in the present embodiment, there are three field, two of them altogether in database Field is for saving weld image feature and weld seam character string, and third field is for saving weld seam egative film image；
S6: the weld seam egative film image newly submitted is split using Radon transform, the weld seam negative map newly submitted The small image of multiple characters of picture, specific segmentation step are shown in above-mentioned steps a, step b and step c；
S7: the small image of character obtained in S6 is inputted into weld seam character recognition model, identification extracts new submission weld seam bottom Each weld seam character in picture；Specifically, the small image of character in S6 is inputted into character recognition model, by character recognition mould Type automatically identifies the corresponding character of the small image of character, and the character identified is combined in order, is newly submitted The character string of weld seam egative film image, wherein identify (every weld seam egative film image after the centralized positioning label of weld seam egative film image In only one centralized positioning label, be a fixed spcial character), the location information of records center telltale mark, rear Continue and is used when being split to weld image.
S8: the weld seam character string extracted in S7 and the weld seam character string saved in database are compared, and ask two welderings The Hamming distance between character string is stitched, Hamming distance is the number of the not identical characters of two weld seam character string corresponding positions, such as Fruit Hamming distance is equal to zero and matches, if Hamming distance is greater than zero, mismatches.Compare processing result are as follows: if result is not Match, the new weld seam character string for submitting weld seam egative film image is stored in the respective field of database；If result matches, report It is alert, it detects to mark identical weld seam egative film；When detecting first weld seam egative film image, because there is no weld seam character in database String, therefore the weld seam character string of first weld seam egative film image can automatically save in the database, specific identification and process flow are such as Shown in Fig. 4, which can rapidly and accurately detect to mark the fraud weld seam egative film for shooting new weld seam with old font.
S9: image segmentation, scaling are carried out to the weld seam egative film image newly submitted, detailed process is as shown in figure 5, obtain multiple The small image of weld seam；
Specifically comprise the following steps:
S91. weld seam egative film image degree of the comparing stretch processing to newly submitting；
S92. to treated, weld seam egative film image uses Radon transform to be divided into the small image of multiple weld seams, specific steps Are as follows: it is primarily based on Radon transform and carries out horizontal direction projection, determine the up-and-down boundary of weld seam, the spacing of weld seam up-and-down boundary is The height H of weld seam obtains one and eliminates character denotation information, only comprising weld seam according to the up-and-down boundary cutting image of weld seam Weld image.Weld image height is H, is started according to the center character marking recorded in S7, respectively to the left and right sides with H Divide weld image for step-length, the small image of weld seam after obtaining multiple segmentations；
S93. image scaling is carried out to the small image of weld seam after segmentation, as shown in fig. 6, the size of the unified small image of weld seam is 224×224；
S94. negative process is carried out to the small image of weld seam after scaling.
S10: deep learning feature extraction is carried out to the small image of weld seam in S9 using VGGNet-16 convolutional neural networks.
Because the present invention mainly extracts weld image feature and carries out aspect ratio pair, do not need to classify, so taking Disappeared softmax layers, directly using the full articulamentum of third output as weld image feature vector, described VGGNet-16 volumes Product neural network is made of five convolutional layer groups and three full articulamentums.Specifically as shown in figure 8, five convolutional layer groups are as follows: ConvA, ConvB, ConvC, ConvD and ConvE, each convolutional layer group include multiple convolutional layers and a pond layer, specifically Ground, ConvA and ConvB are made of two convolutional layers and a maximum pond layer, and ConvC, ConvD and ConvE are by three Convolutional layer and a maximum pond layer composition, the convolution kernel size of all convolutional layers is 3 × 3, and convolution step-length is 1, is owned The window size of pond layer is 2 × 2, and step-length is 2, and each layer is all made of ReLU as activation primitive.Convolutional layer and pond The renewal process of layer is as follows:
During the propagated forward of convolutional layer, upper one layer of feature maps is rolled up by the convolution kernel that one can learn Then product just obtains output feature map by activation primitive.
Wherein: MjIndicate the input feature vector set of graphs of selection, l indicates current level number, kij (l)Indicate l layers of convolution kernel, * Indicate convolution algorithm, bj (l)Indicate biasing, f () indicates activation primitive；
For pond layer,
xj (l)=down (xj (l-1)) (2)
Wherein down () indicates down-sampling operation；
In the present embodiment, input that (224 × 224 be the resolution ratio of image, and 3 be figure for 224 × 224 × 3 small image of weld seam As depth or port number, that is, store digit used in each pixel), it include two convolutional layers in ConvA, each convolutional layer includes 64 3 × 3 convolution kernels, step-length 1, are filled with 1, the characteristic pattern that output is 64 224 × 224 after two convolutional layers. For pond layer using the method in maximum pond, the window size of down-sampling is 2 × 2, and the data that next layer is inputted behind pond are 64 112 × 112 characteristic pattern.Similarly, after two convolutional layers of ConvB (128 convolution kernels) and a pond layer, input The characteristic pattern that next layer of data are 128 56 × 56.ConvC, ConvD and ConvE include three convolutional layers, are passed through After ConvC, the characteristic pattern for 256 28 × 28 is exported, after ConvD, exports the characteristic pattern for 512 14 × 14, is passed through After three convolutional layers and a pond layer of ConvE, the characteristic pattern for 512 7 × 7 is exported.Resampling letter is passed through in this output Number carries out flattenings, is converted into one-dimensional vector, and length is 7 × 7 × 512=25088, then being input to nodal point number is 4096 complete Articulamentum FC1 exports the feature vector for 4096 × 1, then is input to the full articulamentum FC2 that nodal point number is 4096, output one A 4096 × 1 feature vector, the last one full articulamentum FC3 include 1000 neurons, the feature that output is one 1000 × 1 Vector.The depth of one 1000 dimension is finally extracted from a small image of weld seam using VGGNet-16 convolutional neural networks Practise feature；The characteristic sequence extracted from the small image of each weld seam is merged (if the weld seam egative film image that will newly submit If being divided into the small image of m weld seam, a small image of weld seam is characterized in that 1000 dimensions, the feature of the small image of m weld seam merge That m × 1000 is tieed up together), obtain the feature of whole weld image, and with save weld image feature in the database into Row compares to determine whether matching.
Specifically, the characteristic matching judgment basis of weld image is Pearson correlation coefficient J, if A is by VGGNet-16 The characteristics of weld seam vector for the weld seam egative film image that convolutional neural networks extracted newly submit, B are the weld seam saved in database The characteristics of weld seam vector of egative film image, A=(a1,a2,...,an), B=(b1,b2,...,bn), Pearson correlation coefficient J passes through Following formula obtains:
Wherein:
If J < 0.6, then it represents that do not match, submit weld seam egative film image and the characteristics of weld seam extracted to be stored in data for new In the respective field of library.When detecting first weld seam egative film image, because there is no characteristics of weld seam in database, therefore first weld seam bottom The characteristics of weld seam of picture can automatically save in the database, if J >=0.6, then it represents that have matching, detect the identical weldering of weld seam Egative film image is stitched, that is, detects to mark the fraud weld seam egative film for shooting old weld seam with new font.
The above are preferred embodiments of the present invention, all any changes made according to the technical solution of the present invention, and generated function is made When with range without departing from technical solution of the present invention, all belong to the scope of protection of the present invention.

Claims (7)

1. a kind of automatic testing method that the weld seam egative film based on deep learning is faked, characterized by the following steps:
S1: by X-ray colloid film scanning at digitlization weld seam egative film image；
S2: the weld seam egative film image of a batch 200 or more is screened, should include all weld seams to be detected in this batch of weld seam egative film image All characters being likely to occur in egative film；
S3: being split this batch of weld seam egative film image using Radon transform, obtains the small image of multiple characters, completes training character The production of collection；
S4: training, the study of deep learning are carried out to training character set using LeNet-5 convolutional neural networks, obtains weld seam word Accord with identification model；
S5: establishing database, for saving the new weld image feature and weld seam character string for submitting weld seam egative film image；
S6: being split the weld seam egative film image newly submitted using Radon transform, the weld seam egative film image newly submitted Multiple small images of character；
S7: the small image of character in S6 is inputted into weld seam character recognition model, identification extracts new submission weld seam egative film image Weld seam character string；
S8: the weld seam character string extracted in S7 and the weld seam character string saved in database are compared, and compare processing result Are as follows: if result mismatches, and the new weld seam character string for submitting weld seam egative film image is saved in the database；If result matches, Alarm, detects to mark identical weld seam egative film；
S9: image segmentation, scaling are carried out to the weld seam egative film image newly submitted, obtain the small image of multiple weld seams；
S10: carrying out deep learning feature extraction to the small image of weld seam in S9 using improved VGGNet-16 network, will be from every The characteristic sequence extracted in a small image of weld seam merges, and obtains the feature of whole weld image, and be stored in data Weld image feature in library is compared, comparison result processing are as follows: if result mismatch, will newly submit weld seam egative film image and The characteristics of weld seam extracted saves in the database, if result matches, alarms, detects the identical weld seam egative film of weld seam.
2. the automatic testing method that the weld seam egative film according to claim 1 based on deep learning is faked, it is characterised in that: The VGGNet-16 convolutional neural networks are made of five convolutional layer groups and three full articulamentums, the LeNet-5 convolutional Neural Network is made of two convolutional layers, two pond layers, a full articulamentum and one softmax layers.
3. the automatic testing method that the weld seam egative film according to claim 1 based on deep learning is faked, it is characterised in that: The specific steps that weld seam negative map picture is split using Radon transform are as follows:
Step a. is to weld seam negative map picture degree of comparing stretch processing；
Step b. carries out binary conversion treatment using OTSU algorithm to weld seam negative map picture；
Step c. automatically splits all characters, label in weld seam egative film image using Radon transform, obtains multiple characters Small image.
4. the automatic testing method that the weld seam egative film based on deep learning is faked as claimed in claim 3, it is characterised in that: into The small image of weld seam character split is handled using mathematical Morphology Algorithm after the row step c.
5. the automatic testing method that the weld seam egative film based on deep learning is faked as claimed in claim 4, it is characterised in that: right Using the small image rotation processing of Morphology Algorithm treated weld seam character, the small image of multiple characters is obtained.
6. the automatic testing method that the weld seam egative film based on deep learning is faked as described in claim 1, it is characterised in that: institute S7 is stated to specifically comprise the following steps:
The small image of character in S6 is inputted into character recognition model, the corresponding character of the small image of character is identified, what is identified Character is combined classification in order, obtains a character string, i.e., the weld seam character for the weld seam egative film image newly submitted.
7. the automatic testing method that the weld seam egative film based on deep learning is faked as described in claim 1, it is characterised in that: institute S9 is stated to specifically comprise the following steps:
S91. weld seam egative film image degree of the comparing stretch processing to newly submitting；
S92. to treated, weld seam egative film image uses Radon transform to be divided into the small image of multiple weld seams；
S93. image scaling, the size of the unified small image of weld seam are carried out to the small image of weld seam after segmentation；
S94. negative process is carried out to the small image of weld seam after scaling.
CN201910370302.0A 2019-05-06 2019-05-06 Automatic detection method for welding seam negative film counterfeiting based on deep learning Active CN110097547B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910370302.0A CN110097547B (en) 2019-05-06 2019-05-06 Automatic detection method for welding seam negative film counterfeiting based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910370302.0A CN110097547B (en) 2019-05-06 2019-05-06 Automatic detection method for welding seam negative film counterfeiting based on deep learning

Publications (2)

Publication Number Publication Date
CN110097547A true CN110097547A (en) 2019-08-06
CN110097547B CN110097547B (en) 2022-09-06

Family

ID=67446908

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910370302.0A Active CN110097547B (en) 2019-05-06 2019-05-06 Automatic detection method for welding seam negative film counterfeiting based on deep learning

Country Status (1)

Country Link
CN (1) CN110097547B (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111091538A (en) * 2019-12-04 2020-05-01 上海君睿信息技术有限公司 Method and device for automatically identifying and detecting pipeline welding seam and defect
CN111199194A (en) * 2019-12-25 2020-05-26 吉林大学 Automobile intelligent cabin instrument testing method based on machine vision and deep learning
CN112102255A (en) * 2020-08-21 2020-12-18 杭州培慕科技有限公司 Intelligent defect rating method based on X-ray imaging image in industrial scene
CN113376172A (en) * 2021-07-05 2021-09-10 四川大学 Welding seam defect detection system based on vision and eddy current and detection method thereof
CN113392833A (en) * 2021-06-10 2021-09-14 沈阳派得林科技有限责任公司 Method for identifying type number of industrial radiographic negative image
CN115690447A (en) * 2022-11-04 2023-02-03 北京远舢智能科技有限公司 Image quality meter detection method and device, computer equipment and storage medium
CN116110071A (en) * 2023-04-07 2023-05-12 济南大学 Image format pipeline and instrument diagram pipeline identification method based on deep learning
CN113392833B (en) * 2021-06-10 2024-06-28 沈阳派得林科技有限责任公司 Industrial ray film image type number identification method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2009066023A2 (en) * 2007-09-05 2009-05-28 Institut De Soudure Collimator for a non-destructive testing device using gamma radiography
CN109285139A (en) * 2018-07-23 2019-01-29 同济大学 A kind of x-ray imaging weld inspection method based on deep learning

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2009066023A2 (en) * 2007-09-05 2009-05-28 Institut De Soudure Collimator for a non-destructive testing device using gamma radiography
CN109285139A (en) * 2018-07-23 2019-01-29 同济大学 A kind of x-ray imaging weld inspection method based on deep learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
赵淑珍: "驻厂监检中射线底片审查过程不容忽视的问题", 《无损探伤》 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111091538A (en) * 2019-12-04 2020-05-01 上海君睿信息技术有限公司 Method and device for automatically identifying and detecting pipeline welding seam and defect
CN111091538B (en) * 2019-12-04 2023-06-09 上海君睿信息技术有限公司 Automatic identification and defect detection method and device for pipeline welding seams
CN111199194A (en) * 2019-12-25 2020-05-26 吉林大学 Automobile intelligent cabin instrument testing method based on machine vision and deep learning
CN112102255A (en) * 2020-08-21 2020-12-18 杭州培慕科技有限公司 Intelligent defect rating method based on X-ray imaging image in industrial scene
CN112102255B (en) * 2020-08-21 2024-01-23 杭州培慕科技有限公司 Intelligent defect grading method based on X-ray imaging image in industrial scene
CN113392833A (en) * 2021-06-10 2021-09-14 沈阳派得林科技有限责任公司 Method for identifying type number of industrial radiographic negative image
CN113392833B (en) * 2021-06-10 2024-06-28 沈阳派得林科技有限责任公司 Industrial ray film image type number identification method
CN113376172A (en) * 2021-07-05 2021-09-10 四川大学 Welding seam defect detection system based on vision and eddy current and detection method thereof
CN115690447A (en) * 2022-11-04 2023-02-03 北京远舢智能科技有限公司 Image quality meter detection method and device, computer equipment and storage medium
CN116110071A (en) * 2023-04-07 2023-05-12 济南大学 Image format pipeline and instrument diagram pipeline identification method based on deep learning
CN116110071B (en) * 2023-04-07 2023-09-12 济南大学 Image format pipeline and instrument diagram pipeline identification method based on deep learning

Also Published As

Publication number Publication date
CN110097547B (en) 2022-09-06

Similar Documents

Publication Publication Date Title
CN110097547A (en) A kind of automatic testing method that the weld seam egative film based on deep learning is faked
CN111091538B (en) Automatic identification and defect detection method and device for pipeline welding seams
CN105069790A (en) Rapid imaging detection method for gear appearance defect
Gu et al. Automatic and robust object detection in x-ray baggage inspection using deep convolutional neural networks
Tyystjärvi et al. Automated defect detection in digital radiography of aerospace welds using deep learning
CN109544522A (en) A kind of Surface Defects in Steel Plate detection method and system
Dong et al. A random forest-based automatic inspection system for aerospace welds in X-ray images
CN114372955A (en) Casting defect X-ray diagram automatic identification method based on improved neural network
Nizam et al. Vision based identification and classification of weld defects in welding environments: A review
CN115601355A (en) Method and device for detecting and classifying product surface defects and storage medium
CN113643268A (en) Industrial product defect quality inspection method and device based on deep learning and storage medium
CN116091505B (en) Automatic defect detection and classification method and system for sapphire substrate
Yao et al. A feature memory rearrangement network for visual inspection of textured surface defects toward edge intelligent manufacturing
CN115184380A (en) Printed circuit board welding spot abnormity detection method based on machine vision
CN113947583B (en) Weld joint nondestructive testing method based on deep learning two-dimensional time sequence image
CN112017154A (en) Ray defect detection method based on Mask R-CNN model
Patil et al. An autonomous technique for multi class weld imperfections detection and classification by support vector machine
Muresan et al. Automatic vision inspection solution for the manufacturing process of automotive components through plastic injection molding
CN114511543A (en) Intelligent defect evaluation system and method for radiographic negative of long-distance pipeline
CN113506281A (en) Bridge crack detection method based on deep learning framework
CN111179278B (en) Image detection method, device, equipment and storage medium
CN116721080A (en) Small-diameter pipe welding defect detection method based on shape and position sensing and balanced sampling
Yang et al. Weld defect cascaded detection model based on bidirectional multi-scale feature fusion and shape pre-classification
CN114078106A (en) Defect detection method based on improved Faster R-CNN
US20230298152A1 (en) Method for analyzing minor defect based on progressive segmentation network

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