CN116665009A - Pipeline magnetic flux leakage image detection method based on multi-scale SSD network - Google Patents

Pipeline magnetic flux leakage image detection method based on multi-scale SSD network Download PDF

Info

Publication number
CN116665009A
CN116665009A CN202310411146.4A CN202310411146A CN116665009A CN 116665009 A CN116665009 A CN 116665009A CN 202310411146 A CN202310411146 A CN 202310411146A CN 116665009 A CN116665009 A CN 116665009A
Authority
CN
China
Prior art keywords
image
convolution
magnetic flux
ssd network
flux leakage
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310411146.4A
Other languages
Chinese (zh)
Inventor
王竹筠
孙天贺
袁浩
刘斌
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.)
Shenyang Aerospace University
Original Assignee
Shenyang Aerospace 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 Shenyang Aerospace University filed Critical Shenyang Aerospace University
Priority to CN202310411146.4A priority Critical patent/CN116665009A/en
Publication of CN116665009A publication Critical patent/CN116665009A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G01N27/83—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws by investigating stray magnetic fields
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/52—Scale-space analysis, e.g. wavelet analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/70—Labelling scene content, e.g. deriving syntactic or semantic representations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/30—Assessment of water resources

Abstract

The invention belongs to the technical field of magnetic flux leakage image processing, and discloses a pipeline magnetic flux leakage image detection method based on a multi-scale SSD network, which comprises the following steps of S1: acquiring a magnetic leakage signal in the pipeline through a magnetic leakage detector; s2: image enhancement is carried out by adopting a pseudo-color processing technology, and pseudo-color enhancement is carried out by adopting density segmentation; s3: adding a porous convolution and attention residual error module into the SSD network to obtain a new SSD network model; s4: and detecting the magnetic flux leakage image through the improved SSD network. The invention has scientific and reasonable design, the porous convolution is used for extracting the low-level and high-level semantic information characteristics of the small target, and then the image details and the context semantic information can be effectively learned by fusing the characteristics, so that the detection precision of the small target defect can be effectively improved, and the invention has great guiding effect on the efficient and accurate detection of the magnetic flux leakage signal.

Description

Pipeline magnetic flux leakage image detection method based on multi-scale SSD network
Technical Field
The invention belongs to the technical field of magnetic flux leakage image processing, and particularly relates to a pipeline magnetic flux leakage image detection method based on a multi-scale SSD network.
Background
The frequent occurrence of pipeline leakage accidents such as corrosion, abrasion, accidental damage and the like causes huge life and property loss and environmental pollution, and the magnetic flux leakage detection technology plays an important role in long-distance oil and gas exploitation. At present, most of the identification and analysis of magnetic leakage adopt manual work, but the magnetic leakage image has large data volume and various types, so that problems such as time consumption, leakage judgment, misjudgment and the like exist. Therefore, how to efficiently and accurately detect the magnetic flux leakage signal is a current urgent problem to be solved. The object detection is one of important research directions in the field of computer vision, and the traditional object detection method uses a classifier to classify the features after feature descriptors, so as to realize feature extraction. In the feature extraction stage, manual intervention is required to obtain feature information in an original image corresponding to the target. The algorithm depends on the characteristic design to a great extent, and has low efficiency and high error rate. Aiming at the defects of the traditional detection method, a target detection algorithm based on deep learning has become a new direction in the field of computer vision.
The SSD network model can accurately and rapidly detect a plurality of target objects with different dimensions, however, the SSD model has a problem of poor detection capability on small targets, and there is still room for improvement.
Based on the above, the invention provides an improved magnetic flux leakage image detection algorithm based on an SSD model.
Disclosure of Invention
In order to overcome the technical problems, the invention provides a pipeline magnetic flux leakage image detection method based on a multi-scale SSD network, which solves the problem that the current SSD model has poor detection capability on a small target.
The invention adopts the following technical scheme:
a pipeline magnetic flux leakage image detection method based on a multi-scale SSD network comprises the following steps:
s1: acquiring a magnetic leakage signal in the pipeline through a magnetic leakage detector;
s2: image enhancement is carried out by adopting a pseudo-color processing technology, and pseudo-color enhancement is carried out by adopting density segmentation;
s3: adding a porous convolution and attention residual error module into the SSD network to obtain a new SSD network model;
s4: and detecting the magnetic flux leakage image through the improved SSD network.
In S3, the SSD network consists of the first five layers and four additional convolution layers of the VGG-16 network, the porous convolution is added to Conv4_3 layers in the SSD network and changed into multi-scale expansion convolution layers, then attention is paid to feature graphs generated by the two convolution layers, a clipping layer is added to adapt to any size input, and therefore the size of a null convolution feature graph is identical to that of the former one, and the two feature graphs are multiplied pixel by pixel to obtain a fusion feature graph of semantic information of a bottom layer and a high layer.
In S3, the porous convolution is added with an r parameter on the basis of the original convolution, convolution fields with different sizes are obtained by controlling the size of the r parameter, and the new network acts on the same input through the convolution of 4 different r parameters to feel the characteristics of different fields.
In S3, each attention residual module contains mask branches and a trunk branch, which may be any current convolutional neural network model, the mask branches process the feature map and output attention feature maps having the same dimensions, and then combine the feature maps of the two branches by using a point multiplication operation to obtain a final output feature map.
In S3, in the shallow structure, the attention of the network is focused on the background and other areas, while in the deep structure, the attention profile of the network is focused on the object to be detected.
In mask branching, the processing of the feature map includes downsampling to quickly encode and acquire global features of the feature map, and upsampling combines the downsampled extracted global high-dimensional features with previously non-sampled features to produce context and high-low latitude features.
In S4, after the pseudo-color enhanced image is put into an improved model, the image is formed by 1,2, 4, 8 and 4-level porous convolution, simultaneously, conv3_3 and conv4_3 are used for extracting low-level features of the image, the conv3_3 layer stride is 4, the conv4_3 layer stride is 8, attention is paid to feature graphs generated by two convolution layers, a clipping layer is added to adapt to any size input, the size of the empty convolution feature graph is the same as that of the previous one, and the two feature graphs are multiplied pixel by pixel to obtain a fusion feature graph of bottom-layer and high-layer semantic information
Compared with the prior art, the invention has the beneficial effects that:
the multi-scale SSD network model is provided by integrating the multi-scale SSD network model with the multi-scale SSD network model, the multi-scale SSD network model is used for extracting low-level and high-level semantic information features of a small target, then image details and context semantic information can be effectively learned by fusing the features, the detection precision of the defect of the small target is improved, and the multi-scale SSD network model has a great guiding effect on detecting magnetic flux leakage signals efficiently and accurately.
Drawings
FIG. 1 is a schematic diagram of the steps of the present invention;
FIG. 2 is the pipeline leakage flux detection data in the actual project;
FIG. 3 is a detection result image; wherein (a) and (d) are input images, (b) and (e) are detection results of the SSD model, and (c) and (f) are detection results of the improved model.
Detailed Description
Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the drawings are illustrative only and are not to be construed as limiting the invention.
As shown in fig. 1, a method for detecting a pipeline magnetic flux leakage image based on a multi-scale SSD network includes the following steps:
s1: acquiring a magnetic leakage signal in a pipeline through a magnetic leakage detector:
the pipeline magnetic leakage internal detection is to fully magnetize the ferromagnetic pipeline by using a permanent magnet through a rigid electric brush to enable the ferromagnetic pipeline to be saturated or nearly saturated, and judge whether the pipeline is defective or not through the change of magnetic induction lines. If the pipe wall is free from defects, the magnetic induction lines are parallel to the inside of the pipe; if the pipe wall is defective, magnetic induction lines leak from the surface of the pipe, and magnetic leakage is caused. The magnetic leakage signals acquired by the magnetic sensor are imaged to determine the characteristics and the position of the defect.
S2: image enhancement is performed by adopting a pseudo-color processing technology, and pseudo-color enhancement is performed by adopting density segmentation:
after the magnetic flux leakage image is input, since the magnetic flux leakage signal is a three-dimensional signal, the curve and the gray scale image cannot completely display all data information. The pseudo-color processing technique is an effective image enhancement method. The resolution of image details can be improved, and clear and natural images can be obtained. The color values can be used as three-dimensional features of the leakage signal so that the complete data of the leakage signal can be displayed by color. The invention adopts pseudo-color processing technology to enhance the image and adopts density segmentation to enhance the pseudo-color. The gray level of the image is changed from 0 to M 0 Divided into N intervals I i (i=1, 2,3, n., and color C i Each section is assigned so that a pseudo color image can be generated. Assuming that the gray scale range of the original gray scale image is 0.ltoreq.f (x, y). Ltoreq.50, the gray scale range is divided into k segments, [ (I) 0 ，I 1 ，…，I k ),I 0 ＝0,I k ＝L]To map each gray level to one color, and the mapping relationship is as follows:
wherein g (x, y) is an output pseudo color image; c (C) i To map colors.
S3: adding a porous convolution and attention residual error module into the SSD network to obtain a new SSD network model;
the SSD network consists of the first five layers and four additional convolutional layers of the VGG-16 network. The porous convolution is added to Conv4_3 layer in the SSD network and modified to a multi-scale extended convolution layer. Then, note the feature map generated by the two convolution layers. The attention module enhances the saliency of the target; a clipping layer is added to accommodate any size input so that the size of the null convolution feature map is the same as the previous one. And multiplying the two feature maps pixel by pixel to obtain a fused feature map of the bottom-layer semantic information and the high-layer semantic information.
The enhanced pipeline magnetic flux leakage signal image is directly input into a new network model, so that the effect of improving the detection precision of the small target defect is achieved.
Wherein the porous convolution is added with an r parameter (rate) based on the original convolution, and convolution fields of different sizes are obtained by controlling the size of the rate.
Assuming that the standard convolution kernel size is k, the convolution kernel size of the porous convolution is shown in equation (2).
k n ＝k+(k-1)×(r-1) (2)
The new network acts on the same input through convolutions of 4 different rates to perceive the features of different views. The receptive field size calculation formula is shown in formula (3).
v ＝ ((k size + 1) × (r rate - 1) + k size ) 2 (3)
After porous convolution is carried out on the low-resolution feature, the resolution of the feature map is required to be consistent with that of the high-resolution feature map so as to finish feature map merging. The resolution calculation formula of the characteristic diagram after the porous convolution is shown in a formula (4).
O ＝S ×( L - 1)+H - 2 × P (4)
Wherein O is the resolution of the porous convolution output feature map; s is the step length; l represents the input feature map resolution; h represents the convolution kernel size; p represents the edge complement size.
To enhance the selection of the region of interest, an attention residual module is added to the original SSD network. Each attention residual module contains mask branches, which may be any current convolutional neural network model, and a backbone branch, which processes the feature maps and outputs attention feature maps having the same dimensions, and then combines the feature maps of the two branches by using a point multiplication operation to obtain a final output feature map.
In shallow structures, the attention of the network is focused on the background and other areas, while in deep structures, the attention profile of the network is focused on the object to be detected, which enables the deep profile to have higher abstract and semantic expressive power, and the attention value between 0 and 1 is obtained through a convolution kernel sigmoid function with a convolution kernel size of 1×1. The attention map of each convolution layer feature map is learned and weighted using the original feature map. The output characteristics of the remaining attention module are shown in fig. 5:
H i,c (x)＝T i,c (x)*M i,c (x). (5)
wherein the output characteristic diagram of the trunk branch is T i，c (x) The output characteristic diagram of the mask branch is M i And c (x) is [0,1 ]]Values in the interval.
To output a feature map with normalized weights, sigmoid is required as an activation function after mask branching. In mask branching, the processing of the feature map mainly includes downsampling and upsampling. Downsampling is used to quickly encode and acquire global features of feature maps, and upsampling combines the downsampled global high-dimensional features with previously non-sampled features to produce contextual and high-low latitude features. They can be better combined together.
S4: and (3) performing magnetic flux leakage image detection by using the improved SSD network:
after placing the pseudo-color enhanced image into the improved model, the image is formed by level 1,2, 4 and 8, 4 multi-hole convolution. Meanwhile, the low-level features of the image are extracted by using conv3_3 and conv4_3, the conv3_3 layer stride is 4, the conv4_3 layer stride is 8, and attention is paid to the feature map generated by the two convolution layers. The attention residual module enhances the saliency of the object, adding a clipping layer to accommodate any size input, so that the size of the null convolution feature map is the same as the previous one. And multiplying the two feature images pixel by pixel to obtain a fused feature image of the bottom layer semantic information and the high layer semantic information.
Sample { (x) for training set m group marker (1) ,y (1) ),…(x (m) ,y (m) ) Using y (i) ∈ {1,2, … k }, weight vector ω∈R n Mapping feature vectors to class label space x by inner product of vectors and input feature values x (i) ∈R n :
Given the input image I (I, j, k), the training network f 1 Parameter θ 1 Where W represents a matrix of network weights and b represents a network bias vector W m And H m-1 For the matrix, the calculation formula is:
f 1 (I (i,j,k) ,θ 1 )＝W m H m-1 (7)
the output capacity of the m hidden layer is as follows:
H m ＝max(0,(pool(W m H m-1 +b m ))) (8)
wherein m=2, h 0 ＝I b (i,j,k)。b m Is the offset vector of layer m, W m Is a weight matrix. Pool (-) is the largest pooling operation and max (-) is the activation function.
The characteristics of the network captured image are denoted as f 2 (I (i,j,k) ,θ 2 ) The output in combination with the low-level feature network is represented as follows:
F c ＝[f 1 (I (i,j,k) ,θ 1 ),f 2 (I (i,j,k) ,θ 2 )] (9)
symbol F c Representing the combined features, input I (i,j,k) 。
Will feature F c And transmitting the visual characteristic information to a subsequent network, and effectively fusing the visual characteristic information of the multi-scale characteristic map by expanding the convolution and modifying the linear activation function. In a multi-scale network, an output feature vector is obtained, and the feature vector F is trained by the network c Converting to conditional probabilities. The conditional probability distribution of each class C normalized prediction is calculated using a soft maximum regression function.
The invention is verified as follows:
where p is a conditional probability and c is a category.
The training dataset adopts a network dependent monitoring method and uses a minimum negative log likelihood function for training. The calculation formula is as follows:
I f(θ) ＝-∑lnp(l (i,j,k) |I (i,j,k) ,θ) (11)
wherein I is (i,j,k) Is image I k The correct class label for the (i, j) position. Learning optimal parameters using a random gradient descent algorithm and an inverse conduction iteration update error:
in SSD networks, a smooth loss function is employed for location regression training. The calculation formula is as follows:
g i for a region candidate box in four dimensions (x, y, w, h), l is the prediction offset and d is the region candidate box.The candidate box representing the i region matches the j true value, otherwise its value is 0.
Confidence loss is a multi-category problem with softmax. The calculation formula is as follows:
where default box i matches true bounding box j of category p isOtherwise, 0.
The objective loss function is a weight of the confidence loss and the position loss, and thus the objective loss function is obtained as:
wherein L is conf (x, c) represents confidence loss, L log (x, l, g) represents a position loss, and N is a default box number.
Finally, the location of the object in the image and the specific category are obtained by a non-maximal suppression (NMS).
The pipeline magnetic flux leakage image detection method based on the multi-scale SSD network provided by the invention is verified.
As shown in fig. 2, the experimental data is derived from the actual pipeline leakage detection data in the actual project. The data mileage is obtained by multiplying the sampling time by the number of data points acquired per unit time of the probe. The dataset was included in 2000 magnetic leakage image samples including 1800 girth welds, 600 spiral welds, 700 defects, and 500 non-target samples for experimental testing. During the training process, the algorithm needs to calibrate the positive and negative samples of the classification. The positive and negative samples are determined by the target boundary (groudtluth) of the marked picture and the predicted target boundary. If the IOU (intersection ratio) closure value of the two is 0.6, then the positive sample is set, otherwise the negative sample is set.
In fig. 3, a multi-target detection result image with a sampling point of 256×3200 is shown. The two multi-target images (a) and (d) are respectively input into the improved model and the SSD model, the images (b) - (f) are detection results, (b), (e) are detection results of the SSD model, and (c), (f) are detection results of the improved model. As can be seen from the figure, the prediction accuracy of the small defects in the improved model can reach over 96%, while in the original model, the prediction accuracy of the small defects is 85%. The prediction of the improved model on the girth weld can also reach 99%, while the original model is only 89%. By comparing the results, the accuracy of the improved model is higher than the original model. As can be seen from comparison of (b) and (c), the predicted value of the SSD model (b) of the small defect 1 is only 0.54, the small defect 2 and the girth weld 1 are omitted, the small defect 3 is misdiagnosed as the girth weld, and the predicted value of the small defect in the improved model (c) is higher than that in the SSD model; as can be seen from a comparison of (e) and (f), SSD (e) did not detect spiral weld 2, while improved model (f) was predicted with greater accuracy. By comparing the results of the two models, it can be found that the detection capability of the improved model is higher than that of the SSD model, especially for small target defects.
While embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that: many variations, modifications, substitutions and alterations are possible to the above embodiments without departing from the spirit and principles of the invention, the scope of which is defined by the claims and their equivalents.

Claims (7)

1. A pipeline magnetic flux leakage image detection method based on a multi-scale SSD network is characterized by comprising the following steps:
s1: acquiring a magnetic leakage signal in the pipeline through a magnetic leakage detector;
s2: image enhancement is carried out by adopting a pseudo-color processing technology, and pseudo-color enhancement is carried out by adopting density segmentation;
s3: adding a porous convolution and attention residual error module into the SSD network to obtain a new SSD network model;
s4: and detecting the magnetic flux leakage image through the improved SSD network.
2. The method for detecting pipeline magnetic flux leakage image based on multi-scale SSD network according to claim 1, wherein in S3, SSD network is composed of the first five layers and four additional convolution layers of VGG-16 network, porous convolution is added to Conv4_3 layer in SSD network and changed into multi-scale extended convolution layer, then note feature map generated by two convolution layers, and clipping layer is added to adapt to any size input, so that size of empty convolution feature map is same as that of the former one, and two feature maps are multiplied pixel by pixel to obtain fusion feature map of bottom layer and high layer semantic information.
3. The method for detecting pipeline magnetic flux leakage image based on multi-scale SSD network according to claim 1, wherein in S3, the porous convolution is added with an r parameter based on the original convolution, different-sized convolution fields are obtained by controlling the r parameter, and the new network acts on the same input through convolution of 4 different r parameters to feel the characteristics of different fields.
4. The method for detecting pipeline leakage magnetic image based on multi-scale SSD network according to claim 1, wherein in S3, each attention residual module comprises a mask branch and a main branch, the main branch may be any current convolutional neural network model, the mask branch processes the feature map and outputs the attention feature map having the same dimension, and then the feature maps of the two branches are combined by using a dot multiplication operation to obtain a final output feature map.
5. The method for detecting pipeline magnetic flux leakage image based on multi-scale SSD network according to claim 1, wherein in S3, the attention of the network is focused on the background and other areas in the shallow structure, and the attention profile of the network is focused on the object to be detected in the deep structure.
6. The method of claim 4, wherein in mask branching, the processing of the feature map includes downsampling and upsampling, the downsampling being used to quickly encode and obtain global features of the feature map, and the upsampling combining the downsampled extracted global high-dimensional features with previously un-sampled features to produce context and high-low latitude features.
7. The method for detecting pipeline magnetic flux leakage image based on multi-scale SSD network according to claim 1, wherein in S4, after the pseudo-color enhanced image is put into an improved model, the image is formed by 1,2, 4 and 8, 4-level multi-hole convolution, meanwhile, conv3_3 and conv4_3 are used for extracting low-layer features of the image, conv3_3 is 4 in layer stride, conv4_3 is 8 in layer stride, attention is paid to feature images generated by two convolution layers, a clipping layer is added to adapt to any size input, the size of the empty convolution feature image is the same as that of the previous one, and the two feature images are multiplied pixel by pixel to obtain a fusion feature image of bottom-layer semantic information and high-layer semantic information.
CN202310411146.4A 2023-04-18 2023-04-18 Pipeline magnetic flux leakage image detection method based on multi-scale SSD network Pending CN116665009A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310411146.4A CN116665009A (en) 2023-04-18 2023-04-18 Pipeline magnetic flux leakage image detection method based on multi-scale SSD network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310411146.4A CN116665009A (en) 2023-04-18 2023-04-18 Pipeline magnetic flux leakage image detection method based on multi-scale SSD network

Publications (1)

Publication Number Publication Date
CN116665009A true CN116665009A (en) 2023-08-29

Family

ID=87721336

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310411146.4A Pending CN116665009A (en) 2023-04-18 2023-04-18 Pipeline magnetic flux leakage image detection method based on multi-scale SSD network

Country Status (1)

Country Link
CN (1) CN116665009A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117420196A (en) * 2023-11-20 2024-01-19 中磁数智(北京)科技有限公司 Pipeline defect identification positioning method based on target detection and field weakening detection

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117420196A (en) * 2023-11-20 2024-01-19 中磁数智(北京)科技有限公司 Pipeline defect identification positioning method based on target detection and field weakening detection
CN117420196B (en) * 2023-11-20 2024-04-16 中磁数智(北京)科技有限公司 Pipeline defect identification positioning method based on target detection and field weakening detection

Similar Documents

Publication Publication Date Title
Yang et al. Pipeline magnetic flux leakage image detection algorithm based on multiscale SSD network
CN108830285B (en) Target detection method for reinforcement learning based on fast-RCNN
CN110135366B (en) Shielded pedestrian re-identification method based on multi-scale generation countermeasure network
CN109784203B (en) Method for inspecting contraband in weak supervision X-ray image based on layered propagation and activation
CN113436169B (en) Industrial equipment surface crack detection method and system based on semi-supervised semantic segmentation
CN111160311A (en) Yellow river ice semantic segmentation method based on multi-attention machine system double-flow fusion network
CN112364931B (en) Few-sample target detection method and network system based on meta-feature and weight adjustment
CN111368769B (en) Ship multi-target detection method based on improved anchor point frame generation model
CN110598767A (en) SSD convolutional neural network-based underground drainage pipeline defect identification method
CN111738113B (en) Road extraction method of high-resolution remote sensing image based on double-attention mechanism and semantic constraint
CN110543911A (en) weak supervision target segmentation method combined with classification task
CN112488025B (en) Double-temporal remote sensing image semantic change detection method based on multi-modal feature fusion
CN114627052A (en) Infrared image air leakage and liquid leakage detection method and system based on deep learning
CN110659601B (en) Depth full convolution network remote sensing image dense vehicle detection method based on central point
CN112766136B (en) Space parking space detection method based on deep learning
CN113313094B (en) Vehicle-mounted image target detection method and system based on convolutional neural network
CN116310850B (en) Remote sensing image target detection method based on improved RetinaNet
CN114913498A (en) Parallel multi-scale feature aggregation lane line detection method based on key point estimation
CN111259808A (en) Detection and identification method of traffic identification based on improved SSD algorithm
CN111539456B (en) Target identification method and device
Fan et al. A novel sonar target detection and classification algorithm
CN116665009A (en) Pipeline magnetic flux leakage image detection method based on multi-scale SSD network
CN115424237A (en) Forward vehicle identification and distance detection method based on deep learning
CN115147644A (en) Method, system, device and storage medium for training and describing image description model
CN115019201A (en) Weak and small target detection method based on feature refined depth 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