CN113409276A - Model acceleration method for eliminating redundant background based on mutual information registration - Google Patents
Model acceleration method for eliminating redundant background based on mutual information registration Download PDFInfo
- Publication number
- CN113409276A CN113409276A CN202110689843.7A CN202110689843A CN113409276A CN 113409276 A CN113409276 A CN 113409276A CN 202110689843 A CN202110689843 A CN 202110689843A CN 113409276 A CN113409276 A CN 113409276A
- Authority
- CN
- China
- Prior art keywords
- image
- data set
- sample
- mutual information
- redundant
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 66
- 230000001133 acceleration Effects 0.000 title claims abstract description 15
- 239000013598 vector Substances 0.000 claims abstract description 24
- 238000004519 manufacturing process Methods 0.000 claims abstract description 16
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 11
- 238000001514 detection method Methods 0.000 claims abstract description 10
- 238000009776 industrial production Methods 0.000 claims abstract description 10
- 238000004364 calculation method Methods 0.000 claims abstract description 7
- 239000011159 matrix material Substances 0.000 claims description 51
- 238000012549 training Methods 0.000 claims description 28
- 230000008569 process Effects 0.000 claims description 15
- 238000001914 filtration Methods 0.000 claims description 6
- 238000002372 labelling Methods 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 238000003384 imaging method Methods 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 238000013473 artificial intelligence Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 206010021033 Hypomenorrhoea Diseases 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000008280 blood Substances 0.000 description 1
- 210000004369 blood Anatomy 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Images
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- 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
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
-
- 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
- G06T2207/20076—Probabilistic image processing
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Health & Medical Sciences (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)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a model acceleration method for eliminating redundant background based on mutual information registration, which is characterized in that a data set of a target detection task in an industrial scene is collected on the basis of the fixation of collection equipment and under the light diversity of a production environment; selecting a batch data set, selecting an image as a reference after determining the position of a camera on site, wherein the data set is hereinafter referred to as a reference, and carrying out mutual information image registration on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed; and calculating joint probability density of two-dimensional vectors of the two gray-scale images by using a histogram estimation method for rigid registration with reference to the image and the sample image so as to determine and skip redundant backgrounds for inference calculation. The invention improves the calculation efficiency of the convolutional neural network and meets the real-time property of target detection of the industrial production line.
Description
Technical Field
The invention relates to the technical field of industrial computer application, in particular to a model acceleration method for eliminating redundant background based on mutual information registration.
Background
The industrialization and the informatization are continuously fused and developed, and fresh blood is injected into industrial production by the Internet of things and the artificial intelligence technology. The complexity of industrial field environments and the need for real-time production line product testing present challenges to the application of artificial intelligence technology. The deployment edge equipment of the artificial intelligence technology needs to meet the requirements of reliability, safety, certain computing power and the like, particularly certain computing performance and low-delay conditions, and the reasoning speed of the algorithm model target detection directly influences the efficiency of the production line. The model inputs the acquired industrial field real-time image and needs to perform some mathematical calculations, such as convolution, integral transformation, norm and the like. The convolution plays an important role in extracting target features in digital image processing, and is also one of main components of reasoning time, and particularly when the features of a complex target are extracted, a deeper convolutional neural network is needed to better extract the features, and meanwhile, the reasoning time is greatly increased, so that a plurality of challenges are brought to the algorithm model deployment meeting the real-time requirements of a production line.
Disclosure of Invention
The invention aims to provide a model acceleration method for eliminating redundant background based on mutual information registration, and in order to realize the aim, the invention provides the following technical scheme: the model acceleration method for eliminating the redundant background based on mutual information registration comprises the following steps:
s1, acquiring a data set of a target detection task in an industrial scene, fixing the position of image acquisition equipment, and making the proportion of a product in which a detected target is located in an image as large as possible under the condition of meeting the field requirement of an industrial production line; the data set acquisition needs to be carried out on the basis of the fixed acquisition equipment and under different light conditions, such as the data sets of light diversity of production environments including daytime and night lamplight and the like in different production time periods;
s2, selecting a batch data set, selecting an image as a reference after the position of the camera is determined on site, wherein the data set is referred to as a reference in the following, and carrying out mutual information image registration on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed;
s3, calculating the joint probability density of the two-dimensional vectors of the two gray-scale images by using a histogram estimation method due to the rigid registration of the reference image and the sample image;
s4, according to the result of the histogram estimation method, when the probability density reaches a certain value, the overlapping degree of the sample image and the reference image is high; making a difference between the corresponding positions of the sample picture identified as the same background and the reference picture, judging whether the difference is within a certain range according to the difference, determining a redundant background, and using matrix binarization representation to obtain a binarization matrix T (T belongs to C)n×n)；
S5, labeling the acquired data set, inputting the data set into a convolutional neural network for training, and correspondingly multiplying the data set by a binarization matrix T' when the data set is trained and read so as to accelerate the training process;
s6, deploying the trained model on edge computing equipment to process inference of single image input, and filtering redundant backgrounds of input images through a binarization matrix T' after the single image input to accelerate the computational inference of the model.
Preferably, in step S1, the ratio of the product fixed by the image capturing device and the detected object in the image is as large as possible, which includes:
s11, the industrial production line is composed of a series of determined procedures, and the position change of products on the production line is relatively small, so that the position of the image acquisition equipment can be determined;
s12, the definition of the target in the image is influenced by the position of the image acquisition equipment, the larger the proportion of the target in the acquired image in the whole image is, the smaller the influence of noise points on a target detection task is, the easier the target characteristic is extracted, namely, the acquired image is an imaging target product as full as possible;
s13, the data set is collected at different periods of the production phase, including the daytime period and the night light period, so as to ensure the richness of the data set.
Preferably, a batch data set is selected, after the position of the camera is determined on site, an image is selected as a reference, hereinafter referred to as a reference, and mutual information image registration is carried out on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed, wherein the two-dimensional vectors are characterized in that:
s21, selecting a sample fixed by an industrial field camera as a reference;
s22, registering the reference image and each sample in the acquired data set through mutual information images, and graying the reference image and the current single contrast sample;
s23, constructing a two-dimensional vector (as shown in figure 4) corresponding to each pixel of the current single contrast sample by reference;
preferably, since the reference image and the sample image are in rigid registration, a joint probability density of two-dimensional vectors of the two grayscale images is calculated by using a histogram estimation method, and the method includes:
s31, because redundant backgrounds need to be separated from a single sample, images only undergo operations such as rotation and translation, and the distance between any two points of the target before and after transformation is unchanged, a rigid registration method is adopted;
s32, dividing the value range of each position value in a two-dimensional vector constructed by a single sample in a reference picture and a data set into equally spaced intervals on the basis of enough samples by adopting a non-parameter estimation method histogram estimation method, counting the number of samples in each interval, and calculating the probability density of each interval;
preferably, according to the result of the histogram estimation method, when the probability density reaches a certain value, the degree of overlap between the representation sample map and the reference map is high; the sample picture identified as the same background is differed with the corresponding position of the reference picture, whether the difference is within a certain range is judged according to the difference, and redundancy is determinedBackground, and using matrix binarization representation, a binarization matrix T is obtained (T is belonged to C)n×n)；
S41, determining that the sample image is a similar image if the probability density calculated by the histogram estimation method reaches a certain value and the overlap degree between the sample image and the reference image is high;
s42, making difference between the sample image and the corresponding pixel point of the reference image, determining whether the pixel point is a redundant background according to whether the pixel difference of the point is in a certain range, if the pixel difference of the point is in a certain range, setting 1 corresponding to the point in the matrix binarization, otherwise setting 0, to form a binarization matrix TkWherein k is represented as the kth sample;
s43, taking the intersection of the binarization matrix set T, namely performing OR operation on corresponding positions of any two binarization matrices to determine a final redundancy background binarization matrix T';
T＝{T1,T2,…，Tn},T’＝T1∩T2∩…∩Tn
preferably, target labeling is performed on the acquired data set, the data set is input into a convolutional neural network for training, and when the data set is trained and read, the data set is correspondingly multiplied by a binarization matrix T' so as to accelerate the training process, and the method is characterized by comprising the following steps of:
s51, firstly, negating each position of the binarization matrix T', wherein a certain pixel point value is 0 and represents a redundant background pixel, otherwise, reasoning calculation is needed;
s52, when a training sample is read in the training process, firstly, the training sample is multiplied by the corresponding position of the binarization matrix T ', and the pixel at the redundant position is multiplied by 0 at the corresponding position of the matrix T', so that the purpose of filtering the redundant background is achieved; multiplying the interested target area by 1 of the corresponding position of the matrix T', and reserving the pixel information of the original image; during training, the redundant background is skipped to carry out convolution operation, so that an interested target area is reduced, and the training of the model is accelerated;
preferably, the trained model is deployed on an edge computing device to process inference of single-image input, and after the single-image input, the input image is correspondingly multiplied by a binarization matrix T', so as to accelerate the computational inference of the model, and the method is characterized by comprising the following steps:
the trained model is deployed in edge computing equipment, when a single image is input for reasoning, the input image is correspondingly multiplied by a binary matrix T', redundant background pixels are filtered, invalid operation on the background in the convolutional neural network reasoning is avoided by eliminating the redundant background, the reasoning area is reduced, the convolutional operation scale is reduced, and the purposes of accelerating reasoning and reducing time delay are achieved.
Compared with the prior art, the invention has the following beneficial effects: when the model is deployed and reasoned on the industrial production line, the waste of computing power caused by invalid computation on redundant backgrounds is avoided, and the reasoned computation area is reduced. Before deployment, a collected industrial field data set is used for obtaining a union set of the redundant backgrounds under different environmental conditions at different time periods of the industrial field in a mutual information image registration mode. When the model is deployed for carrying out input real-time image reasoning, only convolution operation is carried out on image areas except for the redundant background, so that the calculation efficiency of the convolution neural network is improved, and the real-time performance of target detection of the industrial production line is met.
Drawings
FIG. 1 is a flow chart of mutual information image registration and redundant background extraction according to the present invention;
FIG. 2 is a schematic diagram of mutual information image registration according to the present invention;
FIG. 3 is a schematic diagram of the present invention for eliminating redundant background;
FIG. 4 is a diagram of the construction of a two-dimensional vector with reference to a reference image and a sample according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1 to 4, the present invention provides a technical solution of a model acceleration method for eliminating redundant background based on mutual information registration: the model acceleration method for eliminating the redundant background based on mutual information registration comprises the following steps:
s1, acquiring a data set of a target detection task in an industrial scene, fixing the position of image acquisition equipment, and making the proportion of a product in which a detected target is located in an image as large as possible under the condition of meeting the field requirement of an industrial production line; the data set acquisition needs to be carried out on the basis of the fixed acquisition equipment and under different light conditions, such as the data sets of light diversity of production environments including daytime and night lamplight and the like in different production time periods;
s2, selecting a batch data set, selecting an image as a reference after the position of the camera is determined on site, wherein the data set is referred to as a reference in the following, and carrying out mutual information image registration on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed;
s3, calculating the joint probability density of the two-dimensional vectors of the two gray-scale images by using a histogram estimation method due to the rigid registration of the reference image and the sample image;
s4, according to the result of the histogram estimation method, when the probability density reaches a certain value, the overlapping degree of the sample image and the reference image is high; making a difference between the corresponding positions of the sample picture identified as the same background and the reference picture, judging whether the difference is within a certain range according to the difference, determining a redundant background, and using matrix binarization representation to obtain a binarization matrix T (T belongs to C)n×n)；
S5, labeling the acquired data set, inputting the data set into a convolutional neural network for training, and correspondingly multiplying the data set by a binarization matrix T' when the data set is trained and read so as to accelerate the training process;
s6, deploying the trained model on edge computing equipment to process inference of single image input, and filtering redundant backgrounds of input images through a binarization matrix T' after the single image input to accelerate the computational inference of the model.
In this embodiment, further, in step S1, the ratio of the product where the image capturing device is fixed and the detected target is located in the image is as large as possible, which includes:
s11, the industrial production line is composed of a series of determined procedures, and the position change of products on the production line is relatively small, so that the position of the image acquisition equipment can be determined;
s12, the definition of the target in the image is influenced by the position of the image acquisition equipment, the larger the proportion of the target in the acquired image in the whole image is, the smaller the influence of noise points on a target detection task is, the easier the target characteristic is extracted, namely, the acquired image is an imaging target product as full as possible;
s13, the data set is collected at different periods of the production phase, including the daytime period and the night light period, so as to ensure the richness of the data set.
In this embodiment, further, a batch data set is selected, after the position of the camera is determined on site, an image is selected as a reference, hereinafter referred to as a reference, and mutual information image registration is performed on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed, wherein the two-dimensional vectors are characterized in that:
s21, selecting a sample fixed by an industrial field camera as a reference;
s22, registering the reference image and each sample in the acquired data set through mutual information images, and graying the reference image and the current single contrast sample;
s23, constructing a two-dimensional vector (as shown in figure 4) corresponding to each pixel of the current single contrast sample by reference;
in this embodiment, further, since the reference image and the sample image are in rigid registration, a joint probability density of two-dimensional vectors of the two grayscale images is calculated by using a histogram estimation method, which includes:
s31, because redundant backgrounds need to be separated from a single sample, images only undergo operations such as rotation and translation, and the distance between any two points of the target before and after transformation is unchanged, a rigid registration method is adopted;
s32, dividing the value range of each position value in a two-dimensional vector constructed by a single sample in a reference picture and a data set into equally spaced intervals on the basis of enough samples by adopting a non-parameter estimation method histogram estimation method, counting the number of samples in each interval, and calculating the probability density of each interval;
in the present embodiment, further, according to the result of the histogram estimation method, when the probability density reaches a certain value, the degree of overlap between the representation sample map and the reference map is high; making a difference between the corresponding positions of the sample picture identified as the same background and the reference picture, judging whether the difference is within a certain range according to the difference, determining a redundant background, and using matrix binarization representation to obtain a binarization matrix T (T belongs to C)n ×n)；
S41, determining that the sample image is a similar image if the probability density calculated by the histogram estimation method reaches a certain value and the overlap degree between the sample image and the reference image is high;
s42, making difference between the sample image and the corresponding pixel point of the reference image, determining whether the pixel point is a redundant background according to whether the pixel difference of the point is in a certain range, if the pixel difference of the point is in a certain range, setting 1 corresponding to the point in the matrix binarization, otherwise setting 0, to form a binarization matrix TkWherein k is represented as the kth sample;
s43, taking the intersection of the binarization matrix set T, namely performing OR operation on corresponding positions of any two binarization matrices to determine a final redundancy background binarization matrix T';
T＝{T1,T2,…,Tn},T’＝T1∩T2∩…∩Tn
in this embodiment, further, target labeling is performed on the collected data set, the data set is input into a convolutional neural network for training, and when the data set is read in training, the data set is multiplied by a binarization matrix T' correspondingly to accelerate the training process, which is characterized by comprising:
s51, firstly, negating each position of the binarization matrix T', wherein a certain pixel point value is 0 and represents a redundant background pixel, otherwise, reasoning calculation is needed;
s52, when a training sample is read in the training process, firstly, the training sample is multiplied by the corresponding position of the binarization matrix T ', and the pixel at the redundant position is multiplied by 0 at the corresponding position of the matrix T', so that the purpose of filtering the redundant background is achieved; multiplying the interested target area by 1 of the corresponding position of the matrix T', and reserving the pixel information of the original image; during training, the redundant background is skipped to carry out convolution operation, so that an interested target area is reduced, and the training of the model is accelerated;
in this embodiment, further, after a trained model is deployed on an edge computing device to process inference of single image input, and after the single image input, an input image is correspondingly multiplied by a binarization matrix T', so as to accelerate computational inference of the model, which is characterized by including:
the trained model is deployed in edge computing equipment, when a single image is input for reasoning, the input image is correspondingly multiplied by a binary matrix T', redundant background pixels are filtered, invalid operation on the background in the convolutional neural network reasoning is avoided by eliminating the redundant background, the reasoning area is reduced, the convolutional operation scale is reduced, and the purposes of accelerating reasoning and reducing time delay are achieved.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (7)
1. The model acceleration method for eliminating the redundant background based on mutual information registration comprises the following steps:
s1, acquiring a data set of a target detection task in an industrial scene, fixing the position of image acquisition equipment, and making the proportion of a product in which a detected target is located in an image as large as possible under the condition of meeting the field requirement of an industrial production line; the data set acquisition needs to be carried out on the basis of the fixed acquisition equipment and under different light conditions, such as the data sets of light diversity of production environments including daytime and night lamplight and the like in different production time periods;
s2, selecting a batch data set, selecting an image as a reference after the position of the camera is determined on site, wherein the data set is referred to as a reference in the following, and carrying out mutual information image registration on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed;
s3, calculating the joint probability density of the two-dimensional vectors of the two gray-scale images by using a histogram estimation method due to the rigid registration of the reference image and the sample image;
s4, according to the result of the histogram estimation method, when the probability density reaches a certain value, the overlapping degree of the sample image and the reference image is high; making a difference between the corresponding positions of the sample picture identified as the same background and the reference picture, judging whether the difference is within a certain range according to the difference, determining a redundant background, and using matrix binarization representation to obtain a binarization matrix T (T belongs to C)n×n)。
S5, labeling the acquired data set, inputting the data set into a convolutional neural network for training, and correspondingly multiplying the data set by a binarization matrix T' when the data set is trained and read so as to accelerate the training process;
s6, deploying the trained model on edge computing equipment to process inference of single image input, filtering redundant backgrounds of input images through a binarization matrix T' after the single image input, and accelerating the computational inference of the model;
2. the mutual information registration-based model acceleration method for eliminating redundant backgrounds of claim 1, wherein in step S1, the ratio of the product where the image acquisition device is fixed and the detected object is located in the image is as large as possible, and the method comprises:
s11, the industrial production line is composed of a series of determined procedures, and the position change of products on the production line is relatively small, so that the position of the image acquisition equipment can be determined;
s12, the definition of the target in the image is influenced by the position of the image acquisition equipment, the larger the proportion of the target in the acquired image in the whole image is, the smaller the influence of noise points on a target detection task is, the easier the target characteristic is extracted, namely, the acquired image is an imaging target product as full as possible;
s13, collecting the data set at different time intervals in the production stage, wherein the data set comprises a day time interval and a night light time interval, so as to ensure the richness of the data set;
3. the mutual information registration-based model acceleration method for eliminating redundant backgrounds of claim 2, wherein a batch data set is selected, an image is selected as a reference after the position of a camera is determined on site, the image is hereinafter referred to as a reference, and mutual information image registration is performed on samples in the data set and the reference; when mutual information images are registered, a reference image and a certain sample image are converted into gray level images, and two-dimensional vectors of the two gray level images are constructed, wherein the two-dimensional vectors are characterized in that:
s21, selecting a sample fixed by an industrial field camera as a reference;
s22, registering the reference image and each sample in the acquired data set through mutual information images, and graying the reference image and the current single contrast sample;
s23, constructing a two-dimensional vector corresponding to each pixel of the current single contrast sample by reference;
4. the mutual information registration-based model acceleration method for eliminating redundant background according to claim 3, wherein a histogram estimation method is used to calculate the joint probability density of two-dimensional vectors of two gray images due to the rigid registration of the reference image and the sample image, and the method comprises:
s31, because redundant backgrounds need to be separated from a single sample, images only undergo operations such as rotation and translation, and the distance between any two points of the target before and after transformation is unchanged, a rigid registration method is adopted;
s32, dividing the value range of each position value in a two-dimensional vector constructed by a single sample in a reference picture and a data set into equally spaced intervals on the basis of enough samples by adopting a non-parameter estimation method histogram estimation method, counting the number of samples in each interval, and calculating the probability density of each interval;
5. the mutual information registration-based model acceleration method for eliminating redundant background according to claim 4, wherein according to the result of the histogram estimation method, when the probability density reaches a certain value, the overlap degree between the sample graph and the reference graph is high; making a difference between the corresponding positions of the sample picture identified as the same background and the reference picture, judging whether the difference is within a certain range according to the difference, determining a redundant background, and using matrix binarization representation to obtain a binarization matrix T (T belongs to C)n×n) The method comprises the following steps:
s41, determining that the sample image is a similar image if the probability density calculated by the histogram estimation method reaches a certain value and the overlap degree between the sample image and the reference image is high;
s42, making difference between the sample image and the corresponding pixel point of the reference image, determining whether the pixel point is a redundant background according to whether the pixel difference of the point is in a certain range, if the pixel difference of the point is in a certain range, setting 1 corresponding to the point in the matrix binarization, otherwise setting 0, to form a binarization matrix TkWherein k is represented as the kth sample;
s43, taking the intersection of the binarization matrix set T, namely performing OR operation on corresponding positions of any two binarization matrices to determine a final redundancy background binarization matrix T';
T＝{T1，T2，…，Tn}，T′＝T1∩T2∩…∩Tn
6. the mutual information registration-based model acceleration method for eliminating redundant backgrounds according to claim 5, wherein the collected data set is subjected to target labeling, the data set is input into a convolutional neural network for training, and when the data set is trained and read, the data set is correspondingly multiplied by a binarization matrix T' to accelerate the training process, and the method comprises the following steps:
s51, firstly, negating each position of the binarization matrix T', wherein a certain pixel point value is 0 and represents a redundant background pixel, otherwise, reasoning calculation is needed;
s52, when a training sample is read in the training process, firstly, the training sample is multiplied by the corresponding position of the binarization matrix T ', and the pixel at the redundant position is multiplied by 0 at the corresponding position of the matrix T', so that the purpose of filtering the redundant background is achieved; multiplying the interested target area by 1 of the corresponding position of the matrix T', and reserving the pixel information of the original image; during training, the redundant background is skipped to carry out convolution operation, so that an interested target area is reduced, and the training of the model is accelerated;
7. the mutual information registration-based model acceleration method for eliminating redundant backgrounds according to claim 6, wherein the trained model is deployed on an edge computing device to process inference of single image input, and after the single image input, the input image is correspondingly multiplied by a binarization matrix T', so as to accelerate the computational inference of the model, and the method is characterized by comprising the following steps:
the trained model is deployed in edge computing equipment, when a single image is input for reasoning, the input image is correspondingly multiplied by a binary matrix T', redundant background pixels are filtered, invalid operation on the background in the convolutional neural network reasoning is avoided by eliminating the redundant background, the reasoning area is reduced, the convolutional operation scale is reduced, and the purposes of accelerating reasoning and reducing time delay are achieved.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110689843.7A CN113409276A (en) | 2021-06-22 | 2021-06-22 | Model acceleration method for eliminating redundant background based on mutual information registration |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110689843.7A CN113409276A (en) | 2021-06-22 | 2021-06-22 | Model acceleration method for eliminating redundant background based on mutual information registration |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113409276A true CN113409276A (en) | 2021-09-17 |
Family
ID=77682206
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110689843.7A Pending CN113409276A (en) | 2021-06-22 | 2021-06-22 | Model acceleration method for eliminating redundant background based on mutual information registration |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113409276A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473755A (en) * | 2013-09-07 | 2013-12-25 | 西安电子科技大学 | SAR image sparsing denoising method based on change detection |
| CN106803070A (en) * | 2016-12-29 | 2017-06-06 | 北京理工雷科电子信息技术有限公司 | A kind of port area Ship Target change detecting method based on remote sensing images |
| CN106874949A (en) * | 2017-02-10 | 2017-06-20 | 华中科技大学 | A kind of moving platform moving target detecting method and system based on infrared image |
| CN112862866A (en) * | 2021-04-13 | 2021-05-28 | 湖北工业大学 | Image registration method and system based on sparrow search algorithm and computing equipment |
-
2021
- 2021-06-22 CN CN202110689843.7A patent/CN113409276A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473755A (en) * | 2013-09-07 | 2013-12-25 | 西安电子科技大学 | SAR image sparsing denoising method based on change detection |
| CN106803070A (en) * | 2016-12-29 | 2017-06-06 | 北京理工雷科电子信息技术有限公司 | A kind of port area Ship Target change detecting method based on remote sensing images |
| CN106874949A (en) * | 2017-02-10 | 2017-06-20 | 华中科技大学 | A kind of moving platform moving target detecting method and system based on infrared image |
| CN112862866A (en) * | 2021-04-13 | 2021-05-28 | 湖北工业大学 | Image registration method and system based on sparrow search algorithm and computing equipment |
Non-Patent Citations (2)
| Title |
|---|
| 李俊: ""动平台红外成像运动目标检测方法研究"", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
| 黄心汉: "《微装配机器人》", 31 July 2019 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378264B (en) | Target tracking method and device | |
| CN111046880B (en) | Infrared target image segmentation method, system, electronic equipment and storage medium | |
| CN111626176B (en) | Remote sensing target rapid detection method and system based on dynamic attention mechanism | |
| CN112733950A (en) | Power equipment fault diagnosis method based on combination of image fusion and target detection | |
| CN111369581A (en) | Image processing method, device, equipment and storage medium | |
| CN112446436A (en) | Anti-fuzzy unmanned vehicle multi-target tracking method based on generation countermeasure network | |
| Shu et al. | LVC-Net: Medical image segmentation with noisy label based on local visual cues | |
| CN115578378A (en) | Infrared and visible light image fusion photovoltaic defect detection method | |
| CN117561540A (en) | System and method for performing computer vision tasks using a sequence of frames | |
| CN110969173B (en) | Target classification method and device | |
| CN111415370A (en) | Embedded infrared complex scene target real-time tracking method and system | |
| Wang et al. | Intrusion detection for high-speed railways based on unsupervised anomaly detection models | |
| Sureshkumar et al. | Deep learning framework for component identification | |
| CN107358625B (en) | SAR image change detection method based on SPP Net and region-of-interest detection | |
| CN117576724A (en) | Unmanned plane bird detection method, system, equipment and medium | |
| CN112967293A (en) | Image semantic segmentation method and device and storage medium | |
| Maddileti et al. | Pseudo Trained YOLO R_CNN Model for Weapon Detection with a Real-Time Kaggle Dataset | |
| CN115147450B (en) | Moving target detection method and detection device based on motion frame difference image | |
| CN113409276A (en) | Model acceleration method for eliminating redundant background based on mutual information registration | |
| CN113506289B (en) | Method for classifying false positives of lung nodules by using double-flow network | |
| CN115311680A (en) | Human body image quality detection method and device, electronic equipment and storage medium | |
| CN113256556A (en) | Image selection method and device | |
| Truong et al. | A study on visual saliency detection in infrared images using Boolean map approach | |
| CN114373071A (en) | Target detection method and device and electronic equipment | |
| CN116612390B (en) | Information management system for constructional engineering |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210917 |
|
| RJ01 | Rejection of invention patent application after publication |