阅读说明：本技术 一种工业射线底片的假片识别方法 (False film identification method for industrial radiographic negative film ) 是由 杨波 刘智 袁博 李超 高武奇 郭翔宇 许韫韬 李健衡 于 2021-07-20 设计创作，主要内容包括：本发明涉及深度学习及图像处理技术领域,具体涉及一种工业射线底片的假片识别方法。本发明提出的假片识别方法进行输入一张基准片和一张待判别片,首先进行两次高斯差分法分别提取两张底片的焊缝区域,在焊缝区域提取包含中心点及以焊缝中心点对称的五块感兴趣区域,并对感兴趣区域进行伪彩色处理；其次,将伪彩色处理过的感兴趣区域送入改进过的孪生神经网络训练识别,得出两两对比后的数值；最后,建立相似度评判体系,进行自定义权值运算,大于设定的阈值判定为假片。经试验验证,该方法可以有效的判别出工业底片中的假片,在261张测试集样本中,准确率可达99.1%以上,该方法简洁有效,具有一定的实际可行性。(The invention relates to the technical field of deep learning and image processing, in particular to a false film identification method for an industrial radiographic negative film. The false film identification method provided by the invention inputs a reference film and a to-be-distinguished film, firstly, a Gaussian difference method is carried out twice to respectively extract the welding seam areas of two negative films, five interesting areas which comprise a central point and are symmetrical by the central point of the welding seam are extracted in the welding seam areas, and the interesting areas are subjected to pseudo-color processing; secondly, sending the region of interest processed by the pseudo color into an improved twin neural network for training and recognition to obtain a numerical value after pairwise comparison; and finally, establishing a similarity evaluation system, carrying out self-defined weight calculation, and judging that the similarity is greater than a set threshold value as a false film. Tests prove that the method can effectively judge the false film in the industrial negative film, the accuracy rate can reach more than 99.1% in 261 test set samples, and the method is simple and effective and has certain practical feasibility.)

1. A false film identification method of an industrial radiographic film is characterized in that: the method comprises the following steps:

1) performing Gaussian difference method to extract a weld joint region, extracting five regions of interest which comprise a central point and are symmetrical by the central point of the weld joint, and performing pseudo-color processing on the regions of interest;

2) sending the region of interest processed by the pseudo color into an improved twin neural network for training and recognition to obtain the numerical values after two-by-two judgment;

3) and performing self-defined weight calculation, and judging that the weight is greater than the threshold value to be false.

2. The method for identifying the false negative of the industrial radiographic negative of claim 1, wherein the method comprises the following steps: the specific method comprises the following steps:

1) inputting two negative film images which are respectively a reference film and a to-be-distinguished film, respectively dividing the two images into a plurality of square areas with the width of the negative film as the side length, extracting a welding seam area in the square by using a Gaussian difference method, extracting a complete welding seam area by counting the longitudinal upper and lower coordinates of each welding seam area, establishing an interested area extraction rule, extracting five interested areas of each welding seam, and calibrating the sequence number of the interested areas;

2) carrying out pseudo-color processing on the five regions of interest extracted from each welding seam, then sending the regions of interest into an improved twin neural network for training and recognition according to a calibrated corresponding relation, and carrying out quantization processing on the result to obtain one-to-one corresponding similarity probability values;

3) and performing weighted operation on the obtained similarity probability values and the self-defined weight one by one, outputting a real similarity judgment value, and judging as a false film if the similarity probability value is greater than a similarity judgment threshold value.

3. A method for identifying false negative of industrial radiographic film according to claim 1 or 2, wherein: the method for extracting the weld joint region by the Gaussian difference method in the step 1) comprises the following steps:

1-1) regional block extraction: establishing a coordinate system by taking the upper left corner of the extracted square region as a coordinate origin, performing quartering treatment on the transverse direction and the longitudinal direction respectively, taking the length of the vertex at the upper right corner and the vertex at the lower left corner of the square region from the coordinate origin as 1, and taking a part region of the transverse direction and the longitudinal direction (1/4, 3/4);

2-2) two Gaussian difference processes: inputting P partial areas, i.e. input matrix is I "_{local1/2×1/2}Smoothing the image using two Gaussian low-pass filters, respectively, with a smoothing parameter of δ₁＝5、δ₂10, then, the image after smoothing is subjected to subtraction, so that the image characteristics are more stable, and obvious wave crests appear in an effective welding line area, so that the image is easier to distinguish;

3-3) determining X-axis intersection coordinates: determining the coordinates of the intersection point of the curve and the X axis, wherein the left side coordinates are longitudinal coordinates of the upper coordinates of the welding seam, and the right side coordinates are longitudinal coordinates of the lower coordinates;

4-4) positioning of a welding seam area: the weld joint region located by the gaussian difference method is a region contained in two longitudinal line segments.

4. The method for identifying the false negative of the industrial radiographic negative of claim 3, wherein: the extraction method of the welding seam area in the step 1) comprises the following steps:

and pre-extracting the weld joint region of the block region to obtain upper and lower coordinates for positioning the weld joint region, sequencing longitudinal coordinates of all the upper and lower coordinates, determining a longitudinal coordinate interval according to the minimum value of the longitudinal coordinate of the upper coordinate and the maximum value of the longitudinal coordinate of the lower coordinate, and accurately extracting the weld joint region according to the longitudinal interval of the upper and lower coordinates.

5. The method for identifying the false negative of the industrial radiographic negative of claim 4, wherein the method comprises the following steps: the extraction rule of the region of interest in the step 1) is as follows: 5 square areas are extracted from the two substrates at equal intervals, and the side length of the five square areas is the width of the extracted welding line.

6. The method for identifying the false negative of the industrial radiographic negative of claim 5, wherein: in the step 2):

the pseudo color processing is to transform the monochrome image into an image with a given color distribution;

the twin neural network is a coupling framework established based on two artificial neural networks, takes two samples as input, outputs the characterization of embedding high-dimensional space of the two samples, compares the similarity degree of the two samples, performs small sample/single sample learning, and is not easily interfered by wrong samples;

the trunk feature extraction network of the twin neural network is RepVGG.

7. The method for identifying the false negative of the industrial radiographic negative of claim 6, wherein the method comprises the following steps: the specific method of the step 3) is as follows:

comparing every two of five regions of interest respectively extracted from welding seams of the reference sheet and the sheet to be distinguished, carrying out point multiplication on the similarity probability value and a custom weight to obtain a real similarity judgment value, wherein the custom weight is w_Ai-BiThe similarity probability value is sim (A)_i-B_i) The true similarity evaluation value is Sim (A)_i-B_i)，(A_i-B_i) Represents A_iAnd B_iI represents the region of interest number. The operation method is as the formula (5):

wherein, the weight value w is customized according to the invention_Ai-BiThe weighted values are respectively: w is a_A1-B1＝0.125，w_A2-B2＝0.125，w_A3-B3＝0.5，w_A4-B4＝0.125，w_A5-B5＝0.125

According to the formula, a real similarity evaluation value is calculated, the threshold value is set to be 0.900, and when the real similarity evaluation value of the two industrial radiographic films is larger than the threshold value of 0.900, the two industrial radiographic films are judged to be similar, namely, the two industrial radiographic films are judged to be false films.

Technical Field

The invention relates to the technical field of deep learning and image processing, in particular to a false film identification method for an industrial radiographic film.

Background

With the rapid development of welding process, the quality of welding quality becomes an important concern for the manufacturing industry to judge welding safety. When the number of the welding seams is too large, the situations that part of the welding seams are not shot and one welding seam is shot for many times can occur during the digital imaging of the welding seams, so that industrial false pieces are generated, a large number of welding seams are not evaluated, and huge safety and economic losses are caused.

The problem of false-film identification in the industry is the similarity of the welds in the two negative images. The current working mode of distinguishing false films by human eyes is commonly adopted in the industry, so that the time and labor are wasted, the workload is large, the complexity is high, and the accuracy is low;

at present, the mainstream similarity judgment method is to compare by a histogram method, an image template matching method, a PSNR peak signal-to-noise ratio, ssim (structural similarity) structural similarity, a perceptual hash algorithm and the like, and the test has extremely poor robustness and low accuracy.

The conventional method has the problem that the welding quality evaluation is influenced by the judgment error, so that the safety problem is caused.

In order to overcome the defects of the existing method, the invention provides a false film identification method of an industrial radiographic film.

Disclosure of Invention

In view of the above, the present invention provides a false negative identification method for an industrial radiographic negative, which aims to solve the problems of various defects of the existing manual means and low accuracy rate of similarity evaluation of negative with a large length-width ratio by other methods at present, and ensures accuracy rate and efficiency by constructing a twin neural network, and then optimizing a data set and improving a trunk feature extraction network.

In order to solve the problems in the prior art, the technical scheme of the invention is as follows: a false film identification method of an industrial radiographic film is characterized in that: the method comprises the following steps:

1) performing Gaussian difference method to extract a weld joint region, extracting five regions of interest which comprise a central point and are symmetrical by the central point of the weld joint, and performing pseudo-color processing on the regions of interest;

2) sending the region of interest processed by the pseudo color into an improved twin neural network for training and recognition to obtain the numerical values after two-by-two judgment;

3) and performing self-defined weight calculation, and judging that the weight is greater than the threshold value to be false.

The specific method comprises the following steps:

1) inputting two negative film images which are respectively a reference film and a to-be-distinguished film, respectively dividing the two images into a plurality of square areas with the width of the negative film as the side length, extracting a welding seam area in the square by using a Gaussian difference method, extracting a complete welding seam area by counting the longitudinal upper and lower coordinates of each welding seam area, establishing an interested area extraction rule, extracting five interested areas of each welding seam, and calibrating the sequence number of the interested areas;

2) carrying out pseudo-color processing on the five regions of interest extracted from each welding seam, then sending the regions of interest into an improved twin neural network for training and recognition according to a calibrated corresponding relation, and carrying out quantization processing on the result to obtain a one-to-one corresponding similarity probability value;

3) and performing weighted operation on the obtained similarity probability values and the self-defined weight one by one, outputting a real similarity judgment value, and judging as a false film if the similarity probability value is greater than a similarity judgment threshold value.

The method for extracting the weld joint region by the Gaussian difference method in the step 1) comprises the following steps:

1-1) regional block extraction: establishing a coordinate system by taking the upper left corner of the extracted square region as a coordinate origin, performing quartering treatment on the transverse direction and the longitudinal direction respectively, taking the lengths of the vertex at the upper right corner and the vertex at the lower left corner of the square region from the coordinate origin as 1, and taking a part region of the transverse direction and the longitudinal direction (1/4, 3/4);

2-2) two Gaussian difference processes: inputting P partial areas, i.e. input matrix is I "_{local1/2×1/2}Smoothing the image using two Gaussian low-pass filters, respectively, with a smoothing parameter of δ₁＝5、δ₂10, then, the image after smoothing is subjected to subtraction, so that the image characteristics are more stable, and obvious wave crests appear in an effective welding line area, so that the image is easier to distinguish;

3-3) determining X-axis intersection coordinates: determining the coordinates of the intersection point of the curve and the X axis, wherein the left side coordinates are longitudinal coordinates of the upper coordinates of the welding seam, and the right side coordinates are longitudinal coordinates of the lower coordinates;

4-4) positioning of a welding seam area: the weld joint region located by the gaussian difference method is a region contained in two longitudinal line segments.

The method for extracting the weld joint region in the step 1) comprises the following steps:

and pre-extracting the weld zones of the partitioned zones to obtain upper and lower coordinates for positioning the weld zones, sequencing the longitudinal coordinates of all the upper and lower coordinates, determining a longitudinal coordinate interval according to the minimum value of the longitudinal coordinate of the upper coordinate and the maximum value of the longitudinal coordinate of the lower coordinate, and accurately extracting the weld zones according to the longitudinal intervals of the upper and lower coordinates.

The extraction rule of the region of interest in the step 1) is as follows: 5 square areas are extracted from the two substrates at equal intervals, and the side length of the five square areas is the width of the extracted welding line.

In the step 2) above:

the pseudo color processing is to transform the monochrome image into an image with a given color distribution;

the twin neural network is a coupling framework established based on two artificial neural networks, takes two samples as input, outputs the characterization of embedding high-dimensional space, compares the similarity degree of the two samples, performs small sample/single sample learning, and is not easily interfered by wrong samples;

the trunk feature extraction network of the twin neural network is RepVGG.

The specific method of the step 3) comprises the following steps:

comparing every two of five regions of interest respectively extracted from welding seams of the reference sheet and the sheet to be distinguished, carrying out point multiplication on the similarity probability value and the user-defined weight value to obtain a real similarity judgment value, wherein the user-defined weight value is w_Ai-BiThe similarity probability value is sim (A)_i-B_i) The true similarity evaluation value is Sim (A)_i-B_i)，(A_i-B_i) Represents A_iAnd B_iI represents the region of interest number. The operation method is as the formula (5):

wherein, the weight value w is customized according to the invention_Ai-BiThe weighted values are respectively: w is a_A1-B1＝0.125， w_A2-B2＝0.125，w_A3-B3＝0.5，w_A4-B4＝0.125，w_A5-B5＝0.125

According to the formula, a real similarity evaluation value is calculated, the threshold value is set to be 0.900, and when the real similarity evaluation value of the two industrial ray negative films is larger than the threshold value of 0.900, the two industrial ray negative films are judged to be similar, namely, the two industrial ray negative films are judged to be false.

Compared with the prior art, the invention has the following advantages:

1) aiming at the problems that the traditional threshold segmentation algorithm is not obvious in overlapping segmentation of different target gray values, is not obvious in gray difference and the like, the traditional welding seam threshold segmentation problem is converted into the image information statistical analysis problem, and the problem that a ray negative film, particularly an excessively bright welding seam area and an excessively dark welding seam area, is difficult to position is solved;

2) the invention adopts the twin neural network based on improvement and designs a set of negative film similarity evaluation system based on the self-defined weight, thereby ensuring the authenticity and reliability of similarity evaluation and providing a theoretical thought for the tasks of searching pictures by pictures and verifying human faces in the field of image processing;

3) the false sheet identification method and the welding seam area positioning method provided by the invention replace manpower, greatly improve the working efficiency and reduce the energy consumption;

4) the method can effectively judge the false film in the industrial negative film, the accuracy rate can reach more than 99.1% in 261 test set samples, and the method is simple and effective and has certain practical feasibility.

5) Because the length-width ratio of the industrial negative plates is large, the memory is large, the number of pixels is large, the number of characteristics is large, for a twin neural network, the judgment precision can be greatly influenced by directly inputting two negative plates for training and identification, and the judgment accuracy is further influenced, the method adopts a mode of extracting areas in blocks, and the time overhead can be greatly increased if the areas in blocks are too many; the situation that the block area is too small and the contingency is too large can occur, so that the judgment error is caused, and therefore under the condition that the precision is fully protected and the time overhead is reduced, through a large number of tests, the method selects and extracts five interesting areas, and is the selection with the highest cost performance.

Description of the drawings:

FIG. 1 is a flow chart of false positive assessment;

FIG. 2 is a schematic view of a negative image segmentation; wherein (a), (b), (c) and (d) are square areas with the width of the negative film image as the side length, and (e) is an extracted part with insufficient length;

FIG. 3 is a block extraction diagram;

FIG. 4 is a difference plot of the effective weld area; wherein the sample difference output image (a) is a square area difference image extracted from the film image segmentation image (a); the sample difference output image (b) is a square area difference image extracted from the film image segmentation image (b); the sample difference output graph (c) is a square area difference graph extracted from the film image segmentation graph (c); the sample differential output image (d) is a square area differential image extracted from the film image segmentation image (d); the sample difference output image (e) is a region difference image extracted from the film image segmentation image (e).

FIG. 5 is a graph of X-axis intersection coordinates; wherein the X-axis coordinate determination chart (a) is a vertical coordinate chart determined by a square area extracted from the negative film image segmentation chart (a); the X-axis coordinate determination diagram (b) is a vertical coordinate diagram determined by the square area extracted from the negative film image segmentation diagram (b); the X-axis coordinate determination diagram (c) is a vertical coordinate diagram determined by the square area extracted from the negative film image segmentation diagram (c); the X-axis coordinate determination diagram (d) is a vertical coordinate diagram determined by the square area extracted from the negative film image segmentation diagram (d); the X-axis coordinate determination chart (e) is a longitudinal coordinate chart determined by the extracted area of the negative film image segmentation chart (e).

FIG. 6 is a weld region positioning view; the welding seam area positioning diagram (a) is a welding seam positioning diagram in a square area extracted from the negative image segmentation diagram (a); the welding seam area positioning diagram (b) is a welding seam positioning diagram in a square area extracted from the negative image segmentation diagram (b); the welding seam area positioning diagram (c) is a welding seam positioning diagram in the square area extracted from the negative image segmentation diagram (c); the welding seam area positioning diagram (d) is a welding seam positioning diagram in a square area extracted from the base sheet image segmentation diagram (d); the welding seam area positioning diagram (e) is a welding seam positioning diagram in the area extracted from the negative image segmentation diagram (e);

FIG. 7 is a drawing for drawing a region of interest extraction rule; wherein (a) A1, A2, A3, A4 and A5 in a welding seam 1 figure are five square areas which are pre-extracted from a reference sheet and have the width of the welding seam as a side length respectively; (b) in the diagram of the weld joint 2, B1, B2, B3, B4 and B5 are five square areas which are pre-extracted by the judgment piece and take the width of the weld joint as a side length.

FIG. 8 is a region of interest extraction map; wherein (a) A1, A2, A3, A4 and A5 in the weld 1 figure are five square regions with the weld width as side length extracted from a reference sheet respectively; (b) in the weld 2, B1, B2, B3, B4 and B5 are five square regions extracted by the judgment piece and with the weld width as a side length.

FIG. 9 is a similarity comparison graph; wherein (a) is a comparison result chart of the similarity of A1 in a welding seam 1 and B5 in a welding seam 2; (b) the result is a graph comparing the similarity of A2 in weld joint 1 with B4 in weld joint 2; (c) the result is a graph comparing the similarity of A3 in weld joint 1 with B3 in weld joint 2; (d) the similarity comparison result of the A4 in the welding seam 1 and the B2 in the welding seam 2 is shown; (e) the result is a graph comparing the similarity of A5 in weld joint 1 with B1 in weld joint 2;

Detailed Description

The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. It is to be understood that the described embodiments are only a few embodiments of the invention, and not all 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.

The invention relates to a false film identification method of an industrial radiographic negative film, which comprises the following steps:

firstly, extracting a welding seam region by a Gaussian difference method, extracting five interested regions which comprise central points and are symmetrical by the central points of the welding seam in the welding seam region, and carrying out pseudo-color processing on the interested regions; secondly, sending the region of interest processed by the pseudo color into an improved twin neural network for training and recognition to obtain probability values after two-by-two judgment; and finally, performing user-defined weight calculation, and judging that the weight is larger than the threshold value to be a false film, as shown in figure 1.

The method specifically comprises the following steps:

step 1: region of interest extraction stage

Inputting two negative film images which are respectively a reference film and a to-be-distinguished film, respectively dividing the two images into a plurality of square areas with the width of the negative film as the side length, extracting a welding seam area in the square by using a Gaussian difference method, extracting a complete welding seam area by counting the longitudinal upper and lower coordinates of each welding seam area, establishing an interested area extraction rule, extracting five interested areas of each welding seam, and calibrating the interested areas according to the serial numbers; the main steps are as follows:

(1) inputting a sample piece and a discrimination piece;

(2) dividing two negative film images into a plurality of square areas;

as shown in fig. 2, the film image is divided into a plurality of square regions each having a film width as a side length, and the remaining ones having a length that is insufficient are extracted, as shown in fig. 2, the film image is divided into five regions, in fig. 2, (a), (b), (c) and (d) are square regions each having a film image width as a side length, and in fig. 2, (e) is a region where a foot portion is not extracted.

(3) Weld joint region pre-extraction by Gaussian difference method

The method is characterized in that the extraction of the weld area is a key step of the subsequent extraction of the interested area, the weld area is pre-extracted by a Gaussian difference regional blocking method, and the traditional weld threshold segmentation problem is converted into an image information statistical analysis problem. The method comprises the following specific steps:

a. region block extraction

Establishing a coordinate system with the extracted upper left corner of the square region as a coordinate origin, performing quartering processing on the horizontal direction and the vertical direction respectively, taking the length of the vertex at the upper right corner and the vertex at the lower left corner of the square region from the coordinate origin as 1, and taking a horizontal direction and a vertical direction (1/4, 3/4) partial region, namely a P partial region in the graph 3.

b. Two-time Gaussian difference processing

The algorithm inputs a local area of a weld joint to be positioned and outputs a positioned effective weld joint area. The invention sets the input matrix as I'_{local1/2×1/2}Converting the two-dimensional image feature into a one-dimensional image feature according to the formula (1), and recording as H_1/2×1And subjecting it to the treatment of formula (2) wherein H_i1Is H_1/2×1Middle (i)₁Data, H_MeanIs H_1/2×1Mean value of (A), H_stdIs H_1/2×1Standard deviation of (2).

And extracting edge features of the input image by a Gaussian difference operator, smoothing the image by using two Gaussian low-pass filters respectively, and then performing difference on the smoothed image. Input image feature H of the invention_1/2×1Performing two Gaussian differences as input, and performing transverse and longitudinal Gaussian difference processing according to formulas (3) and (4), wherein delta₂＞δ₁， δ₂＝10、δ₁5, the image characteristics are more stable, obvious wave crests appear in the effective welding seam area, and the image characteristics are easier to distinguish, as shown in fig. 4.

WhereinAs shown in equation (4):

c. determining X-axis intersection coordinates

As shown in fig. 5, coordinates of the intersection of the gaussian curve and the X-axis are determined, wherein the left-side coordinate is a longitudinal coordinate of the upper coordinate of the weld and the right-side coordinate is a longitudinal coordinate of the lower coordinate.

d. Weld seam region positioning

As shown in fig. 6, the weld region located by the gaussian difference method is a region included in the two longitudinal line segments.

(4) Weld region extraction

And (3) performing weld joint region pre-extraction on the block regions to obtain upper and lower coordinates for weld joint region positioning, sequencing longitudinal coordinates of all the upper and lower coordinates, determining a longitudinal coordinate interval according to the minimum value of the longitudinal coordinate of the upper coordinate and the maximum value of the longitudinal coordinate of the lower coordinate, and accurately extracting the weld joint region according to the longitudinal interval of the upper and lower coordinates.

(5) Formulating region of interest extraction rules

As shown in fig. 7, the weld 1 extracted in the previous step. A1-A5 are five square areas with the extracted width of the weld as side length, A3 is a square with the middle point of the length of the weld as the center point, and the five rectangles are equally spaced. The same applies to weld 2 of FIG. 7, B1-B5. (Note: weld 2 is an image of weld 1 rotated 180.)

(6) Extracting five regions of interest

Step 2: similarity contrast phase

Carrying out pseudo-color processing on the five regions of interest extracted from each welding seam, sending the regions of interest into an improved twin neural network for training and recognition according to a calibrated corresponding relation of the situation one in the drawing, and carrying out quantitative processing on the result to obtain a one-to-one corresponding similarity probability value; the main steps are as follows:

(1) pseudo-color processing of regions of interest

And performing pseudo-color processing on the extracted region of interest, wherein the pseudo-color processing is to convert a monochrome image into an image with given color distribution, so that the detail resolution capability of human eyes on the image is improved, the purpose of image enhancement is achieved, the visual effect is obviously enhanced as can be seen from fig. 6, and the next processing is facilitated.

(2) Construction of data sets

The twin neural network is a coupling framework established based on two artificial neural networks, takes two samples as input, outputs the characterization of embedding high-dimensional space of the two samples, compares the similarity degree of the two samples, performs small sample/single sample learning, and is not easily interfered by wrong samples.

According to the characteristic features of the twin neural network, the data set of the invention adopts the interested areas under different resolutions and the interested areas processed by pseudo colors, and each group is classified into about 40 pieces, and the total number is 41 groups.

(3) Improved twin neural network based identification

In order to equally improve the speed and the precision of the network, the invention changes the main feature extraction network of the twin neural network from the traditional VGG to the RepMVGG.

The RepVGG is a classification network, is improved on the basis of a VGG network, and the main improvements comprise: (1) an Identity and a residual branch are added into a Block Block of the VGG network, which is equivalent to applying essence in a ResNet network to the VGG network; (2) in the model reasoning phase, all network layers are converted into Conv3 x 3 through an Op fusion strategy, so that the deployment and acceleration of the model are facilitated.

The identification situation is shown in FIG. 9, wherein in FIG. 9, (a), (B), (c), (d), and (e) are similarity comparison results of A1-B5, A2-B4, A3-B3, A4-B2, and A5-B1, respectively.

(4) Outputting similarity probability value after pairwise comparison

And 3, step 3: user-defined weight calculation

And performing weighted operation on the obtained similarity probability values and the self-defined weights one by one, outputting a real similarity judgment value, and judging as a false film if the true similarity judgment value is greater than a similarity judgment threshold value.

(1) Establishment of similarity evaluation system

According to the research content of the invention, a similarity evaluation system is established. Through a large number of tests and data analysis, five regions of interest respectively extracted from the welding seams of the reference sheet and the sheet to be distinguished are compared pairwise. The similarity between the middle region of interest is generally high (namely, A3 in a welding seam 1 is compared with B3 in a welding seam 2 in a welding seam 7), and the invention provides that the similarity probability value is subjected to point multiplication with a user-defined weight value to obtain a real similarity judgment value. Wherein the user-defined weight is omega (A)_i-B_i) The similarity probability value is sim (A)_i-B_i) The true similarity evaluation value is Sim (A)_i-B_i)，(A_i-B_i) Represents A_iAnd B_iI represents the region of interest number. The operation method is as shown in formula (5):

wherein, the weight value w is customized according to the invention_Ai-BiThe weighted values are respectively: w is a_A1-B1＝0.125， w_A2-B2＝0.125，w_A3-B3＝0.5，w_A4-B4＝0.125，w_A5-B5＝0.125。

(2) Outputting a true similarity evaluation value

Calculating the real similarity evaluation value according to the formula

(3) Setting of threshold values

According to a large number of tests and experimental data analysis, the threshold value is set to 0.900.

(4) Evaluation of dummy wafers

When the real similarity evaluation value of the two industrial X-ray negative films is larger than the threshold value of 0.900, the two industrial X-ray negative films are identified as being similar, namely, false negative.

The false film identification method of the industrial radiographic negative provided by the invention converts the false film identification problem of the industrial radiographic negative into the negative interested region positioning problem based on the Gaussian difference method and the image similarity contrast problem based on the twin neural network, decomposes the complex problem into a plurality of simple problems, and greatly simplifies the false film identification problem in the engineering practice. The method provides a theoretical basis and an actual solution for a large number of false film identification works in the industrial radiographic films, has certain practical feasibility, and provides a feasible scheme for false film identification.

It should be understood that the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention, and that various modifications and alterations can be made therein by those skilled in the art without departing from the spirit of the present invention.