CN109493316B - An intelligent detection method for missing strips of cigarette packs in case sealing machine based on similarity analysis and feature incremental learning - Google Patents

Abstract

The invention discloses an intelligent detection method for missing strips in cigarette packs of a case sealing machine based on similarity analysis and feature incremental learning. In the offline modeling stage of this method, grayscale transformation and image segmentation are performed on the collected front and rear pictures of tobacco rods inside the cigarette pack, and the similarity analysis method of horizontal and vertical blocks is used to calculate the similarity of 16 features of the modeled pictures. The degree index is used to determine the threshold of the 16 features of the modeled image by using the kernel density estimation method. Online identification stage: The similarity analysis method of horizontal and vertical blocks is used to calculate the front and rear pictures of the tobacco rod inside the cigarette pack currently obtained in real time, and 16 feature similarity indicators of the current picture are obtained. By comparing with the current threshold , to identify whether there are missing bars in the front and rear pictures. The invention realizes the effective representation and quantification of the picture feature information of the tobacco rod inside the cigarette pack, can accurately judge the occurrence of the quality defect of the missing rod in the cigarette pack, and provides the information basis for the evidence of the missing rod through the picture mark.

Description

Intelligent detection method for cigarette packet missing strip of carton sealing machine based on similarity analysis and feature increment learning

Technical Field

The invention relates to the field of carton missing detection of carton sealing machine tobacco packets, in particular to an intelligent carton missing detection method of carton sealing machine tobacco packets based on similarity analysis and feature increment learning, which is suitable for cigarette carton sealing machine equipment.

Background

The carton cigarette packing is the last procedure of cigarette production, and the quality problem of cigarette packet strip shortage can occur in the packing process. Most cigarette enterprises strengthen process monitoring management, and detect the possible cigarette packet strip shortage quality problem in the packing process by adopting an electronic scale heavy type strip shortage detection method or a machine vision type strip shortage detection method through the transformation and upgrading of a box sealing machine. Along with the popularization of cigarette carton recycling projects in the tobacco industry, the deviation fluctuation of the weight of the recycling carton and the weight of the calibrated carton is large due to the reasons of moisture absorption in the use and recycling processes of the cartons and the like, and the accuracy of the heavy type strip missing detection method of the electronic scale is influenced. Meanwhile, the strip missing detection of the heavy cigarette packet with the electronic scale lacks the storage of strip missing evidences in the production process, and complaints about the strip missing quality problem are difficult to prove. The machine vision type strip missing detection method has the characteristics of accuracy, rapidness, good stability, strong adaptability, easiness in evidential demonstration and the like, and is concerned more recently. With the application of the novel strip-missing detection method and the novel strip-missing detection system, the strip-missing quality problem is reduced, but the strip-missing quality problem is not completely cured all the time, and almost several or even more than ten strip-missing complaints appear in a cigarette factory every year, so that the brand image of an enterprise is negatively influenced. Further study on a machine vision type strip missing detection method is needed, and the problem of strip missing quality of cigarette packets is effectively avoided.

Disclosure of Invention

In order to solve the technical problems, the invention aims to provide an intelligent detection method for cigarette packet missing strip of a carton sealing machine based on similarity analysis and feature increment learning, which realizes effective characterization and quantification of picture feature information; by means of feature increment learning of online recognition of pictures without missing strips, the threshold value of the similarity index is continuously updated, the quality defect of the tobacco packet missing strips is accurately judged, and an information basis is provided for missing strip evidences through picture marking.

In order to achieve the purpose, the invention adopts the following technical scheme:

the intelligent detection method for cigarette packet missing of the carton sealing machine based on similarity analysis and feature increment learning comprises the following steps:

(1) image graying and active area selection

Based on the front and rear imaging devices added on the carton sealing machine, when the push plate pushes the front 25 cigarettes into the carton completely, a front picture OF (N) is obtained_of×M_ofX 3) when the back 25 cigarettes are all pushed into the carton push plate to return, a back bitmap OB (N) is obtained_ob×M_obX 3) of the front and rear pictures into a gray image GF (N)_of×M_of) And GB (N)_ob×M_ob) Because the shooting angles of the front picture and the rear picture are relatively fixed aiming at a certain carton sealing machine, in order to quickly identify the effective area of the cigarette packet, the method of manually marking the effective area of the cigarette packet is adopted to respectively obtain the coordinates (X) of the upper left corners of the front picture and the rear picture_f、x_b) Upper left corner Y-axis coordinate (Y)_f、y_b) Width (w)_f、w_b) Height (h)_f、h_b) Waiting for 8 parameters to obtain a gray image PF (N) of the effective area of the tobacco bale_pf×M_pf) And PB (N)_pb×M_pb)；

(2) Similarity analysis based on blocking idea

Based on the gray level image of the effective area of the cigarette packet of the preceding picture, equally dividing the gray level image into 5 blocks according to the height of the image to obtain 5 lines of sub-images XPF_i(N_xpf×M_pf) (i-1, …,5), calculate each sub-image matrix XPF_iMean value of column vectors, obtaining vector XMPF of 5 rows of sub-images_i(1×M_pf) Calculating the similarity (rho) between every two adjacent rows by using the Pearson correlation coefficient_f12、ρ_f23、ρ_f34、ρ_f45) (ii) a Equally dividing the gray image based on the front picture cigarette packet effective area into 5 blocks according to the width of the image to obtain 5 columns of sub-images YPF_i(N_pf×M_ypf) (i 1, …,5), each sub-image matrix YPF is calculated_iObtaining the vector YMPF of 5 columns of sub-images by averaging the row vectors_i(N_pfX 1), calculating the similarity (rho ') between every two adjacent columns by using Pearson correlation coefficient'_f12、ρ′_f23、ρ′_f34、ρ′_f45)；

Calculating the bitmap cigarette carton by the same methodSimilarity (p) between every two adjacent lines of the active area gray scale image_b12、ρ_b23、ρ_b34、ρ_b45) And a similarity (ρ'_b12、ρ′_b23、ρ′_b34、ρ′_b45)；

(3) Modeling picture threshold calculation based on kernel density estimation

A similarity row vector rho can be calculated according to a front bit picture and a rear bit picture of a cigarette packet₁In turn from rho_f12、ρ_f23、ρ_f34、ρ_f45、ρ′_f12、ρ′_f23、ρ′_f34、ρ′_f45、ρ_b12、ρ_b23、ρ_b34、ρ_b45、ρ′_b12、ρ′_b23、ρ′_b34、ρ′_b45Elemental composition, collecting N_oCalculating the similarity row vector rho of each cigarette packet_j(1×16)(j＝1,…,N_o) Form a similarity matrix

Aiming at each row of data of the matrix, determining the distribution characteristics of the data by adopting a kernel density estimation method because the distribution condition of the data is unknown, and calculating to obtain a kernel density curve corresponding to each row; taking the similarity data at the leftmost side (namely, the minimum similarity) in the nuclear density curve as a threshold corresponding to the position in the modeling image to obtain a threshold vector theta (1 × 16);

(4) tobacco bale picture on-line identification based on similarity

Re-collecting front picture OF OF cigarette packet_TAnd a back bit slice OB_TConvert it into a gray image GF_TAnd GB_TRespectively obtaining gray front image PF of effective area of tobacco bale based on the coordinates of effective area of front image and rear image recognized in advance_TAnd a back picture PB_TThe gray front image PF_TDividing lines according to the height of the image, and calculating the similarity rho between adjacent lines_tf12、ρ_tf23、ρ_tf34、ρ_tf45Then, the image is divided into columns according to the image width, and the similarity ρ 'between adjacent columns is calculated'_tf12、ρ′_tf23、ρ′_tf34、ρ′_tf45Comparing the calculated 8 similarity indexes with the first 8 elements of the threshold vector theta one by one, if all the similarity indexes are larger than the corresponding elements of the threshold vector theta, identifying that the front image has no missing bars and performing corresponding marking on the image, otherwise identifying that the front image has missing bars and performing corresponding marking on the image, and identifying the gray level back image PB by adopting the same method_TWhether the strip is missing or not;

(5) feature incremental learning based threshold updating

For the tobacco bale image without the missing strips, after the identification by the technologist, the similarity row vector rho representing the tobacco bale image characteristics_TAdding the obtained similarity matrix into the existing similarity matrix R based on the updated similarity matrix

Recalculating to obtain a kernel density curve corresponding to each column, taking the leftmost (i.e. the smallest similarity) similarity data in the kernel density curves as the latest threshold corresponding to the position, and obtaining an updated threshold vector theta_T(1×16)。

The invention provides an intelligent detection method for cigarette packet strip missing of a carton sealing machine based on similarity analysis and feature increment learning, which comprises an offline modeling stage and an online identification stage. An off-line modeling stage: carrying out gray level transformation and image segmentation on tobacco bale internal front and rear position picture sheets collected by front and rear imaging devices of a cigarette pack of a carton sealing machine, calculating 16 characteristic similarity indexes of a modeling picture by adopting a picture transverse and longitudinal blocking similarity analysis method, and determining a threshold value of 16 characteristics of the modeling picture by utilizing a kernel density estimation method. And (3) an online identification stage: and calculating the front position and the rear position of the cigarette carton in the currently acquired cigarette packet in real time by adopting a similarity analysis method of horizontal and vertical partitioning, acquiring 16 characteristic similarity indexes of the current picture, and identifying whether the front position and the rear position have carton defects or not by comparing the similarity indexes with a current threshold value. And for the tobacco bale images which are not stripped on line, after the identification by process personnel, adding the similarity row vector which represents the tobacco bale image characteristics to the existing similarity matrix, and updating the threshold value by using a kernel density estimation method.

The invention has the beneficial effects that: the invention provides an intelligent detection method for cigarette packet strip shortage of a carton sealing machine based on similarity analysis and characteristic increment learning, aiming at the problem that the quality defect of the cigarette packet strip shortage occasionally occurs and based on front and rear pictures of cigarette strips in the cigarette packet, which are acquired by front and rear imaging devices of the cigarette packet of the carton sealing machine. Effective characterization and quantification of picture characteristic information are realized through the horizontal and vertical block division of the front bit and the rear bit picture and the similarity index calculation between the blocks; by means of feature increment learning of online recognition of pictures without missing strips, the threshold value of the similarity index is continuously updated, the quality defect of the tobacco packet missing strips is accurately judged, and an information basis is provided for missing strip evidences through picture marking.

Drawings

Fig. 1 is a flow chart of an intelligent detection method for tobacco bale strip missing of a carton sealing machine based on similarity analysis and feature increment learning;

FIG. 2 is an original image of the front and back of a tobacco rod inside a certain cigarette packet;

FIG. 3 is a gray scale image of the front and rear positions of the tobacco rod inside the cigarette packet after gray scale conversion;

FIG. 4 is a tobacco bale effective area gray level image after image segmentation;

FIG. 5 is a graph of kernel density corresponding to 16 similarity indicators for the anterior and posterior images;

FIG. 6 is a diagram illustrating an on-line recognition of a previous original image without a missing bar and a recognized annotation image;

FIG. 7 is a diagram illustrating an on-line recognition of a posterior original image without missing bars and a recognized annotation image;

FIG. 8 is a diagram illustrating an on-line recognition of a previous original image with a missing bar and a recognized annotation image;

FIG. 9 is a diagram illustrating an on-line identification of a post-original image with a missing bar and an identified annotation image.

Detailed Description

In order to better understand the technical scheme of the invention, the following description is further provided for the embodiment of the invention with the accompanying drawings. The implementation is an intelligent detection method for carton missing of a carton sealing machine cigarette packet, and the carton missing detection system of the carton sealing machine cigarette packet comprises a control PLC, front and rear cameras, front and rear light sources, a carton missing identification algorithm and the like. In the process that the front 25 cigarettes and the rear 25 cigarettes are pushed into the packing box by the push plate, when all the front 25 cigarettes enter the packing box, the front camera takes a picture; after the back 25 cigarettes are all pushed into the packing box, the lower part of the push plate is turned upwards in the process of returning the push plate, and at the moment, the back camera takes a picture. And (3) adopting image similarity analysis and feature increment learning technology to collect and judge whether the front and rear bitmap sheets of the tobacco strips in the tobacco bale have strip-lacking quality defects in real time. If the defective strip quality exists, sending a removing instruction to a defective strip removing control system; and if the strip-missing quality defect does not exist, sending a quality qualified signal to a quality tracing code pasting system. The implementation block diagram of the intelligent detection method for cigarette packet missing strip of the carton sealing machine based on similarity analysis and feature increment learning is shown in figure 1, and the method mainly comprises the following steps:

(1) image graying and active area selection

Based on the front and rear imaging devices added on the carton sealing machine, when the push plate pushes the front 25 cigarettes into the carton completely, a front picture OF (N) is obtained_of×M_ofX 3) when the back 25 cigarettes are all pushed into the carton push plate to return, a back bitmap OB (N) is obtained_ob×M_obX 3) of the front and rear pictures into a gray image GF (N)_of×M_of) And GB (N)_ob×M_ob). As the shooting angles of the front picture and the rear picture are relatively fixed aiming at a certain carton sealing machine, in order to quickly identify the effective area of the cigarette packet, the method of manually marking the effective area of the cigarette packet is adopted to respectively obtain the coordinates (X) of the upper left corners of the front picture and the rear picture_f、x_b) Upper left corner Y-axis coordinate (Y)_f、y_b) Width (w)_f、w_b) Height (h)_f、h_b) Waiting for 8 parameters to obtain a gray image PF (N) of the effective area of the tobacco bale_pf×M_pf) And PB (N)_pb×M_pb)。

In this example, the front and back original pictures OF the tobacco rod inside a certain cigarette packet are shown in fig. 2, and the matrices OF the original pictures are OF (652 × 738 × 3) and OB (652 × 738 × 3), respectively. The grayscale images after the grayscale conversion are represented by grayscale image matrices GF (652 × 738) and GB (652 × 738) in fig. 3. The effective area of the front gray level image of the artificial mark is as follows: upper left corner X-axis coordinate X_fIs 121, upper left corner Y axis coordinate Y_fIs 196 and has a width w_fIs 550 and the height h_fIs 290; the effective area of the rear gray level image is as follows: upper left corner X-axis coordinate X_bIs 131, upper left corner Y-axis coordinate Y_bIs 280 and width w_bIs 470 and the height h_bIs 245. Fig. 4 shows a gray scale image of the effective area of the cigarette packet, and gray scale image matrices PF (290 × 550) and PB (245 × 470) of the effective area.

(2) Similarity analysis based on blocking idea

Based on the gray level image of the effective area of the cigarette packet of the preceding picture, equally dividing the gray level image into 5 blocks according to the height of the image to obtain 5 lines of sub-images XPF_i(N_xpf×M_pf) (i-1, …,5), calculate each sub-image matrix XPF_iMean value of column vectors, obtaining vector XMPF of 5 rows of sub-images_i(1×M_pf) Calculating the similarity (rho) between every two adjacent rows by using the Pearson correlation coefficient_f12、ρ_f23、ρ_f34、ρ_f45) The calculation formula is as follows:

cov (XMPF)_i,XMPF_i+1) Representing two adjacent row vectors XMPF_iAnd XMPF_i+1The covariance of (a) of (b),

respectively represent vectors XMPF_i、XMPF_i+1Standard deviation of (2).

Equally dividing the gray image based on the front picture cigarette packet effective area into 5 blocks according to the width of the image to obtain 5 columns of sub-images YPF_i(N_pf×M_ypf) (i 1, …,5), each sub-image matrix YPF is calculated_iObtaining the vector YMPF of 5 columns of sub-images by averaging the row vectors_i(N_pfX 1), calculating the similarity (rho ') between every two adjacent columns by using Pearson correlation coefficient'_f12、ρ′_f23、ρ′_f34、ρ′_f45)。

Calculating the similarity (rho) between every two adjacent rows of the effective area gray level image of the bitmap tobacco lamina package by adopting the same method_b12、ρ_b23、ρ_b34、ρ_b45) And a similarity (ρ'_b12、ρ′_b23、ρ′_b34、ρ′_b45)。

In this example, the 5 lines of sub-images of the gray scale image of the effective area of the cigarette packet in the front picture are XPF_i(58X 550), by calculation XPF_iObtaining the vector of the sub-image of 5 rows by the mean value of the column vectors as XMPF_i(1 × 550), degree of similarity ρ between each two adjacent rows_f12、ρ_f23、ρ_f34、ρ_f450.9181, 0.9375, 0.9472, 0.8916, respectively. 5 columns of sub-images of the gray scale image of the effective area of the front picture cigarette packet are YPFs_i(290X 110) by calculating YPF_iThe vector of 5 columns of sub-images obtained by the mean value of the row vectors is YMPF_i(290 × 1), similarity ρ 'between each two adjacent columns'_f12、ρ′_f23、ρ′_f34、ρ′_f450.7799, 0.8350, 0.9184, 0.9218, respectively.

The 5 lines of subimages of the effective area gray level image of the back bitmap chip tobacco bale are XPB_i(49X 470) by calculating XPB_iObtaining the vector of 5 lines of sub-images as XMPB by the mean value of the column vectors_i(1 × 470), the degree of similarity ρ between each two adjacent rows_b12、ρ_b23、ρ_b34、ρ_b450.9825, 0.9322, 0.9472, 0.9602, respectively. 5 columns of sub-images of the effective area gray scale image of the back bitmap film tobacco package are YPB_i(245X 94), by calculating YPB_iThe vector of 5 columns of sub-images obtained by the mean value of the row vectors is YMPB_i(245 x 1), similarity ρ 'between each two adjacent columns'_b12、ρ′_b23、ρ′_b34、ρ′_b450.7457, 0.6217, 0.8400, 0.9247, respectively.

(3) Modeling picture threshold calculation based on kernel density estimation

A similarity row vector rho can be calculated according to a front bit picture and a rear bit picture of a cigarette packet₁In turn from rho_f12、ρ_f23、ρ_f34、ρ_f45、ρ′_f12、ρ′_f23、ρ′_f34、ρ′_f45、ρ_b12、ρ_b23、ρ_b34、ρ_b45、ρ′_b12、ρ′_b23、ρ′_b34、ρ′_b45And (3) element composition. Collecting N_oCalculating the similarity row vector rho of each cigarette packet_j(1×16)(j＝1,…,N_o) Form a similarity matrix

And aiming at each row of data of the matrix, determining the distribution characteristics of the data by adopting a kernel density estimation method because the distribution condition of the data is unknown, and calculating to obtain a kernel density curve corresponding to each row. The kernel density estimation calculation formula in column k is as follows:

where K (-) is a Gaussian kernel function, ρ_jkIs a similarity index of the jth row and kth column of the similarity matrix R, N_oFor modeling the number of pictures, h is the window width, when h is smaller, the nuclear density estimation curve can reflect more details but has poorer smoothness, and when h is larger, the nuclear density estimation curve is smoother but can cover some details.

The similarity data at the leftmost side (i.e., the minimum similarity) in the kernel density curve is used as a threshold corresponding to the position in the modeled image, and a threshold vector θ (1 × 16) is obtained.

In this example, a front bitmap of 130 groups of cigarette packets without strip missing is collectedCalculating similarity row vector rho of each tobacco packet by using the sheet and the rear bitmap sheet_j(1 × 16), a similarity matrix R (130 × 16) is constructed. And aiming at each row of data of the matrix, selecting a window width h as 3, calculating a kernel density curve of each row according to a kernel density estimation calculation formula, wherein 16 similarity indexes correspond to the kernel density curve and are shown in fig. 4.

The similarity data at the leftmost side (i.e., the minimum similarity) in the nuclear density curve is used as a threshold corresponding to the position in the modeled image, and a threshold vector θ (1 × 16) is obtained, and 16 elements of the threshold vector θ are 0.8477, 0.8397, 0.8139, 0.8246, 0.6315, 0.6740, 0.7602, 0.7621, 0.9572, 0.8831, 0.9122, 0.9096, 0.5443, 0.5367, 0.7551, and 0.8267.

(4) Tobacco bale picture on-line identification based on similarity

Re-collecting front picture OF OF cigarette packet_TAnd a back bit slice OB_TConvert it into a gray image GF_TAnd GB_T. Respectively obtaining gray front image PF of effective area of tobacco bale based on the coordinates of effective areas of front image and rear image which are recognized in advance_TAnd a back picture PB_T. PF (front gray level) image_TDividing lines according to the height of the image, and calculating the similarity rho between adjacent lines_tf12、ρ_tf23、ρ_tf34、ρ_tf45Then, the image is divided into columns according to the image width, and the similarity ρ 'between adjacent columns is calculated'_tf12、ρ′_tf23、ρ′_tf34、ρ′_tf45. And comparing the 8 similarity indexes obtained by calculation with the first 8 elements of the threshold vector theta one by one, if all the similarity indexes are greater than the corresponding elements of the threshold vector theta, identifying that the front image has no missing bars and carrying out corresponding marking on the image, and otherwise, identifying that the front image has missing bars and carrying out corresponding marking on the image. Identifying a grayscale post-image PB using the same method_TWhether the strip missing occurs.

In this example, the front images and the rear images of 1 group of cigarette packets without missing strips are collected again for online identification, and the original front image is shown in fig. 6 (a). Gray level image matrix PF of effective area obtained by gray level transformation and image segmentation_T1(290 × 550), PF is calculated_T18 similarity indexes ρ_tf12、ρ_tf23、ρ_tf34、ρ_tf45、ρ′_tf12、ρ′_tf23、ρ′_tf34、ρ′_tf45Respectively as follows: 0.9028, 0.8716, 0.9407, 0.8638, 0.8168, 0.8085, 0.8457, 0.8888. The 8 similarity indexes are all larger than the first 8 elements of the threshold vector theta (1 × 16), the previous picture is identified that the missing bar does not appear, and the picture is labeled correspondingly, as shown in fig. 6 (b).

The original posterior image is shown in fig. 7 (a). Obtaining a gray image matrix PB of an active area by gray conversion and image segmentation_T1(245X 470) to calculate PB_T18 similarity indexes ρ_tb12、ρ_tb23、ρ_tb34、ρ_tb45、ρ′_tb12、ρ′_tb23、ρ′_tb34、ρ′_tb45Respectively as follows: 0.9768, 0.9272, 0.9548, 0.9514, 0.7241, 0.6833, 0.8539, 0.9264. The 8 similarity indexes are all larger than the last 8 elements of the threshold vector theta (1 × 16), the fact that the back bitmap slice has no missing bars is identified, and corresponding labeling is performed on the picture, as shown in fig. 7 (b).

And then, the image of the front cigarette rod in 1 group of cigarette packets is collected again for online identification, and the original front image is shown in fig. 8 (a). Gray level image matrix PF of effective area obtained by gray level transformation and image segmentation_T2(290 × 550), PF is calculated_T28 similarity indexes ρ_tf12、ρ_tf23、ρ_tf34、ρ_tf45、ρ′_tf12、ρ′_tf23、ρ′_tf34、ρ′_tf45Respectively as follows: 0.4856, 0.9295, 0.8801, 0.9198, 0.8183, 0.8139, 0.8542, 0.6398. Among the 8 similarity indexes, ρ_tf12And ρ'_tf45And identifying the missing bar of the previous picture when the element is smaller than the element corresponding to the threshold vector theta (1 × 16), and labeling the picture correspondingly, as shown in fig. 8 (b).

And then, the image of the back cigarette strip in 1 group of cigarette packets is collected again for online identification, and the original back image is shown in fig. 9 (a). By grey scale transformation and mappingObtaining gray image matrix PB of effective area by image division_T2(245X 470) to calculate PB_T28 similarity indexes ρ_tb12、ρ_tb23、ρ_tb34、ρ_tb45、ρ′_tb12、ρ′_tb23、ρ′_tb34、ρ′_tb45Respectively as follows: 0.8437, 0.9165, 0.9431, 0.9601, 0.5909, 0.5157, 0.8406, 0.9079. Among the 8 similarity indexes, ρ_tb12And ρ'_tb23And identifying the elements smaller than the threshold vector theta (1 × 16), and labeling the picture correspondingly, as shown in fig. 9 (b).

(5) Feature incremental learning based threshold updating

For the tobacco bale image without the missing strips, after the identification by the technologist, the similarity row vector rho representing the tobacco bale image characteristics_TAdding the obtained similarity matrix into the existing similarity matrix R based on the updated similarity matrix

The kernel density curve corresponding to each column is recalculated. Taking the similarity data at the leftmost side (namely, the similarity is minimum) in the nuclear density curve as the latest threshold corresponding to the position, and obtaining an updated threshold vector theta_T(1×16)。

In this example, the front picture and the rear picture of 10 groups of cigarette packets without the occurrence of the missing carton are collected again for online identification, and the occurrence of the missing carton is identified. The similarity row vector rho of the 10 groups of non-missing tobacco bale image features is represented_T1,…,ρ_T10And adding the similarity matrix R to the existing similarity matrix R in sequence. Selecting the window width h as 3, recalculating the kernel density curve of each row by adopting kernel density estimation, and obtaining an updated threshold vector theta_T(1 × 16), 16 elements of which are 0.8466, 0.8390, 0.8116, 0.8222, 0.6297, 0.6771, 0.7608, 0.7625, 0.9568, 0.8831, 0.9128, 0.9104, 0.5408, 0.5375, 0.7550, 0.8273, respectively.

It should be noted that the above embodiments are merely representative examples of the present invention. Many variations of the invention are possible. Any simple modification, equivalent change and modification of the above embodiments according to the spirit of the present invention should be considered to be within the protection scope of the present invention.

Claims (1)

1. The intelligent detection method for cigarette packet missing of the carton sealing machine based on similarity analysis and feature increment learning is characterized by comprising the following steps of:

(1) image graying and active area selection

Based on a front imaging device and a rear imaging device which are additionally arranged on a carton sealing machine, when a push plate pushes all the front 25 cigarettes into a carton, a front picture OF is obtained, when a push plate pushes all the rear 25 cigarettes into the carton push plate to return, a rear picture OB is obtained, and color RGB images OF the front picture and the rear picture are converted into gray level images GF and GB;

(2) similarity analysis based on blocking idea

Based on the gray level image of the effective area of the cigarette packet of the preceding picture, equally dividing the gray level image into 5 blocks according to the height of the image to obtain 5 lines of sub-images XPF_i1, 5, calculate each sub-image matrix XPF_iMean value of column vectors, obtaining vector XMPF of 5 rows of sub-images_iCalculating the similarity rho between every two adjacent lines by using the Pearson correlation coefficient_f12、ρ_f23、ρ_f34、ρ_f45(ii) a Equally dividing the gray image based on the front picture cigarette packet effective area into 5 blocks according to the width of the image to obtain 5 columns of sub-images YPF_iI 1, …,5, calculating the respective sub-image matrix YPF_iObtaining the vector YMPF of 5 columns of sub-images by averaging the row vectors_iCalculating the similarity rho 'between every two adjacent columns by using Pearson correlation coefficients'_f12、ρ′_f23、ρ′_f34、ρ′_f45；

Calculating the effective area of the bitmap cigarette carton package by adopting the same methodSimilarity rho between every two adjacent lines of the domain gray image_b12、ρ_b23、ρ_b34、ρ_b45And a similarity ρ 'between each two adjacent columns'_b12、ρ′_b23、ρ′_b34、ρ′_b45；

(3) Modeling picture threshold calculation based on kernel density estimation

A similarity row vector rho can be calculated according to a front bit picture and a rear bit picture of a cigarette packet₁In turn from rho_f12、ρ_f23、ρ_f34、ρ_f45、ρ′_f12、ρ′_f23、ρ′_f34、ρ′_f45、ρ_b12、ρ_b23、ρ_b34、ρ_b45、ρ′_b12、ρ′_b23、ρ′_b34、ρ′_b45Elemental composition, collecting N_oCalculating the similarity row vector rho of each cigarette packet_j，j＝1,…,N_oForm a similarity matrix

Aiming at each row of data of the matrix, determining the distribution characteristics of the data by adopting a kernel density estimation method because the distribution condition of the data is unknown, and calculating to obtain a kernel density curve corresponding to each row; taking the similarity data at the leftmost side in the nuclear density curve as a threshold value of a corresponding position in a modeling image to obtain a threshold value vector theta;

(4) tobacco bale picture on-line identification based on similarity

Re-collecting front picture OF OF cigarette packet_TAnd a back bit slice OB_TConvert it into a gray image GF_TAnd GB_TRespectively obtaining gray front image PF of effective area of tobacco bale based on the coordinates of effective area of front image and rear image recognized in advance_TAnd a back picture PB_TThe gray front image PF_TDividing lines according to the height of the image, and calculating the similarity rho between adjacent lines_tf12、ρ_tf23、ρ_tf34、ρ_tf45According to the image widthDividing columns, and calculating the similarity rho 'between adjacent columns'_tf12、ρ′_tf23、ρ′_tf34、ρ′_tf45Comparing the calculated 8 similarity indexes with the first 8 elements of the threshold vector theta one by one, if all the similarity indexes are larger than the corresponding elements of the threshold vector theta, identifying that the front image has no missing bars and performing corresponding marking on the image, otherwise identifying that the front image has missing bars and performing corresponding marking on the image, and identifying the gray level back image PB by adopting the same method_TWhether the strip is missing or not;

(5) feature incremental learning based threshold updating

For the tobacco bale image without the missing strips, after the identification by the technologist, the similarity row vector rho representing the tobacco bale image characteristics_TAdding the obtained similarity matrix into the existing similarity matrix R based on the updated similarity matrix

Recalculating to obtain a nuclear density curve corresponding to each column, taking the leftmost similarity data in the nuclear density curve as the latest threshold of the corresponding position, and obtaining an updated threshold vector theta_T。

Images