Coenzyme Q10 soft capsule appearance defect detection method based on image recognition

Abstract

The invention discloses a coenzyme Q10 soft capsule appearance defect detection method based on image recognition, relates to the technical field of capsule detection, and solves the problem that the reliability and pertinence of a detection result are insufficient due to the fact that different types of defects such as size deviation and breakage wrinkles cannot be effectively recognized. According to the method, the standard image and the to-be-analyzed image set are determined from a large number of images, and then the image with the most significant difference between the to-be-analyzed features and the focus is selected for ellipse perimeter comparison; according to the mode, the analysis range can be efficiently narrowed, the area where wrinkles, damage or abnormal sizes possibly exist can be accurately positioned, and the pertinence and efficiency of defect analysis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of capsule detection, in particular to a coenzyme Q10 soft capsule appearance defect detection method based on image recognition. BACKGROUND

[0002] As a common health food and drug dosage form, the appearance quality of coenzyme Q10 soft capsules is directly related to product stability, content safety and consumer trust. In industrial production, soft capsules are prone to deformation, wrinkles, breakage, uneven size and other appearance defects due to factors such as formula (e.g. gelatin viscosity fluctuation), process (e.g. press machine mold precision deviation, drying temperature and humidity out of control), and conveying collision. Such defects not only affect product compliance, but also may cause content leakage, oxidation and deterioration, and even lead to medication safety risks. Therefore, appearance defect detection is a key link in the quality control of coenzyme Q10 soft capsule production. The mainstream soft capsule appearance detection method in the industry is still mainly based on "manual sampling + visual observation": the detection personnel compares the standard sample with the naked eye, combined with magnifying glass, caliper and other simple tools, to judge the shape, color and surface state of the soft capsule. However, this method has significant limitations: on the one hand, manual detection relies on subjective experience, and the recognition accuracy of minor defects (such as 0.1mm or smaller scratches, and local slight deformation) is low, and visual fatigue may lead to missed detection and misjudgment; on the other hand, manual detection is low in efficiency (it takes several hours to detect a single batch of samples), and it is difficult to meet the full detection needs of industrial continuous production, especially it cannot cover the full range of defect inspection of the ring surface of the soft capsule (such as the hidden deformation of the side and back of the capsule). Some enterprises have tried to introduce basic machine vision detection schemes, but existing technologies mainly focus on single-angle image analysis, and have not designed adaptive detection logic for the "ring three-dimensional structure" of soft capsules. For example, only through the front image to extract contour features, it cannot fully reflect the difference in the circumferential shape of the capsule. At the same time, existing schemes lack quantitative analysis of "contour feature dispersion", making it difficult to accurately identify defects with "local deformation but normal appearance in single angle" (such as one side of the capsule is swollen and the other side is flat), and have not established a correlation verification mechanism between "contour features and elliptical circumference", which cannot effectively distinguish different types of defects such as "size deviation" and "breakage and wrinkles", resulting in insufficient reliability and pertinence of the detection results. In summary, the current coenzyme Q10 soft capsule appearance detection technology faces the pain points of "low precision, poor efficiency, incomplete coverage, and weak defect differentiation ability", and there is an urgent need for a detection method that can achieve full-range, high-precision, automatic defect recognition and quantitative verification of defect types to meet the strict quality control needs of industrial production. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a coenzyme Q10 soft capsule appearance defect detection method based on image recognition, which solves the problem that different types of defects such as "size deviation" and "breakage and wrinkle" cannot be effectively distinguished, resulting in insufficient reliability and pertinence of the detection result.

[0004] To achieve the above object, the present application is implemented by the following technical scheme: a coenzyme Q10 soft capsule appearance defect detection method based on image recognition, comprising the following steps: Step one, place the soft capsule at the detection position, and use a high-definition camera to obtain the overall image of the soft capsule, and perform contour detection on the overall image at the set position, confirm the contour features associated with the overall image, and according to the confirmation result, identify whether the current detection process is abnormal, in particular: Extract the overall image at the set position and mark it as a to-be-processed image, confirm the RGB values associated with different pixel points in the to-be-processed image, and according to the preset weight, confirm the gray values associated with the corresponding pixel points in the to-be-processed image, and according to different gray values associated with different pixel points, convert the to-be-processed image into a gray image; Confirm the horizontal gradient and vertical gradient associated with a single pixel point in the gray image using the Sobel algorithm, and according to the confirmed horizontal gradient and vertical gradient, confirm the comprehensive gradient associated with the corresponding pixel point, which Compare the confirmed comprehensive gradient with a preset value Y1, if the comprehensive gradient > Y1, mark the corresponding pixel point as a gradient pixel point, otherwise, do not mark it; Confirm the gradient pixel points associated with the gray image in turn, and connect adjacent gradient pixel points to confirm the gray contour associated with the gray image, and then place the gray image in a two-dimensional coordinate system to confirm the two-dimensional coordinates associated with different points in the gray contour, and perform mean value processing on the confirmed several groups of two-dimensional coordinates to confirm the mean value coordinates, and mark the point of the mean value coordinates as the center point of the current gray image; Confirm the two groups of contour points closest to and farthest from the center point on the gray contour, mark the distance between the closest contour point and the center point as L1, and the distance between the farthest contour point and the center point as L2, and use L2 ÷ L1 = Bz to confirm the contour feature Bz of the corresponding gray image, and then confirm the contour features associated with different to-be-processed images, and perform variance processing on the confirmed several groups of contour features Bz to confirm the calibration variance, if the calibration variance > Y2, it means that the current detection process is abnormal, and a detection abnormality signal is generated directly; If the calibration variance ≤ Y2, it means that the current detection process is normal, and no further processing is needed, where Y2 is a preset value; Step two, when the detection process exists an exception, directly confirm the standard image from the overall image at the set position, and then confirm the image set to be analyzed from the acquired several groups of overall images according to the adjacent features of the standard image, and then select the feature to be analyzed according to the gradient features of the edge profiles of the adjacent images in the image set to be analyzed, and the specific way is: extract the overall image at the set position, and extract the profile feature Bz belonging to the profile feature Bz associated with the corresponding overall image k wherein k represents different set positions, and then compare the confirmed profile feature Bz k with the preset standard feature BJ: if |Bz k -BJ|≤0.02, then mark the corresponding overall image as a standard image, otherwise, do not mark it; The specific way of confirming the image set to be analyzed is: confirm the set position associated with the standard image, identify whether there is a standard image adjacent to the set position, if there is, then according to the direction of image acquisition, the previous group of images of the adjacent standard image is regarded as the initial image, and the next group of images is regarded as the final image, and a group of "initial image-final image" image column is confirmed; sort the several groups of overall images acquired in the detection process, confirm a group of image set, and according to the confirmed image column, remove the initial image, the final image and other images associated between the initial image and the final image from the image set to obtain the image set to be analyzed; The specific way of selecting the feature to be analyzed from the image set to be analyzed is: According to the confirmed image set to be analyzed, the same way of confirming the profile feature of the image to be processed is adopted to confirm the profile features of different single images in the image set to be analyzed, and the maximum value and the minimum value are confirmed from the different profile features belonging to different images, and the single image associated with the maximum value and the single image associated with the minimum value are both recorded as the feature to be analyzed; Step three, based on the different profile features associated with different features to be analyzed, confirm whether the ellipse circumference associated with two features to be analyzed is consistent, and according to the confirmation result, generate an association signal for display, and the specific way is: According to the different profile features associated with different features to be analyzed, confirm the nearest distance L1 and the farthest distance L2 associated with the corresponding profile feature, and adopt: confirm the eccentricity e associated with the corresponding feature to be analyzed, and 0 adopt: wherein t is the integral variable, and dt is the differential of the integral variable t; Confirm the ellipse perimeter associated with the two characteristics to be analyzed, identify whether the two sets of confirmed ellipse perimeters are consistent, if consistent, directly generate a size difference signal for display, if inconsistent, directly generate a capsule abnormal signal.

[0005] The application provides a coenzyme Q10 soft capsule appearance defect detection method based on image recognition. The application can accurately extract contour features by performing grayscale, gradient calculation and other processing on the multi-directional images of the soft capsule, and can sensitively identify whether the detection process is abnormal by combining variance analysis and other means, effectively avoids defect omission caused by single-angle detection or simple visual observation, greatly improves the accuracy of defect detection, and ensures that no potential appearance problem is missed. When the detection process is abnormal, the standard image and the image set to be analyzed can be quickly confirmed from a large number of images, and then the characteristics to be analyzed are selected, and the image with the most significant difference is focused on for ellipse perimeter comparison. This way can efficiently narrow the analysis range and accurately locate the area that may have wrinkles, damage or size abnormalities, improving the targeting and efficiency of defect analysis. The method strictly controls the appearance quality of coenzyme Q10 soft capsules, can timely find unqualified products, prevents soft capsules with appearance defects (such as damage leading to content pollution, size abnormalities affecting drug efficacy, etc.) from flowing into the market, effectively safeguards the safety of consumers using the medicine, and also helps to maintain the product quality reputation and brand image of the enterprise; The entire detection process relies on image recognition, algorithm calculation and other technologies to reduce subjective errors and labor costs of manual detection, realizes automation and intelligentization of detection, meets the needs of modern industrial production for efficient and accurate quality detection, and is conducive to improving the production and detection efficiency and intelligent level of enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0006] Figure 1 The figure is a schematic diagram of the method of the application. DETAILED DESCRIPTION

[0007] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0008] Please refer to Figure 1 The application provides a coenzyme Q10 soft capsule appearance defect detection method based on image recognition, which comprises the following steps: Step one, place the soft capsule in the detection position, and use a high-definition camera to obtain the overall image of the soft capsule (an annular acquisition process can effectively confirm the overall image of the soft capsule), and perform contour detection on the overall image at the set position, confirm the contour features associated with the overall image, and according to the confirmation result, identify whether the current detection process is abnormal, if abnormal, the subsequent analysis process needs to be performed, if normal, no processing is needed, and the detection process belongs to a sampling detection process; Wherein, the specific way to identify whether the current detection process is abnormal is: Extract the overall image at the set position and mark it as a to-be-processed image. Specifically, the set position is set by the operator in advance, which is generally the front, left side, right side and back of the capsule. When confirming the position, there are corresponding angle data. The data can be confirmed and marked according to the corresponding angle characteristics. The RGB values associated with different pixel points in the to-be-processed image are confirmed, and the gray values associated with the corresponding pixel points in the to-be-processed image are confirmed according to the preset weight. The to-be-processed image is converted into a gray image according to different gray values associated with different pixel points. Specifically, the RGB values associated with a single pixel point are R value, G value and B value, respectively. Different values are associated with different weights, and the sum of multiple weights is 1. The gray value = 0.299 * R + 0.587 * G + 0.114 * B. The way of gray processing of the image is common in the prior art, so it will not be described here. The horizontal gradient and vertical gradient associated with a single pixel point in the gray image are confirmed by using the Sobel algorithm (the Sobel algorithm can confirm the vertical gradient and vertical gradient associated with the pixel point according to the different gray values associated with the eight surrounding pixel points. The weight factor is between -2 and 2. The gradient associated with the pixel point is confirmed by convolution summation. The way of gradient data confirmation is common in the prior art, so it will not be described here), and the comprehensive gradient associated with the corresponding pixel point is confirmed according to the confirmed horizontal gradient and vertical gradient. The confirmed comprehensive gradient is compared with the preset value Y1. If the comprehensive gradient > Y1, the corresponding pixel point is marked as a gradient pixel point, otherwise, no marking is performed. The gradient pixel points associated with the gray image are confirmed in sequence, and the adjacent gradient pixel points are connected to confirm the gray contour associated with the gray image. Then the gray image is placed in a two-dimensional coordinate system to confirm the two-dimensional coordinates associated with different points in the gray contour. The confirmed several groups of two-dimensional coordinates are processed by mean value to confirm the mean value coordinates. The point where the mean value coordinates are located is marked as the center point of the current gray image. And on the gray scale profile to confirm the nearest and farthest two groups of profile points with the center point, record the distance between the nearest profile point and the center point as L1, record the distance between the farthest profile point and the center point as L2, adopt: L2 ÷ L1 = Bz to confirm the profile feature Bz of the corresponding gray scale image, and then confirm the profile features associated with different images to be processed, and the confirmed several groups of profile features Bz are subjected to variance processing to confirm the calibration variance, if the calibration variance > Y2, it represents that there is an abnormality in the current detection process, and a detection abnormality signal is directly generated, otherwise, it represents that the current detection process is normal and does not need to be processed, wherein Y2 is a preset value, and its specific value is determined by the operator according to experience; Specifically, one capsule is placed at a detection position, the images of eight specified positions around it are acquired, the image to be processed is confirmed, the confirmed image to be processed is subjected to gray scale processing to obtain the gray scale image associated with the image to be processed, the gradient points existing around the gray scale image are confirmed according to the gray scale values associated with different pixel points in the gray scale image, and the gradient points are taken as the edge profile of the corresponding gray scale image, the center point of the corresponding gray scale image is confirmed according to the associated edge profile, and the two groups of points closest to and farthest from the center point are locked from the edge profile, which are respectively recorded as the nearest point and the farthest point, the profile feature of the corresponding capsule is confirmed according to the distance features associated with the nearest point and the farthest point, the specific difference degree between the capsule profiles is confirmed by analyzing the dispersion degree between the profile feature values, if the difference degree is large, it represents that the shape difference around the corresponding capsule is large, which belongs to the state of not meeting the standard, under normal circumstances, the profile features of the images associated with the four faces of the soft capsule should be relatively consistent, and the profile features are relatively similar, so the difference between the profile features of the images in four different directions can be compared to confirm whether the corresponding detection process is abnormal; Step two, when the current detection process is abnormal, the standard image is directly confirmed from the overall image at the set position, the image set to be analyzed is confirmed from the acquired several groups of overall images according to the adjacent features of the standard image, and the image feature to be analyzed is selected according to the gradient features of the adjacent image edge profiles in the image set to be analyzed. Specifically, different set positions are associated with different overall images, different overall images have different profile features, and the standard image is confirmed from a plurality of different overall images according to a preset standard feature. The specific way of confirming the standard image is as follows: Extract the overall image at the set position, and extract the profile feature Bz associated with the corresponding overall image k Wherein k represents different set positions, and the confirmed profile feature Bz k Is compared with the preset standard feature BJ: if |Bz kIf BJ|≤0.02, the corresponding overall image is marked as a standard image, otherwise, no marking is performed. Specifically, the preset standard feature BJ is a proportional feature associated with a standard size, which is generally 3, that is, the ratio between the length and width of the corresponding capsule is 3; Specifically, the confirmation of the standard image set is as follows: The set position associated with the standard image is confirmed, and it is identified whether there is a standard image adjacent to the set position. If there is, the previous group of images of the adjacent standard image is taken as the initial image and the next group of images is taken as the final image (that is, a part of the area of an elliptical facade, which belongs to the normal state) according to the direction of image acquisition, and a group of "initial image-final image" image column is confirmed. The several groups of overall images obtained in the detection process are sorted, a group of image set is confirmed, and according to the confirmed image column, the initial image, the final image and other images associated with the initial image and the final image are removed from the image set to obtain the image set to be analyzed. Specifically, the set positions are eight directions or four directions, and there are corresponding overall images in the corresponding directions. If there is a standard image, but the set positions associated with the standard image are not adjacent, then the image column cannot be confirmed. If they are adjacent, for example, the overall images in the "east" and "south" directions are standard images, then the several groups of other images associated between the two directions also belong to the standard image and do not need to be confirmed subsequently. During the image acquisition process, not only the images at the set positions are acquired, but also the ring-shaped acquisition is performed. Step one only analyzes the images at the set positions, but does not analyze the overall images. Because the amount of images is too large during overall analysis, but it does not mean that during the image acquisition process, images at different positions are not acquired. Instead, they are all acquired, which can be understood as a continuous acquisition state. Specifically, the confirmation of the standard image set is as follows: According to the confirmed image set to be analyzed, the same confirmation method of the contour feature of the image to be processed is used to confirm the contour features of different single images in the image set to be analyzed, and the maximum value and the minimum value are confirmed from the different contour features confirmed to belong to different images. The single image associated with the maximum value and the single image associated with the minimum value are both recorded as the feature to be analyzed. Specifically, that is, each image has a corresponding contour feature, and in the corresponding image set, there is a minimum contour feature and a maximum contour feature. The minimum contour feature and the maximum contour feature are associated with different images, and the corresponding image is the corresponding feature to be analyzed, which will be analyzed and verified subsequently. Step three, based on different profile features associated with different features to be analyzed, confirming whether the circumferences of the ellipses associated with the two features to be analyzed are consistent, and generating an association signal according to the confirmation result for display; The specific way of confirming whether the circumferences of the ellipses associated with the two features to be analyzed are consistent is as follows: According to different profile features associated with different features to be analyzed, the nearest distance L1 and the farthest distance L2 associated with the corresponding profile features are confirmed, which are as follows: The eccentricity e associated with the corresponding feature to be analyzed is confirmed, and 0 < e < 1; The eccentricity e associated with the corresponding feature to be analyzed is confirmed, and 0 < e < 1; Where t is the integral variable (eccentric angle), which is preset by the operator, and the value range is [0, 0.5π], and dt is the differential of the integral variable t; The circumferences of the ellipses associated with the two features to be analyzed are confirmed, and whether the two sets of confirmed circumferences are consistent is identified. If they are consistent, a size difference signal is directly generated for display. If they are not consistent, a capsule abnormal signal is directly generated (generally, the presence of wrinkles, damage or other conditions will lead to inconsistent circumferences of different faces, so the corresponding signal is directly generated for display); Specifically, in the actual verification process, as long as the circumferences associated with the two faces are not damaged, the circumferences will be consistent regardless of how the two faces change. Therefore, by selecting the most obvious feature faces of the two difference features from the corresponding image set and comparing and verifying the two feature faces, the difference between the corresponding faces can be effectively confirmed according to the feature change process of the circumferences of the two faces, so as to evaluate whether the corresponding capsule is damaged or whether there is an abnormality in the size during the manufacturing process. Therefore, by actual comparison and verification, the corresponding abnormality can be effectively and quickly locked.

[0009] Some data in the above formula are dimensionless numerical calculations, and the contents not described in detail in the specification are all prior art known to those skilled in the art.

[0010] The above embodiments are only used to illustrate the technical method of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A method for detecting appearance defects in coenzyme Q10 soft capsules based on image recognition, characterized in that, Includes the following steps: Step 1: Place the soft capsule at the detection location and use a high-definition camera to acquire an overall image of the soft capsule. Perform contour detection on the overall image at the set location to confirm the contour features associated with the overall image. Based on the confirmation results, identify whether the detection process is abnormal. Step 2: When there is an abnormality in the current detection process, directly confirm the standard image from the overall image at the set location, and then confirm the image set to be analyzed from several sets of overall images obtained according to the adjacent features of the standard image. Then select the features to be analyzed according to the gradual change features of the edge contours of adjacent images in the image set to be analyzed. Step 3: Based on the different contour features associated with different features to be analyzed, confirm whether the circumferences of the ellipses associated with the two features to be analyzed are consistent. Based on the confirmation results, generate the associated signals for display.

2. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 1, characterized in that, In step one, the specific method for identifying whether the current detection process is abnormal is as follows: Extract the overall image at the set location and mark it as the image to be processed. Confirm the RGB values ​​associated with different pixels in the image to be processed, and confirm the gray values ​​associated with corresponding pixels in the image to be processed according to the preset weights. Then, convert the image to be processed into a grayscale image based on the different gray values ​​associated with different pixels. The Sobel algorithm is used to identify the horizontal and vertical gradients associated with individual pixels in a grayscale image. Based on the identified horizontal and vertical gradients, the overall gradient associated with the corresponding pixel is determined. The confirmed comprehensive gradient is compared with the preset value Y1. If the comprehensive gradient is greater than Y1, the corresponding pixel is recorded as the gradient pixel. Otherwise, no marking is performed. The gradient pixels associated with the grayscale image are confirmed sequentially, and adjacent gradient pixels are connected to confirm the grayscale contour associated with the grayscale image. Then, the grayscale image is placed in a set of two-dimensional coordinate systems to confirm the two-dimensional coordinates associated with different points within the grayscale contour. The confirmed sets of two-dimensional coordinates are averaged to confirm the mean coordinates. The point where the mean coordinates are located is recorded as the center point of the current grayscale image. Then, identify the two sets of contour points closest to and farthest from the center point on the grayscale contour. The distance between the closest contour point and the center point is denoted as L1, and the distance between the farthest contour point and the center point is denoted as L2. The contour feature Bz of the corresponding grayscale image is confirmed by using L2÷L1=Bz. Then, the contour features associated with different images to be processed are confirmed. The variance of the confirmed contour features Bz is processed to confirm the calibration variance. If the calibration variance > Y2, it means that there is an abnormality in the detection process, and a detection abnormality signal is directly generated.

3. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 2, characterized in that, If the calibration variance is less than or equal to Y2, it means that the detection process is normal and no processing is required. Y2 is a preset value.

4. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 1, characterized in that, In step two, the specific method for confirming the standard image is as follows: Extract the overall image at a specified location, and extract the contour feature Bz associated with the corresponding overall image. k Where k represents different set positions, and then the confirmed contour feature Bz k Compare with the preset standard feature BJ: If |Bz k If -BJ|≤0.02, the corresponding overall image is marked as a standard image; otherwise, no marking is performed.

5. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 4, characterized in that, In step two, the specific method for confirming the image set to be analyzed is as follows: Confirm the set position associated with the standard image, identify whether there are standard images adjacent to the set position, and if so, take the first set of images of the adjacent standard images as the initial image and the second set of images as the final image according to the image acquisition direction, and confirm a set of "initial image-final image" image columns. The several sets of overall images acquired during the detection process are sorted to confirm an image set. Based on the confirmed image series, the initial image, the final image, and other images associated with the initial and final images are removed from the image set to obtain the image set to be analyzed.

6. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 5, characterized in that, In step two, the specific method for selecting the features to be analyzed from the image set to be analyzed is as follows: Based on the confirmed image set to be analyzed, the contour features of different individual images in the image set to be analyzed are confirmed using the same confirmation method. From the confirmed different contour features belonging to different images, the maximum and minimum values ​​are identified, and the individual images associated with the maximum value and the individual images associated with the minimum value are recorded as features to be analyzed.

7. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 1, characterized in that, In step three, the specific method for confirming whether the perimeters of the two feature ellipses to be analyzed are consistent is as follows: Based on the different contour features associated with different features to be analyzed, the nearest distance L1 and the farthest distance L2 associated with the corresponding contour features are identified, and the following method is used: Confirm the eccentricity e associated with the feature to be analyzed, and 0 < e < 1; use: , where t is the integral variable, and dt is the derivative of the integral variable t; The perimeters of the ellipses associated with the two features to be analyzed are confirmed. It is then determined whether the two confirmed sets of ellipse perimeters are consistent. If they are consistent, a size difference signal is directly generated and displayed.

8. The method for detecting appearance defects of coenzyme Q10 soft capsules based on image recognition according to claim 7, characterized in that, If there is a discrepancy, an abnormal capsule signal will be generated directly.

Images