Product inspection device

Abstract

To provide an article inspection device with which it is possible to easily and accurately select and set an algorithm that suits the article to be inspected from a plurality of image processing algorithms.SOLUTION: Provided is an article inspection device for inspecting an article W to be inspected by applying a prescribed image processing algorithm Pgm to an inspection image Dpx in which the article W to be inspected is captured. The article inspection device comprises: an algorithm storage unit 44 for preliminarily storing a plurality of image processing algorithms; a learning unit 41 for evaluating about the stored image processing algorithms on the basis of performance information indicating inspection performance when the suitability of inspection to an inspection image is applied to the data of a plurality of acquired images obtained by imaging articles other than the article to be inspected, and calculating a plurality of evaluation values representing respective evaluation results about the plurality of image processing algorithms; and an image processing algorithm setting unit 45 for setting a prescribed image processing algorithm used for determining the quality state of the article W to be inspected, on the basis of the plurality of evaluation values calculated by the learning unit 41.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an article inspection apparatus, and more particularly to an article inspection apparatus that inspects the quality state of an inspected article by applying a predetermined image processing algorithm to an inspection image obtained by imaging the inspected article of a predetermined variety.

Background Art

[0002] Conventionally, in an article inspection apparatus, a predetermined image processing filter or a combination thereof corresponding to an inspection item is applied to imaging data of an inspected article so that a predetermined quality state of the inspected article can be inspected with high accuracy.

[0003] In such an article inspection apparatus, it is necessary to select and set an image processing filter or the like corresponding to the inspection item of a specific inspected article from a plurality of image processing filters and the like that are created in advance based on the characteristics of the inspected article and the characteristics of foreign matters to be detected and stored in a memory. Therefore, there is an apparatus in which the function of selecting and setting an image processing filter or the like necessary for inspection is automated so that such a selection and setting operation can be easily performed without being affected by the operator's experience.

[0004] As this type of article inspection apparatus, for example, a plurality of types of filter processing corresponding to characteristics are performed on input inspection image data based on detection data of the X-ray dose transmitted through an inspected article in which no foreign matter is mixed, and a plurality of X-ray image data are generated. A filter that generates X-ray image data such that the maximum pixel value is minimized among the data of the plurality of X-ray images in which the edge portion image of the article has an increased luminance is selected and set as the optimal X-ray image processing filter (see, for example, Patent Document 1).

[0005] Furthermore, in order to extract a desired image processing algorithm that approximates the foreign object detection characteristics capable of detecting the foreign object to be detected from a plurality of image processing algorithms stored in a storage means, some systems have made it easier to select and set the algorithm by displaying a ranking of image processing algorithms that can highlight foreign objects while reducing the influence of the item being inspected (see, for example, Patent Document 2). [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2004-28891 [Patent Document 2] Japanese Patent Publication No. 2012-137387 [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] However, while conventional item inspection devices, as described above, are capable of high-precision foreign object detection, they tend to relatively suppress the detection and display functions of the features of the inspected item in order to emphasize the foreign object. Therefore, it was not easy to accurately select and set the necessary image processing algorithm, that is, the combination of multiple necessary image processing filters, while taking into account differences in the condition of the item itself, such as its properties and form (and the corresponding differences in density, shape, brightness distribution, etc., of the inspection image), which affect the inspection results.

[0008] For example, for inexperienced users, selecting and configuring an image processing algorithm that accurately corresponds to the item being inspected was not easy, and the need for configuration using foreign object samples also made it difficult to accurately select and configure the image processing algorithm.

[0009] Therefore, it was necessary to, for example, have the system perform a selection operation input to roughly classify the condition of the item to be inspected based on typical shapes and sizes that can be visually observed, and then automatically select the image processing algorithm best suited to the selection result, or to confirm that the image processing algorithm is accurate enough that the detection waveform rises clearly when a sample of a foreign object to be detected is attached.

[0010] Furthermore, if a learning stage is included in which X-ray transmission images of the inspected item with foreign matter are learned and stored to obtain high inspection accuracy, there was a concern that if the learning data focused on the defective area becomes overfitted, the influence of the properties and morphology of the item itself, which affect the inspection results, may not be sufficiently learned, and the optimal image processing algorithm may be excluded from the options.

[0011] This invention has been made in view of the above-mentioned unresolved problems of the conventional approach, and aims to provide an article inspection device that can easily and accurately select and set an image processing algorithm suitable for an article to be inspected from among multiple image processing algorithms. [Means for solving the problem]

[0012] The article inspection apparatus according to the present invention, in order to achieve the above objective, is an article inspection apparatus that inspects the quality state of an article to be inspected by applying a predetermined image processing algorithm to an inspection image obtained by imaging an article of a predetermined type, and is characterized by comprising: an image processing algorithm storage unit that stores in advance a plurality of image processing algorithms including the predetermined image processing algorithm; an image processing algorithm evaluation unit that evaluates the suitability of the inspection for the inspection image of the plurality of image processing algorithms of the image processing algorithm storage unit based on performance information that represents the inspection performance when applied to data of a plurality of acquired images obtained by imaging other articles other than the article to be inspected, and calculates a plurality of evaluation values ​​that represent the evaluation results for each of the plurality of image processing algorithms; and an image processing algorithm setting unit that sets the predetermined image processing algorithm used to determine the quality state of the article to be inspected based on the plurality of evaluation values ​​calculated by the image processing algorithm evaluation unit.

[0013] Therefore, in the present invention, the image processing algorithm evaluation unit evaluates the suitability of the inspection to the inspection image data of the item under inspection based on performance information representing the inspection performance of the inspection applied to the data of multiple acquired images obtained by imaging items other than the item under inspection, and calculates multiple evaluation values ​​representing the evaluation results. Then, based on the multiple calculated evaluation values, the image processing algorithm setting unit sets a predetermined image processing algorithm to be used to determine the quality state of the item under inspection. As a result, the optimal image processing algorithm can be easily, quickly, and accurately selected and set from multiple image processing algorithms, resulting in an item inspection device that can quickly and accurately set an inspection algorithm for performing item inspection, including image processing and judgment processing.

[0014] The inspection image data of the inspected item referred to here refers to, for example, the inspection image data where the label is unknown when using a model trained with labeled training data to select an image processing algorithm. In this case, the performance information representing inspection performance refers to, for example, the performance information labeled to the acquired image data, and the evaluation value refers to, for example, an evaluation index equivalent to the accuracy rate for the correct label, which is a value that can be displayed as a score.

[0015] In a preferred embodiment of the present invention, (2) the image processing algorithm evaluation unit calculates the plurality of evaluation values ​​such that the higher the inspection performance represented by the performance information, the greater the evaluation value, and the image processing algorithm setting unit selects a subset of superior image processing algorithms from the plurality of image processing algorithms based on the numerical values ​​of the plurality of evaluation values ​​to set the predetermined image processing algorithm.

[0016] This configuration allows for the precise and easy selection and configuration of the image processing algorithm that offers superior inspection performance from among multiple image processing algorithms.

[0017] In a preferred embodiment of the present invention, (3) the plurality of evaluation values ​​are calculated such that the maximum value is obtained when the performance information represents inspection performance that is superior to a predetermined level, and the image processing algorithm setting unit sets the plurality of evaluation values ​​in order of superiority. Image processing algorithms A predetermined number of image processing algorithms may be displayed on a display device for selection.

[0018] In this case, it becomes possible to extract image processing algorithms that offer superior inspection performance at a predetermined level or higher from among multiple image processing algorithms, and then make a decision and confirm the selection of the necessary inspection performance.

[0019] In a preferred embodiment of the present invention, (4) the image processing algorithm evaluation unit may be configured to evaluate the suitability of each of the plurality of image processing algorithms for the inspection of the inspection images based on a learning model created in advance by learning using inspection images of a plurality of product groups of other articles and image processing algorithms suitable for the inspection images of the product groups as training data.

[0020] In this case, training is performed using training data that associates inspection images of multiple product groups, each representing multiple types of items to be inspected, with an image processing algorithm suited to each inspection image. A trained model is then created. As a result, the trained model can exhibit high-precision classification performance for inspection images of the items to be inspected, enabling the selection and setting of the appropriate image processing algorithm.

[0021] In a preferred embodiment of the present invention, (5) the image processing algorithm setting unit extracts at least one image processing algorithm candidate according to the evaluation value of the plurality of image processing algorithms based on the learning model in the image processing algorithm evaluation unit, and if there are multiple extracted image processing algorithm candidates, the system can be configured to display the extracted image processing algorithm candidates on a display in a selectable manner.

[0022] In this case, when an image processing algorithm with superior inspection performance is clearly identified among multiple image processing algorithms, it becomes possible to automatically select and set that algorithm, or to display multiple candidate image processing algorithms with a high probability of having comparable inspection performance, and wait for the user to input a decision or confirmation of the required inspection performance before making a selection.

[0023] In a preferred embodiment of the present invention, (6) the present invention may have a product group storage means that stores images of a product group to which the optimal image processing algorithm among the plurality of image processing algorithms is associated as training data, and a learning unit that generates the learning model based on the training data.

[0024] In this case, since images of product groups associated with the optimal image processing algorithm among a plurality of image processing algorithms can be used as teaching data, it becomes possible to accurately select and set the optimal image processing algorithm for the inspection image of the inspected article.

[0025] In a preferred embodiment of the present invention, (7) the learning model may be created using a deep learning classification method.

[0026] In this case, the learning model uses a deep learning classification method with inspection images of a plurality of types of product groups that image other articles than the inspected article and the image processing algorithms suitable for the respective inspection images as teaching data, so that accurate classification can be performed for the multi-dimensional features of the inspection images respectively, and based on those classification results, it is possible to more accurately select and set the necessary image processing algorithm that suitably corresponds to the inspection image.

[0027] In a preferred embodiment of the present invention, (8) an X-ray inspection image may be obtained by irradiating the inspected article with X-rays, and the quality state of the inspected article may be X-ray inspected by applying the predetermined image processing algorithm to the X-ray inspection image.

[0028] In this case, it becomes an X-ray inspection apparatus that can easily, quickly, and accurately select and set the optimal algorithm from a plurality of X-ray inspection image processing algorithms.

Advantages of the Invention

[0029] According to the present invention, it is possible to provide an article inspection apparatus that can easily and accurately select and set an image processing algorithm suitable for the inspected article from a plurality of image processing algorithms.

Brief Description of the Drawings

[0030] [Figure 1] It is a schematic configuration diagram of a main part of an article inspection apparatus using an X-ray inspection method according to an embodiment of the present invention. [Figure 2]This is an explanatory diagram illustrating an example of an output format when multiple candidate image processing algorithms suitably corresponding to an input inspection image are extracted by a learning model in an X-ray inspection type article inspection device according to one embodiment of the present invention. [Figure 3] This is a conceptual diagram illustrating the inspection image dataset used as training data for learning during the learning phase in an X-ray inspection type product inspection device according to one embodiment of the present invention. [Figure 4] This is a flowchart showing the learning procedure during the learning phase in an X-ray inspection type article inspection device according to one embodiment of the present invention. [Figure 5] This flowchart shows the procedure for extracting necessary image processing algorithms that suitably correspond to inspection images from a trained model in an X-ray inspection type product inspection device according to one embodiment of the present invention, and selecting and setting them as product type parameters. [Figure 6] This is an explanatory diagram of the operation screen used when extracting necessary image processing algorithms that suitably correspond to inspection images using a pre-trained model in an X-ray inspection type product inspection device according to one embodiment of the present invention, and selecting and setting them as product type parameters. [Figure 7] This schematic diagram of the main components of an X-ray inspection type article inspection apparatus according to another embodiment of the present invention illustrates a case in which an algorithm evaluation unit is provided in the image processing control unit. [Modes for carrying out the invention]

[0031] Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.

[0032] (One embodiment) Figures 1 to 6 show an article inspection device according to one embodiment of the present invention.

[0033] First, let me explain the structure.

[0034] As shown in Figure 1, the article inspection apparatus 1 of this embodiment comprises a transport unit 10, an inspection unit 20, and a control unit 30. The inspection unit 20 irradiates the article to be inspected W (article to be inspected) which is transported by the transport unit 10 on a conveyor belt with X-rays, and detects image data corresponding to the transmitted X-ray dose distribution. Based on the detected image data, the quality state of the article to be inspected W is then inspected. The quality state referred to here refers to whether the article to be inspected W meets the required quality and physical quantities as a product, such as the presence or absence of foreign matter, the presence or absence of missing parts, the pass / fail status of the shape, size, and storage state of the contents, and the distribution of density, thickness, volume, or mass.

[0035] The conveying section 10 is a conveyor that has a loop-shaped conveying belt 11 wrapped around a plurality of conveying rollers 12 and 13, and can sequentially convey the items to be inspected W in the rightward direction in Figure 1 by the upper section 11a of the conveying belt 11, and is supported by a housing (not shown).

[0036] The inspection unit 20 is an X-ray inspection unit comprising an X-ray generator 21 (X-ray source) that generates X-rays in a predetermined energy band that penetrate the item W to be inspected, which is transported by the transport unit 10, and an X-ray detector 23 positioned directly below the upper section 11a of the transport belt 11.

[0037] The X-ray generator 21 generates X-rays of wavelength and intensity corresponding to the tube current and tube voltage of a known X-ray tube 22, and is configured to irradiate the items to be inspected W within a predetermined inspection section Zx on the conveyor belt 11 with fan-beam-shaped X-rays directed in a main observation direction perpendicular to the item conveying direction of the conveyor unit 10, through an X-ray window of an enclosure (details not shown).

[0038] Although not shown in detail, the X-ray detector 23 is composed of an X-ray line sensor camera in which detection elements, for example, a scintillator (a phosphor) and a photodiode or charge-coupled element, are arranged in an array at a predetermined pitch in the width direction of the transport path of the transport unit 10, and X-ray detection is performed at a predetermined resolution. It is positioned at a predetermined location in the transport direction corresponding to the X-ray irradiation position from the X-ray generator 21.

[0039] In other words, the X-ray detector 23 detects the X-rays irradiated from the X-ray generator 21 and transmitted through the item W under inspection for each predetermined transmission region corresponding to the detection element, converts them into an electrical signal corresponding to the amount of X-ray transmission, and outputs an X-ray detection signal for generating an X-ray transmission image in which the direction of X-ray transmission is the observation direction. Here, the X-rays irradiated from the X-ray generator 21 or the X-rays detected by the X-ray detector 23 are assumed to be of a certain beam quality (energy, wavelength) specified according to the quality of the item W under inspection. However, by using multiple types of X-rays with different beam qualities, it may be possible to generate so-called dual-energy or multi-energy X-ray images.

[0040] The control unit 30 includes transport control means for controlling the transport speed and transport interval of the items to be inspected W by the transport belt 11 in the transport unit 10, and inspection control means for controlling the X-ray irradiation intensity and irradiation period in the inspection unit 20, and for controlling the X-ray detection period of the X-ray line sensor of the X-ray detector 23 and the detection period of each item to be inspected W according to the transport speed of the items to be inspected, but detailed illustrations are omitted.

[0041] The control unit 30 also includes: a detection image data acquisition unit 31 that sequentially acquires X-ray detection signals from the X-ray detector 23 at predetermined intervals to acquire and output X-ray transmission images of each item W to be inspected; an image processing unit 32 that acquires the detection image data output from the detection image data acquisition unit 31 and performs image analysis processing such as predetermined filtering (including preprocessing) to extract image features and feature measurement to determine the feature quantities of the extracted image features; an inspection and determination unit 33 that performs a determination process to determine whether or not the item W to be inspected is in a predetermined quality state based on the feature quantity data extracted and measured by the image processing unit 32, for example, the presence or absence of foreign matter, the presence or absence of missing parts, the shape, size or storage condition of the contents, etc.; an image processing control unit 34 that can update and change inspection algorithms, including the image processing algorithm used in the image processing unit 32 and the determination processing algorithm used in the inspection and determination unit 33; and a display operation unit 35 (display) such as a touch panel that can display and output the determination results from the inspection and determination unit 33 and input product registration and other requested operations to the image processing control unit 34.

[0042] The control unit 30 is configured to include, for example, a microcomputer having a CPU, ROM, RAM, and I / O interface (not shown), an auxiliary storage device that stores control programs for performing the functions of the image processing unit 32, inspection and determination unit 33, and image processing control unit 34 in a readable format in cooperation with the ROM, and a timer circuit, etc. The CPU executes predetermined arithmetic processing and the control program while exchanging data with the RAM, etc., according to the control program stored in the ROM, etc.

[0043] The detection image data acquisition unit 31 is an image input unit that, for example, performs A / D conversion on the X-ray detection signals from multiple detection elements of the X-ray detector 23, and at predetermined unit transport times corresponding to the size of the detection elements in the X-ray detector 23, it performs an operation (hereinafter referred to as line scanning) that writes the cumulative transmission amount data within that unit time to the image memory as digital data of density levels representing grayscales from 0 to 1023 for all n detection element regions (n ​​is an integer greater than 1, for example, 640 elements).

[0044] Furthermore, the detection image data acquisition unit 31 has a data processing program and working memory (not shown) that perform the function of generating X-ray image data Dpx corresponding to the dose distribution of X-rays transmitted through the item W, based on the detection data Lx of the line scanning image which is sequentially written to the image memory when the line scanning by the X-ray detector 23 is repeated a predetermined number of times corresponding to the inspection period of the item W to be inspected, and outputting it to the image processing unit 32 as imaging data of the item W to be inspected.

[0045] The image processing unit 32 has a predetermined image processing algorithm pre-set and stored, which combines image processing filters and the like, in order to perform a predetermined item inspection based on the imaging data Dpx (X-ray imaging image data) of the item W to be inspected acquired from the detection image data acquisition unit 31.

[0046] The image processing filter included in the image processing algorithm of the image processing unit 32 is a processing program for extracting image features (e.g., edges, lines, corners, regions, grayscale, texture) necessary for a predetermined item inspection based on the imaging data Dpx of the item W under inspection. If the image processing algorithm includes a filter for detecting foreign objects, it may have, for example, a feature extraction filter that performs edge detection processing to emphasize the contours of foreign objects in the item W under inspection, or a differential filter such as a Sobel filter, which emphasizes the edges of foreign objects by applying differential processing based on a predetermined calculation formula to the neighborhood region of the pixel of interest. Note that "image processing filter, etc." means that pre-processing such as grayscale correction and noise reduction of the imaging data from the detection image data acquisition unit 31 is included in order to improve the accuracy of the image feature detection processing.

[0047] Furthermore, the feature measurement of image features performed by the image processing unit 32 is a process that calculates the features necessary for the judgment process in the inspection judgment unit 33 by performing necessary preprocessing and image processing on the image data Dpx of the item to be inspected W acquired from the detection image data acquisition unit 31, and then calculating attributes related to grayscale features, color features, shape features, etc. (feature quantities that characterize edges, regions, distances, positions, shapes, etc.), calculating feature quantities that represent the spatial relationships between such features, or calculating texture features related to spatial frequency distribution and directional components.

[0048] The inspection and determination unit 33 detects feature shapes, foreign objects, etc., detected in the item W under inspection based on the feature quantities extracted and feature-measured by the image processing unit 32. It also compares the feature quantities of the detected object, such as area, contour length, and density sum, with predetermined judgment criteria values ​​to perform a determination process to determine whether or not local feature shapes, etc., corresponding to foreign objects or defective parts that satisfy the judgment conditions are included in the item W under inspection.

[0049] The item inspection device 1 inspects the item W of a predetermined type by applying a predetermined image processing algorithm Pgm, which combines multiple filtering processes, to the detection image Dpx obtained by imaging the item W of a predetermined type, and to the image data obtained by applying necessary pre-processing and image processing to the detection image Dpx, and to the image data.

[0050] On the other hand, the image processing control unit 34 is capable of data communication with a machine learning learning unit 41, which is configured as an external PC or the like. This learning unit 41 is equipped with a learning model 42 and a learning data storage unit 43 that stores a dataset of a large number of learning samples for training the learning model 42, each containing image data of the sample and a label indicating an image processing algorithm that is suitable for (preferably used) that image data. In this embodiment, the label of each image processing algorithm stored in the learning data storage unit 43 includes at least the number (identification number) of the image processing algorithm, but may also include the name of the algorithm, the name of the product to be inspected, and the inspection items, which indicate the content of the inspection performed by that algorithm.

[0051] The image processing control unit 34 further includes an algorithm storage unit 44 (image processing algorithm storage unit) in which a plurality of image processing algorithms Pgm, including a predetermined image processing algorithm, are pre-stored, and an algorithm setting unit 45 (image processing algorithm setting unit) capable of executing an update setting process to update the image processing algorithm set and stored in the image processing unit 32 to one of the image processing algorithms stored in the algorithm storage unit 44.

[0052] The learning unit 41 has the function of creating a learning model 42 that can associate the type of suitable image processing algorithm with one of several image processing algorithm classifications, each of which is associated with a suitable image processing algorithm, for input inspection image data of varieties that correspond to or approximate any of the image data of the many learning samples. This is achieved by performing learning processing using statistical pattern recognition methods and machine learning processing based on a dataset of numerous learning samples stored in the learning data storage unit 43.

[0053] The statistical pattern recognition method described herein represents the pattern to be classified using a specific high-dimensional feature vector corresponding to the number of classifications, and classifies it based on the position of that vector in the low-dimensional feature space. This method consists of a learning stage and a recognition stage. Furthermore, the input test image data described herein refers to test image data where the labels are unknown, when using a trained model 42 that has been trained on labeled training data related to the algorithm settings. In this case, the performance information representing the test performance is, for example, performance information about the algorithm labeled on the acquired image data, and the evaluation value is, for example, an evaluation index equivalent to the accuracy rate for the correct label, which is a value that can be displayed as a score.

[0054] In the learning phase of statistical pattern recognition performed in the learning unit 41, a dataset of numerous training samples, where the type of suitable image processing algorithm is known as the classification label information, is used to calculate the type parameter of the image processing algorithm to be classified, and parameters representing the distribution state of the image features of each training sample in the feature space, from this dataset.

[0055] Furthermore, in the recognition stage of statistical pattern recognition performed in the learning unit 41, feature selection is performed to extract low-dimensional feature vectors effective for classification based on the feature vectors obtained from the input image pattern. This feature selection is, for example, principal component analysis, which linearly projects feature points in a multidimensional space onto a low-dimensional subspace with high variance (an indicator of the variability of pixel values).

[0056] Next, the extracted low-dimensional feature vectors are classified to determine which class they belong to, based on the distribution in the feature space of each class, which has been learned in advance using a large number of training samples. This identifies the type of image processing algorithm that is suitable for the item W being inspected based on the input image pattern.

[0057] More specifically, the learning unit 41 creates a learning model 42 that can evaluate the suitability of the algorithm in the inspection image and output an evaluation value for each algorithm, based on the data of a sample inspection image linked to an image processing algorithm, and the feature data obtained from that inspection image data, such as statistical features such as image density variance, mean, and hentogram of imaging data of a group of products that are the same as or similar to the item under inspection W but are not the item under inspection W itself, as well as features such as density, contour length, and area for planar and three-dimensional shapes, using predetermined calculation formulas.

[0058] Multiple types of image processing algorithms that can be classified by the learning unit 41 are stored in the algorithm storage unit 44 as programs and setting parameters corresponding to each type of image processing algorithm. When the learning unit 41 identifies a suitable type of image processing algorithm based on the imaging data Dpx of the item W to be inspected acquired from the detection image data acquisition unit 31, the learning model 42 notifies the algorithm setting unit 45 that the suitable type of image processing algorithm is the required image processing algorithm. Upon receiving this notification of the required image processing algorithm, the algorithm setting unit 45 performs an update setting process to update the current image processing algorithm stored in the image processing unit 32 to the suitable type of image processing algorithm stored in the algorithm storage unit 44. The algorithm setting unit 45 may also perform the update setting process not only once, but when the learning unit 41 notifies the suitable type of image processing algorithm a predetermined number of times.

[0059] The multiple image processing algorithms stored in the algorithm memory unit 44 include, for example, image filters and other image processing programs that are suitable for 30 types of articles that may be registered as product types W to be inspected. More specifically, they are image processing algorithms suitable for any of the product groups such as chicken breast, uneven surfaces, flour, noodle products, and packaged chocolate.

[0060] The learning unit 41 primarily performs supervised learning using a dataset of numerous training samples where the type of suitable image processing algorithm is known as label information. However, it is also conceivable to perform semi-supervised learning, which involves labeling unlabeled training samples based on the classification results performed by the trained model 42 to increase the number of training samples. Alternatively, it is conceivable to perform data augmentation by applying transformations such as horizontal / vertical flipping, scaling, or rotation to the image data of the training samples, then re-checking the classification by the trained model 42, and adding a dataset of sample images for which the classification does not change to the training data.

[0061] The learning unit 41 also has the function of an image processing algorithm evaluation unit, which evaluates the suitability of each image processing algorithm for inspection of the input inspection image data from the detection image data acquisition unit 31 using a learning model 42, based on performance information representing the inspection performance when applied to the data of multiple acquired images obtained by imaging a large number of samples other than the item being inspected, and calculates multiple evaluation values ​​representing the evaluation results of each image processing algorithm as values ​​that can be displayed as scores.

[0062] The algorithm setting unit 45 then determines a predetermined image processing algorithm necessary for determining the quality state of the item W under inspection, based on multiple evaluation values ​​calculated by the learning unit 41 for multiple image processing algorithms, and executes an update setting process to update the current image processing algorithm stored in the image processing unit 32 to the image processing algorithm required for the appropriate type.

[0063] More specifically, the learning unit 41 performs training using inspection images of multiple product groups and image processing algorithms that are suitable for those product groups as training data. As a result, the learning unit 41 calculates the aforementioned scoreable evaluation value, which corresponds to the probability (accuracy rate) of falling into the class of the suitable image processing algorithm, and the larger the value, the better the inspection performance represented by the performance information, and the score can be displayed on the display operation unit 35.

[0064] As shown in Figures 2 and 3, the score display 51 on the display operation unit 35 by the learning unit 41 displays a predetermined number of labeled image processing algorithm datasets (3 in Figure 2) ranked in order of score on the algorithm selection operation screen 50 (details will be described later). Starting from the rank column 52 on the left, the columns to the right consist of an algorithm number column 53 that identifies the image processing algorithm, an algorithm detail description column 54 that specifies the target of the image processing algorithm's inspection, and a score display column 55. In addition, the score display in the score display column 55 in Figure 2 shows the probability, as a percentage of 1 or less, that the same suitable image processing algorithm exists in a specific classification class of sample images whose image features are classified in the same way as the data of the input inspection image from the detection image data acquisition unit 31.

[0065] In this illustrated example, regarding the suitability of each image processing algorithm for inspection of the input inspection image data from the detection image data acquisition unit 31, the image processing algorithm with algorithm number "1310," which has the most superior inspection performance as indicated in the performance information, is ranked as the number one candidate image processing algorithm. Its detailed description column 54 states that it is an image processing algorithm "for chicken breast meat," and its score is displayed as 0.92 (92%). The image processing algorithm "for chicken breast meat" referred to here is an image processing algorithm that includes an image processing filter suitable for detecting foreign objects such as bones remaining on or attached to chicken breast meat products.

[0066] In this case, the data from the input inspection image is shown to fit the class to which algorithm number "1310" is classified with a 92% probability. Furthermore, algorithm number "1312" for "2kg pack of chicken breast" is ranked second with a score of 0.05, and algorithm number "1311" for "chicken thigh meat" is ranked third with a score of 0.01. Therefore, it can be seen that the image processing algorithms ranked second and below are not suitable.

[0067] The algorithm setting unit 45 can update the image processing algorithm stored in the image processing unit 32 using the optimal image processing algorithm automatically selected based on the conformity evaluation results in the learning unit 41, without requiring the user to perform a selection operation input to roughly classify the condition of the item W under inspection based on typical shapes, sizes, etc., that can be visually observed, as in the conventional method.

[0068] Alternatively, the algorithm setting unit 45 can select a superior image processing algorithm from among multiple image processing algorithm candidates based on the score display values ​​(numerical values) of multiple evaluation values, for example, the "for chicken breast" image processing algorithm shown in Figures 2 and 3, via input from the display operation unit 35, and update the predetermined image processing algorithm selected in accordance with the operation input when the update button 56 in the figures is operated. The cancel button 57 in both figures is for canceling an operation if the selection operation of one of the multiple image processing algorithm candidates is incorrect.

[0069] The algorithm selection operation screen 50 (see Figure 6) for selecting image processing algorithm candidates during variety registration, as shown in Figure 3, is displayed on the touch panel screen (display) of the display operation unit 35. Below the common information display unit 61 are an image display area 62 of a predetermined screen size and an inspection information display unit 63, and below these are an operation input unit 64 having multiple operation buttons. The inspection information display unit 63 includes a variety number display unit 63a, an inspection result display unit 63b, and a selection content display unit 63c that displays the inspection content, image processing algorithm, or operation target and operation buttons during setting operations. Below the image display area 62 are the operation input unit 64, which includes a menu button 64a, a display switching button 64b, an operation confirmation button 64c, and a setting / adjustment button 64d. Below the inspection information display unit 63 are the operation input unit 64, which includes a stop button 64e and a start button 64f for operating the item inspection device 1.

[0070] The learning unit 41 may calculate a score, which is an evaluation value of inspection performance, such that it takes its maximum value when the performance information regarding inspection performance represents an inspection performance that is superior to a predetermined level. The algorithm setting unit 45 displays a predetermined number of image processing algorithms, for example three, from among several, in order of superiority of evaluation value, on the algorithm selection operation screen 50 (see Figure 6) of the display operation unit 35 for selection.

[0071] In other words, the algorithm setting unit 45 extracts at least one image processing algorithm candidate based on the evaluation value (score) of multiple image processing algorithms based on the learning model 42 in the learning unit 41. If there are multiple extracted image processing algorithm candidates, the unit can display these extracted image processing algorithm candidates in a selectable list on the display unit 50.

[0072] Furthermore, the learning unit 41, which is an image processing algorithm evaluation unit, evaluates the suitability of multiple image processing algorithms (algorithm numbers 1310, 1311, and 1312) for the input inspection image Dpx, based on a learning model 42 that was created in advance by supervised learning using inspection images of multiple types of products other than the item to be inspected W, and image processing algorithms that are suitable for the inspection images of those product groups, as described above.

[0073] The image processing control unit 34 has a learning data storage unit 43 (product group storage means) that stores image data of product groups to which the optimal image processing algorithm from among multiple image processing algorithms is linked as learning data, and a learning unit 41 that generates a learning model 42 based on the learning data stored in the learning data storage unit 43 is provided outside or inside the main body of the product inspection device 1, and is capable of data communication.

[0074] This learning model 42 is described as performing learning processing or machine learning processing using statistical pattern recognition methods, but it may also be created using deep learning classification methods using convolutional neural networks, or support vector machines (SVMs) that perform two-class classification, such as nonlinear SVMs. Furthermore, the input for learning may not be the X-ray image itself, but rather numerical data such as the mean, variance, and maximum values ​​of pixels, or product size. In addition, when using images as learning input, not only X-ray transmission images, but also difference images using transmission images of different energy bands, images from imaging methods of different optical systems such as visible light NIR, and images after filtering of the captured image may be used. Furthermore, images created by combining multiple types of images (the above-mentioned images), such as assigning an arbitrary image to each channel of a color image like an RGB image (see, for example, Japanese Patent Publication No. 2023-114827 and Japanese Patent Publication No. 2023-114828), may also be used.

[0075] In either case, the learning model 42 is typically created using AI technology, specifically machine learning (deep learning or other machine learning), based on learning data of individual items from product groups other than the item W being inspected (a collection of inspection images for which suitable image processing algorithms are already known). The learning data storage unit 43 stores the learning data data of sample items, and the trained learning model 42 allows the system to identify image processing algorithms by obtaining algorithm numbers and names, starting with those with the highest evaluation values. When registering the product type of the item W being inspected, the learning model 42 can be used to extract the optimal image processing algorithms in a ranked format based on the new input inspection images that will be inspected, and numerical data of the evaluation value (probability, etc.) can also be displayed simultaneously.

[0076] In this embodiment, the article inspection apparatus 1 irradiates the article W being transported with X-rays using the inspection unit 20 to acquire data of an input inspection image, which is an X-ray inspection image. The image processing unit 32 then applies a predetermined image processing algorithm to the data of the input inspection image to perform an X-ray inspection of the quality state of the article W being inspected.

[0077] On the other hand, when the set product type of the product to be inspected W is switched or newly registered in the display operation unit 35, the control unit 30 of the product inspection device 1 takes in the input inspection image data from the detection image data acquisition unit 31 of the inspection unit 20 into the learning unit 41 and algorithm setting unit 45 of the image processing control unit 34, and extracts and displays at least one, for example, multiple, suitable image processing algorithm candidates that are estimated by the learning model 42 to be highly likely to fit the input inspection image data. When the optimal image processing algorithm that fits the product to be inspected W of the set product type is automatically selected or selected by the selection operation, the control unit 30 causes the algorithm setting unit 45 to update the image processing algorithm used in the image processing unit 32 to the image processing algorithm that is optimal for the switched product type.

[0078] Next, I will explain the mechanism of action.

[0079] In this embodiment configured as described above, during the learning phase of the learning unit 41, as shown in Figure 4, first, a dataset of image data for each learning sample and a label indicating an image processing algorithm suitable for that image data is acquired for a large number of learning samples to train the learning model 42, and stored in the learning data storage unit 43 (step S11).

[0080] Next, the learning unit 41 uses a dataset of numerous learning samples stored in the learning data storage unit 43 to perform learning processing and machine learning processing using the aforementioned statistical pattern recognition method, thereby creating a learning model 42 that can associate the data of the input inspection image of the item W under inspection with one of several class classifications corresponding to the type of image processing algorithm that is suitable (step S12).

[0081] Next, as shown in Figure 5, the display operation unit 35 inputs a predetermined variety registration setting to the control unit 30, and when the item to be inspected W for variety registration is test transported to the item inspection device 1, the item to be inspected W is irradiated with X-rays by the inspection unit 20 while being transported, and the X-rays that have passed through the item to be inspected W are detected by the X-ray detector 23. The input inspection image data acquired by the detection image data acquisition unit 31 is then taken into the learning unit 41 and the algorithm setting unit 45 of the image processing control unit 34, respectively (step S21).

[0082] Next, the learning model 42 extracts at least one suitable image processing algorithm candidate that is estimated to fit the input inspection image data with a high probability (step S22), and then it is determined whether or not there are multiple extracted image processing algorithm candidates (step S23).

[0083] If there are multiple candidate image processing algorithms (if the answer is YES in step S23), the multiple candidate image processing algorithms are displayed in a ranking (step S24). In this case, the ranking may be displayed only if there are multiple candidate image processing algorithms with a score equal to or greater than a predetermined value.

[0084] Next, the system waits for the user to refer to the multiple image processing algorithm candidates displayed in a ranking as shown in Figure 2 and select one of the image processing algorithm candidates (if the user selects NO in step S25). When an image processing algorithm candidate deemed optimal for the item W to be inspected is selected (if the user selects YES in step S25), or when there are no multiple image processing algorithm candidates (if the user selects NO in step S23), the selected or extracted image processing algorithm is then parameterized as part of an inspection algorithm suitable for the item W to be registered (step S26), and the process ends.

[0085] As described above, in the article inspection apparatus 1 of this embodiment, the learning unit 41, which is an image processing algorithm evaluation unit, evaluates the suitability of the inspection to the inspection image data of the article to be inspected W, based on performance information representing the inspection performance of the inspection applied to the data of multiple acquired images obtained by imaging a group of products other than the article to be inspected W. Multiple evaluation values ​​(scores) representing the evaluation results for multiple image processing algorithms are calculated. Then, based on the multiple calculated evaluation values, the algorithm setting unit 45 sets a predetermined image processing algorithm to be used to determine the quality state of the article to be inspected W as the required image processing algorithm. Therefore, the optimal image processing algorithm can be easily, quickly, and accurately selected from multiple image processing algorithms.

[0086] Furthermore, in this embodiment, the learning unit 41 calculates an evaluation score such that the higher the inspection performance, the greater the score. The algorithm setting unit 45 selects an image processing algorithm with superior inspection performance based on the values ​​of multiple scores and updates the image processing algorithm. This makes it possible to accurately and easily select and set an image processing algorithm with superior inspection performance from among multiple image processing algorithms.

[0087] Furthermore, in this embodiment, the evaluation value is calculated to be the maximum value when the performance information represents inspection performance superior to a predetermined level or higher. The algorithm setting unit 45 displays a predetermined number, for example, three image processing algorithms, in order of superiority of the evaluation value, so that the user can select and confirm the necessary inspection performance after extracting the image processing algorithms that provide inspection performance superior to a predetermined level or higher from among the multiple image processing algorithms.

[0088] In addition, in this embodiment, a learning model 42 is created by training using training data that associates inspection images of multiple types of products W, which are the items to be inspected, with image processing algorithms that are suitable for each inspection image. As a result, the learning model 42 can exhibit highly accurate classification performance for inspection images of the items to be inspected W, and it becomes possible to select and set the appropriate image processing algorithm.

[0089] Furthermore, in this embodiment, if there are multiple extracted image processing algorithm candidates, the algorithm setting unit 45 displays these image processing algorithm candidates on the algorithm selection operation screen 50 of the display operation unit 35 (display). This allows for accurate selection of the image processing algorithm with superior inspection performance among multiple image processing algorithms, and also allows for selection and setting when multiple image processing algorithm candidates with a high probability of having comparable inspection performance are displayed, after waiting for the necessary selection judgment and confirmation operation input.

[0090] Furthermore, in this embodiment, images of a product group to which the optimal image processing algorithm from among multiple image processing algorithms is linked can be used as training data, making it possible to accurately select and set the optimal image processing algorithm for the inspection image of the item to be inspected W.

[0091] Furthermore, in this embodiment, when the learning model is created using a deep learning classification method, the learning model can accurately classify multidimensional features of the inspection image by using a deep learning classification method that uses inspection images of multiple types of products, including images of items other than the item being inspected W, and image processing algorithms suitable for each inspection image as training data. Therefore, based on these classification results, it is possible to more accurately select and set the necessary image processing algorithm that is suitable for the inspection image.

[0092] In this embodiment, the inspection device also irradiates the item W to be inspected with X-rays to acquire an X-ray inspection image, and applies a predetermined image processing algorithm to the X-ray inspection image to inspect the quality state of the item W. This makes it possible to easily, quickly, and accurately select and set the optimal image processing algorithm from among multiple X-ray inspection image processing algorithms.

[0093] As described above, according to this embodiment, it is possible to provide an item inspection device 1 that can easily select and set an appropriate image processing algorithm corresponding to the item W to be inspected from a plurality of image processing algorithms for item inspection.

[0094] (Other embodiments) Figure 7 shows an article inspection apparatus 1A according to another embodiment of the present invention.

[0095] As shown in the figure, the article inspection device 1A of this embodiment differs from the article inspection device 1 shown in Figure 1 in that it is further provided with an algorithm evaluation and determination unit 46 in the image processing control unit 34, but the other configurations are the same as those of the article inspection device 1 of the first embodiment. Therefore, in Figure 7, the same components as those of the first embodiment shown in Figure 1 are indicated by the same reference numerals as in Figure 1. The differences from the first embodiment will be explained below.

[0096] In the article inspection apparatus 1A of this embodiment, the image processing control unit 34 is provided with an algorithm evaluation and determination unit 46 that evaluates the image processing algorithm selected by the user and determines whether the selection is appropriate. This algorithm evaluation and determination unit 46 is activated when a predetermined product registration setting input is made to the control unit 30 by the display operation unit 35, and the article W to be inspected for product registration is test transported, and a candidate image processing algorithm that is estimated to be highly suitable for the data of the input inspection image acquired by the detection image data acquisition unit 31 is extracted by the learning model 42 and selected by the user.

[0097] During this operation, the algorithm evaluation and determination unit 46 requests, for example, on the display screen to test transport an item W to be inspected with a predetermined number of test pieces, and for each test transported item, it has the function of performing image processing in the image processing unit 32 and determination processing in the inspection and determination unit 33 based on the selected image processing algorithm, and further evaluates whether the selected image processing algorithm is appropriate by displaying the results on the inspection result display unit 63b. In addition, while the algorithm evaluation and determination unit 46 is operating, the algorithm setting unit 45 does not finalize the update setting of the selected image processing algorithm, but enters an operating state in which the image processing unit 32 can be restored to its state before the update.

[0098] The user can refer to the evaluation and judgment results from the algorithm evaluation and judgment unit 46 on the screen and, if the evaluation and judgment results are satisfactory, operate the update button 56. In addition, if the evaluation values ​​of the two image processing algorithms displayed in the ranking do not differ significantly, the user can operate the cancel button 57 to re-select another image processing algorithm whose evaluation value is second only to the previously selected image processing algorithm.

[0099] Alternatively, it is possible to perform additional setting operations, such as fine-tuning some of the parameters when using the image processing algorithm selected in the algorithm setting unit 45 by operating the setting / adjustment button 64d, and then confirming the updated setting state by operating the update button 56.

[0100] In this embodiment as well, the algorithm evaluation and determination unit 46, which can function as an image processing algorithm evaluation unit, evaluates the suitability of the inspection to the inspection image data of the item under inspection W based on performance information representing the inspection performance of the inspection applied to the data of multiple acquired images obtained by imaging a group of products other than the item under inspection W. Multiple evaluation values ​​(scores) representing the evaluation results for multiple image processing algorithms are calculated. Based on the multiple calculated evaluation values, the algorithm setting unit 45 sets a predetermined image processing algorithm to be used to determine the quality state of the item under inspection W as the required image processing algorithm. Therefore, this embodiment as well can achieve the same effects as the embodiment described above.

[0101] Furthermore, in this embodiment, when an image processing algorithm suitable for the input inspection image data of a new type of item W to be inspected is selected, the selected image processing algorithm can be evaluated by the algorithm evaluation and determination unit 46 using a predetermined number of test items prior to the update setting in the algorithm setting unit 45, thereby improving its suitability.

[0102] In the above-described embodiment, the article inspection device 1 was an X-ray inspection device, but the present invention is applicable to any other article inspection device that uses imaging data to perform article inspection. Furthermore, as an example of an image processing algorithm, an image processing algorithm suitable for foreign object detection was given as an example, but as mentioned above, it may also be used to inspect the quality and physical quantities required of the article W to be inspected, such as whether there are any missing parts, whether the shape, size, and storage condition of the contents are acceptable, and the distribution of density, thickness, volume, or mass.Of course, the selection of the image processing algorithm may also be performed in units of inspection algorithms, including the image processing algorithm used in the image processing unit 32 and the judgment processing algorithm used in the inspection judgment unit 33.

[0103] As described above, the present invention provides an article inspection device that can easily and accurately select and set an image processing algorithm suitable for an article to be inspected from among a plurality of image processing algorithms, and is useful for article inspection devices in general that inspect the quality state of an article to be inspected by applying a predetermined image processing algorithm to an inspection image obtained by imaging an article of a predetermined type to be inspected. [Explanation of symbols]

[0104] 1. 1A Item inspection device 10 Conveying section 11. Conveyor belt 11a Upper running section 12, 13 Conveyor rollers 20. Inspection Department (X-ray Inspection Department) 21 X-ray generator 22 X-ray tube 23 X-ray detector 30 Control Unit 31. Detection Image Data Acquisition Unit (Image Input Unit) 32 Image Processing Unit 33. Inspection and Judgment Unit 34 Image Processing Control Unit 35 Display operation unit (display unit) 41. Learning Unit (Image Processing Algorithm Evaluation Unit) 42 Learning Models 43. Learning Data Storage Unit 44. Algorithm Memory Unit (Image Processing Algorithm Memory Unit) 45. Algorithm Setting Section (Image Processing Algorithm Setting Section) 46. ​​Algorithm Evaluation and Judgment Unit (Image Processing Algorithm Evaluation Unit) 50. Algorithm selection operation screen (touch panel, display) 51 Score Display 52 Ranking column 53 Algorithm number field 54 Detailed description section 55 Score display area 56 Update button 57 Cancel button 61 Common information display section 62 Image display area 63 Inspection Information Display Unit 63a Product number display section 63b Inspection Result Display Unit 63c Selection Display Section 64 Operation Input Section 64a Menu button 64b Display Switching Button 64c operation check button 64d Settings / Adjustment Buttons 64e Stop button 64f Start button Dpx imaging data (examination images, captured image data, X-ray image data) Lx line scan image detection data W: Items under inspection Zx Inspection Area

Claims

1. An article inspection device that inspects the quality state of an article by applying a predetermined image processing algorithm (Pgm) to an inspection image (Dpx) obtained by imaging an article (W) of a predetermined type, An image processing algorithm storage unit (44) that stores in advance a plurality of image processing algorithms, including the predetermined image processing algorithm, Image processing algorithm evaluation unit (41; 46) evaluates each of the plurality of image processing algorithms in the image processing algorithm storage unit based on performance information representing the inspection performance when the suitability of the inspection for the inspection image is applied to the data of a plurality of acquired images obtained by imaging other items other than the item to be inspected, and calculates a plurality of evaluation values ​​representing the evaluation results for each of the plurality of image processing algorithms. An article inspection apparatus comprising: an image processing algorithm setting unit (45) that sets a predetermined image processing algorithm used for determining the quality state of the article to be inspected, based on the plurality of evaluation values ​​calculated by the image processing algorithm evaluation unit.

2. The image processing algorithm evaluation unit calculates the multiple evaluation values ​​such that the higher the inspection performance represented by the performance information, the greater the value. The article inspection apparatus according to claim 1, characterized in that the image processing algorithm setting unit selects a superior portion of the multiple image processing algorithms based on the numerical values ​​of the multiple evaluation values ​​and sets the predetermined image processing algorithm.

3. The aforementioned multiple evaluation values ​​are calculated such that they reach their maximum value when the performance information represents inspection performance superior to a predetermined level or higher. The article inspection apparatus according to claim 1 or 2, characterized in that the image processing algorithm setting unit displays a predetermined number of image processing algorithms from among the plurality of image processing algorithms on a display in order of priority of the plurality of evaluation values.

4. The aforementioned image processing algorithm evaluation unit is: The article inspection apparatus according to claim 1 or 2, characterized in that it evaluates the suitability of each of the multiple image processing algorithms for the inspection of the inspection images based on a learning model created in advance by learning using inspection images of multiple types of products, which are other articles, and image processing algorithms that are suitable for the inspection images of the product groups, as training data.

5. The article inspection apparatus according to claim 4, characterized in that the image processing algorithm setting unit extracts at least one image processing algorithm candidate according to the evaluation value of the plurality of image processing algorithms based on the learning model in the image processing algorithm evaluation unit, and if there are multiple extracted image processing algorithm candidates, the extracted image processing algorithm candidates are displayed on a display in a selectable manner.

6. The article inspection apparatus according to claim 4, further comprising a product group storage means that stores images of a product group associated with the optimal image processing algorithm from among the plurality of image processing algorithms as training data, and a learning unit that generates the learning model based on the training data.

7. The article inspection apparatus according to claim 4, characterized in that the learning model is created using a deep learning classification method.

8. The article inspection apparatus according to claim 1, characterized in that it irradiates the article to be inspected with X-rays to acquire an X-ray inspection image, and applies the predetermined image processing algorithm to the X-ray inspection image to perform an X-ray inspection of the quality state of the article to be inspected.

Images