Article inspection device

Abstract

The invention provides an article inspection device which can improve the accuracy of learning or performance verification even if an inspection image used in learning or performance verification for setting inspection conditions is not time-consuming and labor-consuming to acquire. An article inspection device is provided with an inspection unit (20) for inspecting the quality state of an article (P) using an inspection image (Dpx) obtained by capturing an image of the article (P) being conveyed, and the inspection unit (20) sets inspection conditions for the quality state of the article (P) on the basis of a virtual image (Dpp) of the inspection image generated by a generative expression AI (51). The inspection unit (20) can inspect the quality state of the article (P) by applying a learned model (35) which is created as an inspection condition for learning the virtual image (Dpp).

Description

Technical Field

[0001] This invention relates to an article inspection device, and more particularly to an article inspection device that uses an inspection image obtained by photographing the article to be inspected and a learning model or inspection algorithm to inspect the quality status of the article. Background Technology

[0002] Recently, an item inspection device has been known that uses a machine learning-completed model, also known as an AI (artificial intelligence) model, or an inspection algorithm to inspect the quality status of an item by applying the inspection image or sensor signal containing feature quantities corresponding to the quality status of the item, and displays the image containing the inspection result information as an inspection image.

[0003] As such an article inspection device, for example, it includes: an information storage unit that stores multiple image processing algorithms; a setting operation unit that selects a foreign object detection characteristic that represents the ability to detect foreign objects in the inspected article; a control unit that extracts at least one image processing algorithm from the information storage unit that is similar to the selected foreign object detection characteristic; and a display unit that displays the extracted image processing algorithm in a prescribed format, making it easy to execute the setting operation of the optimal image processing algorithm (for example, see Patent Document 1).

[0004] Furthermore, in order to improve the accuracy of item inspection, there is a device that: captures multiple images from different input systems under specified shooting conditions corresponding to each input system, acquires image data as a group of multiple images of the item to be inspected, and stores it in an image storage unit; on the other hand, it creates a machine learning model for inspection judgment using learning image data acquired under the same shooting conditions as the image data of the item to be inspected stored in the image storage unit, processes the image data of the item to be inspected acquired during actual inspection by each pixel using this model, calculates the degree of quality defect, and compares the degree of quality defect with a preset threshold to determine the quality status of the item to be inspected (for example, see Patent Document 2).

[0005] Furthermore, there is an apparatus comprising an X-ray generator, an X-ray detector, and a determination unit for determining the quality status of an inspected object. The apparatus includes: an X-ray image storage unit for storing image data from the X-ray detector; and an image production unit for virtually generating a virtual image model that virtually generates X-ray image data of other energy bands based on learning results of X-ray image data of multiple energy bands related to the variety of the learning object. Based on the X-ray image data of the inspected object, the virtual image production unit produces virtual transmission images of other energy bands using the virtual image production model. The determination unit performs a determination based on the X-ray image data of a specified energy band and the virtual transmission images of other energy bands produced by the image production unit (for example, see Patent Document 3).

[0006] Patent Document 1: Japanese Patent Application Publication No. 2012-137387

[0007] Patent Document 2: Japanese Patent Application Publication No. 2023-114828

[0008] Patent Document 3: Japanese Patent Application Publication No. 2021-148486

[0009] However, in conventional product inspection devices as described above, when there are large deviations in the products judged as qualified or when there are many types of defects such as foreign objects, a large number of images corresponding to the quantity of deviations or types are required. In particular, for example, in product inspection devices equipped with image inspection functions that effectively utilize AI, in addition to using only qualified product images as learning images, it is also necessary to obtain images of defective products such as those with foreign objects mixed in or those with defective product shapes based on the qualified product images, thus requiring even more images.

[0010] In contrast, inspection equipment that uses cameras or X-ray images can, for example, prepare an image containing only foreign objects and composite that image into an image of a qualified product, thereby making it easy to create an image of a product with foreign objects mixed in.

[0011] However, on the product manufacturing line, it is impossible to know in advance what kind of foreign objects will be mixed into the manufactured products and what kind of defects will occur. For example, in the case of bone foreign objects that may be mixed into meat, the bones that are cut along with the meat are of various shapes. Therefore, it is difficult to prepare learning images that correspond to all of them, and a large number of learning images are still needed.

[0012] Furthermore, even for qualified products, such as cooked foods or processed sausages, the size or shape of the individual contents may differ from that of typical industrial products. In cases where contents overlap in packaged products, the inspection images will be entirely different depending on the degree of overlap. Therefore, a large number of training images are also required in such situations.

[0013] Moreover, when detecting defective products with abnormal shapes, the frequency of defective products is low and it is difficult to prepare physical samples. Therefore, even if the shape is not good, it is difficult to create fragmented or distorted images of defective products in a natural way.

[0014] On the other hand, when an item inspection device incorporates an image inspection function that effectively utilizes AI (artificial intelligence) or performs its performance verification, or when inspection images are used in performance verification within rule-based processing, it is necessary to improve the accuracy of learning or performance verification by using as many patterned images of foreign objects or defective products as possible in a balanced manner. Therefore, in conventional item inspection devices, it is not possible to set inspection conditions with a small number of images, and the acquisition of inspection images used in learning or performance verification and the setting of inspection conditions based on those images require time. Summary of the Invention

[0015] Therefore, the object of the present invention is to provide an item inspection apparatus that can improve the accuracy of learning or performance verification even without the time and effort required to acquire inspection images used in learning or performance verification for setting inspection conditions for the items to be inspected.

[0016] (1) In order to achieve the above objective, the article inspection apparatus of the present invention includes an inspection unit that uses an inspection image obtained by photographing the article being transported to inspect the quality status of the article. The article inspection apparatus is characterized in that the inspection unit sets inspection conditions for the quality status of the article based on a virtual image of the inspection image generated by generative AI.

[0017] According to this structure, in this invention, in order to set inspection conditions for the quality status of an item, virtual images of the inspection image are generated by generative AI. Therefore, based on the size of the item being inspected, deviations in shape including fragmentation or deformation, or various configurations or shapes of its contents, or unpredictable shapes or sizes of foreign objects mixed in, multiple virtual images can be easily generated as simulation images of various inspection images (or segmented images). Therefore, data expansion of the original image can be effectively performed to set inspection conditions, and the accuracy of machine learning or performance verification can be effectively improved.

[0018] (2) In a preferred embodiment of the present invention, the structure can be configured as follows: the inspection unit uses a learned model created by learning the virtual image to set the inspection conditions.

[0019] In this case, during the learning phase after the model has been learned, it is possible to learn in a balanced manner various shape deviations, such as fragmentation or deformation of the inspection object, various configurations or shapes of the contents of the inspection object, and unpredictable shapes or sizes of foreign objects mixed in, thereby ensuring the accuracy required in model learning or performance verification.

[0020] (3) In a preferred embodiment of the present invention, the structure may be configured as follows: it further includes: an image processing algorithm storage unit that stores a plurality of image processing algorithms in advance; and an evaluation unit that evaluates the suitability of the inspection of the quality status using a virtual image of the inspection image for each of the plurality of image processing algorithms and calculates a plurality of evaluation values, wherein the inspection unit applies an image processing algorithm selected based on the plurality of evaluation values ​​to inspect the quality status of the article.

[0021] (4) In a preferred embodiment of the present invention, the structure can be set as follows: the dataset used in the learning of the learned model includes data that creates these virtual images based on the images of qualified samples and automatically adds OK tags.

[0022] (5) In a preferred embodiment of the present invention, the dataset used in the learning of the learned model includes data that creates virtual images based on images of defective samples and automatically adds defective labels; and data that creates virtual images based on images of defective samples and automatically adds foreign object labels.

[0023] (6) In a preferred embodiment of the present invention, the structure may be configured as follows: the evaluation unit evaluates whether the qualified product determination can be performed with a probability of a specified accuracy rate or higher and whether the defective product determination can be performed with a probability of a specified accuracy rate or higher.

[0024] In this case, multiple virtual images can be used as simulation images of various inspection images to evaluate multiple image processing algorithms, and based on the evaluation values ​​of each image processing algorithm, any image processing algorithm suitable for the quality state of the inspected item can be selected. Therefore, this becomes an item inspection device that can improve the accuracy of performance verification even without the time and effort required to obtain the inspection images used in the performance verification of the image processing algorithms.

[0025] Invention Effects

[0026] According to the present invention, an item inspection apparatus can be provided that improves the accuracy of learning or performance verification even without the time and effort required to acquire inspection images used in learning or performance verification for setting inspection conditions for the item to be inspected. Attached Figure Description

[0027] Figure 1 This is a schematic structural diagram of an article inspection device according to an embodiment of the present invention.

[0028] Figure 2This is an illustrative diagram illustrating an example of a qualified product group, which is an original image of a qualified product sample used in the learning process of a judgment learning model in an article inspection apparatus according to an embodiment of the present invention, simulates various qualified product groups with deviations in individual posture or configuration, and generates a virtual image of a qualified product group through image generation AI.

[0029] Figure 3 This is an illustrative diagram of an example of a defective sample group using an original image of a shape defective sample in the learning process of a judgment learning model in an article inspection apparatus according to an embodiment of the present invention, which simulates the posture or configuration deviation of an individual, and generates a virtual image of a defective sample group through image generation AI.

[0030] Figure 4 This is an explanatory diagram of an example of a foreign object group, which is the original image of the foreign object group used in the learning of the determination learning model in the article inspection device according to an embodiment of the present invention, and a virtual image of the foreign object group generated by image generation AI to simulate various foreign objects.

[0031] Figure 5 This is a flowchart illustrating the general steps of a setup operation including model learning in an item inspection apparatus according to an embodiment of the present invention.

[0032] Figure 6 This is a schematic structural diagram of an article inspection device according to another embodiment of the present invention. Detailed Implementation

[0033] Hereinafter, the methods for carrying out the present invention will be described with reference to the accompanying drawings.

[0034] (One implementation method)

[0035] Figures 1 to 5 An article inspection device according to one embodiment of the present invention is shown.

[0036] First, let's explain the structure.

[0037] like Figure 1 As shown, the article inspection apparatus 1 of this embodiment includes: a conveying unit 10 for conveying the object to be inspected, i.e., article P; an inspection unit 20 having the function of photographing the article P during conveying; a control unit 30 including a main control unit for controlling the conveying unit 10 and the inspection unit 20; a display unit 40 with operable inputs such as a touch panel; and a learning unit 50 for causing the control unit 30 to perform predetermined machine learning. Furthermore, the article inspection apparatus 1, through the inspection unit 20, for example, irradiates the article P conveyed by the conveying unit 10 with X-rays and photographs it, while simultaneously detecting image data corresponding to the distribution of transmitted X-rays, and inspects the quality status of the article P based on this image data.

[0038] In addition, the quality status mentioned here refers to whether the quality or physical quantity required for the article P as a product is appropriate, such as whether there are foreign objects mixed in, whether there are missing parts, whether the shape / size / containment state of the contents is qualified, and the distribution of density / thickness / volume or quality.

[0039] The conveying unit 10 winds the annular conveyor belt 11 around the conveyor roller 12 on the drive side and the conveyor roller 13 on the driven side, and feeds the items P that are sequentially fed into the upward section 11a of the conveyor belt 11 from the upstream side. Figure 1 It is conveyed to the right and passed through the camera section of the inspection unit 20 and moved to the downstream conveyor, and supported on a frame not shown.

[0040] The inspection unit 20 includes, for example, an X-ray generator (X-ray source) 21 that generates X-rays of a specified energy band that transmit through the article P conveyed by the transport unit 10, and an X-ray detector 23 disposed directly below the upward section 11a of the conveyor belt 11. Furthermore, the inspection unit 20 is not limited to a component that irradiates the article P with X-rays to obtain an inspection image Dpx. For example, it could be a component that uses NIR (near-infrared) light to obtain an inspection image from a camera showing the appearance or transmission, or it could be a component that obtains an inspection image from a color image of the article's appearance using other light sources such as visible light.

[0041] The X-ray generator 21 generates X-rays with wavelengths and intensities corresponding to the tube current and tube voltage through the X-ray tube 22, and irradiates the article P on the conveyor belt 11 with fan-shaped beams of X-rays directed toward the main observation direction orthogonal to the article transport direction of the conveyor section 10 through the X-ray window of the peripheral equipment (details not shown).

[0042] Although not detailed in the illustration, the X-ray detector 23 is, for example, composed of an X-ray linear sensor camera and is positioned at a predetermined position corresponding to the X-ray irradiation position from the X-ray generator 21. The X-ray linear sensor camera is equipped with detection elements, such as phosphors (scintillators) and photodiodes or charge-coupled devices, arranged in an array at predetermined intervals along the width of the transport path of the transport section 10, and outputs an X-ray detection signal Lx at a predetermined resolution equivalent to the transmittance.

[0043] That is, the inspection unit 20 can detect X-rays transmitted from the X-ray generator 21 to the article P in each predetermined transmission area corresponding to the detection element, convert them into electrical signals corresponding to the transmission amount of the X-rays, and output an X-ray detection signal Lx for generating an X-ray transmission image with the transmission direction of the X-rays as the observation direction. Furthermore, the X-ray detector 23 performs a main scan in the width direction corresponding to the conveyor belt 11's conveying speed and sequentially outputs X-ray detection signals Lx. If the article detection sensor 28 detects that the article P has been placed on the conveyor belt 11 and is being conveyed towards the predetermined inspection interval Zx, it outputs the X-ray detection signal Lx of the article P detected after a predetermined time.

[0044] The control unit 30 has the functions of a conveying control mechanism that controls the conveying speed or conveying interval of the article P based on the conveyor belt 11 in the conveying unit 10, and an inspection control mechanism that controls the X-ray irradiation intensity or irradiation period in the inspection unit 20, or controls the X-ray detection cycle and detection period of each article P in the X-ray linear sensor of the X-ray detector corresponding to the conveying speed of the article P.

[0045] The control unit 30 is configured to include, for example, a microcomputer (processor) having a CPU, ROM, RAM and I / O interface (not shown); a program device that readablely stores control programs for performing the functions of the multiple functional units described later in the ROM or auxiliary storage device and other recording media or downloads them from other computers via data communication; and a timer circuit, etc., and according to the control program stored in the ROM, etc., the CPU performs prescribed arithmetic processing while sending and receiving data with the RAM, etc., and executes the control programs of the multiple functional units.

[0046] The control unit 30 has an inspection image storage unit 31, an inspection processing unit 32, and a learned model 35 as the main functional units for performing the functions of the aforementioned inspection control mechanism. The inspection processing unit 32 is configured to include an image processing unit 33 and a determination unit 34.

[0047] The inspection image storage unit 31 sequentially reads the X-ray detection signal Lx from the X-ray detector 23 of the inspection unit 20, temporarily stores the image data representing the X-ray transmittance distribution of each item P in the memory, and outputs the image data as the image data of the inspection image Dpx.

[0048] The image processing unit 33 of the inspection processing unit 32 performs image analysis processing for extracting global or local features of the image (e.g., extracting feature quantities of local regions based on pixel values ​​or brightness gradients, extracting frequency feature quantities of the image as a whole such as spatial spectrum) by setting one or more predetermined filtering processes that sequentially read image data of the inspection image Dpx output from the inspection image storage unit 31. Furthermore, the predetermined filtering process described here is a filtering process that detects or emphasizes image features (e.g., edges or spots) that indicate a deviation in quality state from the normal state, i.e., a degree different from the normal level, using the aforementioned predetermined image processing algorithm.

[0049] Furthermore, the determination unit 34 of the inspection processing unit 32 performs inspection image processing, which can determine whether the item P has a specified quality state based on the image processing result in the image processing unit 33 and a specified image processing algorithm.

[0050] Furthermore, the image processing unit 33 and the determination unit 34 of the inspection processing unit 32 also have the function of performing deep learning-based classification or anomaly detection by cooperating with the learned model 35 and using the inspection image Dpx of the item P obtained from the inspection image storage unit 31 or the inspection image after the aforementioned filtering process, thereby determining with confidence whether the item P has a specified quality state and whether its quality state is normal. Here, classification is a process that performs image class classification by extracting features from the input image and learning decision boundaries, for example, to determine the type of the inspected item. Anomaly detection is a process that detects abnormal parts in the input image, such as partial defects in the contents of the inspected item or unevenness outside the normal range, as abnormalities.

[0051] The learned model 35 is a program module (artificial intelligence software) that enables the inspection processing unit 32 to perform deep learning-based classification or anomaly detection functions based on the inspection image or the inspection image after specified filtering. The cooperation between the inspection processing unit 32 and the learned model 35 means that the inspection processing algorithm Pgm, which is based on a specified image processing algorithm, is input from the learned model 35 into the memory of the inspection processing unit 32 as a program module, and the program module performs the aforementioned classification or anomaly detection functions.

[0052] The model 35, during its learning phase, is a model that learns the features of qualified product images based on image data of qualified product images without abnormalities such as foreign objects, or further learns the features of defective product images and their defective parts based on image data of defective product images with defective parts caused by abnormalities such as foreign objects.

[0053] Specifically, this is a learning model that inputs a predetermined number (e.g., more than 1000) of qualified images for training into the learned model 35, and adjusts these parameters using parameters such as the weights between layers of a multi-layer neural network, for example, the weighted weights in the hidden layer (intermediate layer) relative to any i-th neuron of the input layer, and the weighted weights in the output layer of any k-th neuron relative to any j-th neuron of the hidden layer. Here, i, j, and k are arbitrary natural numbers.

[0054] like Figure 2 As shown, the dataset used in the learning of Model 35 is, for example, a dataset that creates virtual images Dpp1 based on the qualified sample image Dps1, and automatically appends OK labels that may be used for image classification or anomaly detection. Furthermore, as... Figure 3 and Figure 4 As shown, the dataset used in the learning of Model 35 includes data that creates virtual images Dpp2 based on defective sample images Dps2 and automatically adds defective labels for image classification, and data that creates virtual images Dpp3 based on defective sample images Dps3 and automatically adds the categories and labels of defective parts such as foreign objects for object detection.

[0055] In addition, Figure 1 For ease of illustration, virtual images Dpp1 to Dpp3 are referred to as multiple virtual images Dpp. However, in the following description, the dataset obtained by attaching OK labels to virtual images Dpp1 of a specified number of qualified products is called dataset Dpp1, the dataset obtained by attaching defective labels to virtual images Dpp2 of a specified number of defective products is called dataset Dpp2, and the dataset obtained by attaching foreign object categories and labels to images Dps3 of a specified number of defective samples is called dataset Dpp3.

[0056] After learning the image data of each item P for model 35, the weights and other parameters between the layers are adjusted so that the output value of the multilayer neural network is distributed in the attribute area of ​​the main feature quantity distribution of the qualified product label of each qualified product image or in the attribute area of ​​the main feature quantity distribution of the defective product label of each defective product image in the feature space based on the global or local features of the above images.

[0057] During the utilization phase, when the inspection image data of the qualified product inspection image Dpx is input from the inspection image storage unit 31 to the inspection processing unit 32, the output value of the neural network determines whether the centroid or distribution pattern of the distribution of feature quantities (feature vectors) of the inspection image Dpx and its respective judgment pixel regions in the feature space based on the global or local features of the aforementioned image is located within the attribute region of the qualified product label distributed in the qualified product image of the learning process or within the attribute region of the defective product label distributed in the main feature quantities of each defective product image.

[0058] In addition, after learning, Model 35 can not only perform the functions of an object detection model, but also learn whether the rectangle in the image contains an object or the background. In the case of an object, it can learn to reduce the error between the object category and the ground truth label.

[0059] The learning unit 50 is configured to include: a processor, such as a GPU (Graphics Processing Unit) or VPU (Visual Processing Unit), which is an arithmetic module that performs image processing functions in cooperation with or independently of the CPU of the control unit 30; a program device that can readablely store a control program for performing multiple learning functions in a recording medium such as ROM or auxiliary storage device or download it from another computer via data communication; and a timing circuit, etc., and is capable of executing the control program for performing multiple learning functions according to the control program stored in ROM, etc.

[0060] The learning unit 50 has an image generation AI model, namely generative AI 51, which is a main functional unit that performs multiple functions for learning; a virtual image storage unit 52 that stores a dataset of virtual images generated by generative AI 51; and a learning processing unit 53 that performs learning processing, such as adjusting the weights between layers of each neuron, on the dataset of virtual images stored in the virtual image storage unit 52 during the learning phase of the learning completed model 35.

[0061] Generative AI51 generates images using the following datasets: datasets used in the learning of Model 35, such as dataset Dpp1, which appends OK labels (potentially used for image classification or anomaly detection) to virtual qualified product images (feature vectors in a multi-dimensional (hereinafter, also referred to as n-dimensional) latent space) that are similar to the qualified product sample image Dps1; dataset Dpp2, which appends defective product labels for image classification to virtual defective product images (feature vectors in a multi-dimensional (hereinafter, also referred to as n-dimensional) latent space); and dataset Dpp3, which automatically appends the category and label of defective parts, such as foreign objects, for object detection to virtual defective product images (defective product images Dps3) that are similar to the defective product sample image Dps3. Furthermore, the similarity of semantic feature vectors can be determined based on the size of the inner product of multiple feature vectors compared in the n-dimensional latent space.

[0062] The generative AI51 includes, for example, the following functions: an image classification model that learns paired data of input text (prompt words) and images associated with their semantic content, and estimates the similarity between the feature vectors of the images and text; an encoder that converts the input text into feature vectors associated with its semantic content, or compresses the input image according to specified parameters and converts it into multidimensional feature vectors associated with its semantic content; a diffusion model that performs a forward diffusion process and a reverse denoising process after setting the multidimensional feature vectors (representation vectors in the latent space) obtained by compressing the input image, etc., as a list of latent variables; and a decoder that decodes and generates an image based on the denoised latent variables and outputs it as the output image. The specified settings mentioned here are, for example, settings in the image generation AI model "Stable Diffusion" that specify the degree to which the original image (raw image) is maintained or noise-enhanced, such as Prompt / Denoising strength, Number of Outputs, Steps, and Guidance Scale for adjusting the similarity between the prompt and the generated image.

[0063] Furthermore, the generative AI51 estimates the similarity between the semantic feature vector of the output image after denoising in the above diffusion model and the semantic feature vector corresponding to the request content based on the input text, and the similarity between the output image and the semantic feature vector of the image classification model. It then repeatedly generates similar images until their similarity reaches or exceeds a pre-set similarity threshold (e.g., based on the settings of the above Prompt strength and Guidance Scale) to generate a virtual image. The generation of virtual images with similarity reaching or exceeding the pre-set similarity threshold is repeated until the required number of learning images is reached, as specified by the above-defined settings (e.g., Number of Outputs).

[0064] The generative AI51 is installed after the image classification model for general images has been sufficiently improved through pre-learning, and the input sample image to be referenced and its supplementary description as text and the original image are used.

[0065] For example, regarding the segmented images of the original image (e.g., pixels of various categories assigned through semantic segmentation), generative AI51 creates a modified image whose shape, size, configuration, etc., are randomly changed within a specified range that can be obtained as a product, based on the request content in the text input, i.e., the prompt words, to create the datasets Dpp1 to Dpp3 of the aforementioned first to third virtual images.

[0066] Therefore, the aforementioned variation images generated by generative AI51 are obtained by randomly changing the location, shape, or configuration, posture, shape, and packaging of defective parts of the product, or the contents of the product, according to the request content in the text input prompt, and thus changing the image in various ways as a whole. Of course, preprocessing has been performed to standardize each image in the dataset to the same scale.

[0067] The datasets Dpp1 to Dpp3 of the first to third virtual images generated by the generative AI 51 are sequentially stored in the virtual image storage unit 52. The datasets Dpp1 to Dpp3 of the various virtual images Dpp that are required for the learning process are provided to the learning model 35 through the learning processing unit 53.

[0068] Next, the function will be explained.

[0069] In the article inspection apparatus 1 of this embodiment configured as described above, before the article inspection begins, as follows: Figure 5 The outline steps shown are for performing the setup operation.

[0070] First, the generative AI 51 of the learning unit 50 is introduced into a program device that fully enhances the function of a general image classification model through pre-learning, or a web service that enables the image generation AI to be used via a PC or server connected to the company's internal network (step S11).

[0071] Next, a sample image of the image to be used as a reference for the virtual image, along with its supplementary description, will be used as follows: Figure 1 The text and original image shown are input into the generative AI51 (step S12), which generates images to create the first to third datasets Dpp1 to Dpp3 of various virtual images Dpp (step S13).

[0072] At this time, for example, regarding the segmented images of the original images Dps1 to Dps3, deformed images whose shape, size, configuration, etc., are randomly changed within the prescribed limits that can be obtained as products are created according to the request content in the prompt words, and the datasets Dpp1 to Dpp3 of the first to third virtual images are stored sequentially in the virtual image storage unit 52.

[0073] Next, the datasets Dpp1 to Dpp3 of the first to third virtual images are provided to the learned model 35 through the learning processing unit 53. As a result, the weights and other parameters of each neuron of the learned model 35 are adjusted in a balanced and effective manner, thereby obtaining the image classification function or anomaly detection function based on the inspection image Dpx of the learned model 35 with high accuracy (step S14).

[0074] If the parameter adjustment is completed, during the utilization phase of the learned model 35, after setting the type of the item P to be inspected, the image data of the inspection image Dpx of item P is input from the inspection image storage unit 31.

[0075] Specifically, initially, for example, in order to confirm the operation under the inspection conditions of the product being inspected, multiple foreign object samples with different diameters are prepared and attached to a test sample defective product of the product being inspected. This test sample defective product is inserted and photographed by the inspection unit 20, and the image data of the inspection image Dpx is acquired by the inspection image storage unit 31. Moreover, it is known that if the image data of the inspection image Dpx is input into the inspection processing unit 32, the determination unit 34 of the inspection processing unit 32, which cooperates with the learned model 35, determines that the photographed test sample defective product is abnormal, and performs the required foreign object detection function (step S15).

[0076] If the inspection conditions for this type of inspection object are set and confirmed, the setting is complete. Then, the inspected items, i.e., articles P, of the inspection object type are sequentially fed in and photographed at a specified conveying interval or conveying distance unit. The image data of the multiple articles P fed in are sequentially input from the inspection image storage unit 31 to the inspection processing unit 32 as the image data of the inspection image Dpx.

[0077] At this time, it can be seen that in the inspection processing unit 32, the determination unit 34 of the inspection processing unit 32, which cooperates with the learned model 35, determines whether the image data of the inspection image Dpx from the inspection image storage unit 31 is a normal quality state in which there are no foreign objects or defective parts in each item P, and displays the inspection results in the display unit 40, thus performing the required item inspection function.

[0078] In this embodiment, to set the inspection conditions for the quality status of item P, a dataset of virtual images Dpp1 to Dpp3, the first to third virtual images of inspection image Dpx, is generated by generative AI51. Therefore, based on the size of item P, the deviation of its shape (including fragmentation or deformation), the various configurations or shapes of its contents, unpredictable shapes or sizes of foreign objects, and poor packaging, the semantic changes in the image based on the prompt words can be made. Thus, multiple virtual images Dpp can be easily generated as simulation images of various inspection images (which may be segmented images), i.e., images matching the request content. Therefore, a learning dataset corresponding to a wide range of actual morphological changes that cannot be obtained by simply expanding the data to the extent of inverting, rotating, scaling, parallel shifting, brightness changes, and contrast changes of the original image, can be obtained from the original image data effective for setting inspection conditions. This effectively improves the accuracy of the learned model 35 in learning or performance verification.

[0079] Furthermore, in this embodiment, the control unit 30 of the inspection unit 20 uses a learned model 35, which is created as inspection conditions for learning the virtual image Dpp, to perform processing of the quality status of the inspection item P. Therefore, during the learning phase of the learned model 35, it is possible to learn in a balanced manner various shape deviations, including fragmentation or deformation of the inspection item, various configurations or shapes of the contents of the inspection item, and unpredictable shapes or sizes of foreign objects mixed in. This learned model 35 ensures the accuracy required for machine learning or performance verification.

[0080] Thus, according to this embodiment, during the learning phase of the learned model 35, the datasets Dpp1 to Dpp3 of the first to third virtual images, generated evenly by the generative AI 51, are provided to the learned model 35 for learning through the learning processing unit 53. As a result, the weights and other parameters of each neuron in the learned model 35 are adjusted evenly and effectively, thereby ensuring image classification or anomaly detection functions based on the learned model 35 with high accuracy.

[0081] Therefore, an item inspection device 1 can be provided that improves the accuracy of learning or performance verification even without the time and effort required to acquire inspection images used in learning or performance verification for setting inspection conditions for item inspection.

[0082] (Another implementation method)

[0083] Figure 6 The main structural components of an article inspection device according to another embodiment of the present invention are shown.

[0084] Figure 6 The article inspection device 2 of another embodiment shown has a structure similar to that of the article inspection device 1 of one embodiment. Therefore, the similar structural parts are marked with the same reference numerals as the corresponding constituent elements of one embodiment, and repeated descriptions are omitted.

[0085] The item inspection device 2 of this embodiment includes: a main control unit 70, which includes a control transport unit 10 and an inspection unit 20; and an image processing control unit 75, which has an inspection image storage unit 31 and an inspection processing unit 32, and switchesably supplies the inspection processing unit 32 with an inspection processing algorithm Pgm (a program module including an image processing algorithm).

[0086] Specifically, the image processing control unit 75 includes a generative AI 81, a virtual image storage unit 82, an algorithm evaluation unit 83, an algorithm storage unit 84, and an algorithm selection unit 85.

[0087] The generative AI81 can input prompts from the text input on the display unit 40, which is operable and can input image data of qualified samples and / or defective samples (which may only contain the defective parts) captured by the inspection unit 20 from the X-ray detector 23 in the form of X-ray detection signal Lx, and standardize them with the same scale and grayscale density as the inspection image Dpx output from the inspection image storage unit 31.

[0088] Furthermore, the generative AI81 has the same image classification model function, encoder function, diffusion model function, and decoder function as the generative AI51 in one embodiment. It estimates the semantic feature vector of the output image after denoising in the diffusion model and decoding and output by the decoder function, and the similarity between the feature vector corresponding to the request content based on the input text and the image classification model function. It then repeatedly generates similar images until their similarity reaches or exceeds a preset similarity determination threshold to generate a similar image. The generation of similar images is then repeated until the number of images whose similarity reaches or exceeds the preset similarity determination threshold becomes the required number of learning images specified by the above-mentioned setting parameters, such as "Number of Outputs".

[0089] For the generative AI81, the image data of qualified samples and / or defective samples (which may only contain the defective parts) are required to be learned as the original image input, and a prompt word is input from the text input of the display unit 40, which is operable input such as a touch panel, as the request content. The request content is the request content that causes the location and shape of the defective part of the target product, or the configuration, posture, shape, packaging form of the contents of the product, to change randomly for each individual in the image and to change in various ways in the image as a whole.

[0090] At this time, the modified images generated in the generative AI 81 based on the original image and the prompt words are sequentially stored in the virtual image storage unit 82 and maintained in a storage state that can be read by the algorithm evaluation unit 83.

[0091] The algorithm evaluation unit 83 sequentially reads from the algorithm storage unit 84 a plurality of rule-based inspection processing algorithms, mainly image processing algorithms included in the inspection algorithm of the item inspection device 2, and evaluates (sensitivity evaluates) whether the inspection processing algorithms read in the evaluation (sensitivity evaluation) of a predetermined number of virtual qualified product images equivalent to the first virtual image dataset Dpp1 can perform the determination of the truth label, i.e., qualified product determination, with a probability of a predetermined positive resolution rate or higher, and / or whether the inspection processing algorithms read in the evaluation (sensitivity evaluation) of a predetermined number of virtual defective product images equivalent to the second virtual image dataset Dpp2 can perform the determination of the truth label, i.e., defective product determination, with a probability of a predetermined positive resolution rate or higher.

[0092] The algorithm storage unit 84 has a storage capacity such that it can store multiple rule-based inspection processing algorithms that can be used in the inspection processing unit 32, such as foreign object detection algorithms or missing part detection algorithms, in a identifiable manner according to multiple algorithm numbers, and can add / update these multiple inspection processing algorithms as part of the inspection algorithm of the item inspection device 2.

[0093] When requesting a new variety selection setting or performance verification of a specific inspection processing algorithm from the display unit 40, the algorithm selection unit 85 evaluates, based on the dataset of multiple virtual images Dpp stored in the algorithm storage unit 84, whether the algorithm can perform inspection processing with a specified accuracy rate or higher for any category or specific inspection processing algorithm stored in the algorithm storage unit 84, and selects the best inspection processing algorithm Pgm for the variety or variety group of the inspection object based on the results.

[0094] The algorithm selection unit 85 can also replace the inspection processing algorithm of the image processing unit 33 installed in the inspection processing unit 32 with an inspection processing algorithm selected by the algorithm selection unit 85 as the optimal algorithm. Therefore, the image processing control unit 75 can verify the performance of image processing and judgment processing based on the inspection processing unit 32.

[0095] Thus, the item inspection device 2 of this embodiment includes: an algorithm storage unit 84 (image processing algorithm storage unit) that pre-stores inspection processing algorithms based on multiple image processing algorithms; and an algorithm evaluation unit 83 (evaluation unit) that evaluates the suitability of quality status inspection for multiple inspection processing algorithms using datasets of multiple virtual images Dpp to determine the correctness rate of the labels, and calculates the correctness rate of the determination for multiple inspection processing algorithms as multiple evaluation values ​​representing each evaluation result. The inspection unit 20 can use the algorithm selection unit 85 to select the set inspection processing algorithm Pgm based on the multiple evaluation values ​​obtained by the algorithm evaluation unit 83 to inspect the quality status of the item P to be inspected.

[0096] Therefore, in this embodiment, multiple virtual images (Dpp) can be used as simulation images of various inspection images to evaluate multiple image processing algorithms, and an inspection processing algorithm Pgm can be selected and set based on the evaluation value (correctness rate) of each image processing algorithm to include any image processing algorithm suitable for the quality state of the inspected item P. Therefore, even without the time and effort required to obtain inspection images for performance verification of the inspection processing algorithm Pgm, the accuracy of the performance verification can be improved.

[0097] Thus, in this embodiment, similar to the first embodiment, an item inspection apparatus can be provided that improves the accuracy of learning or performance verification even without the time and effort required to acquire inspection images used in learning or performance verification for setting inspection conditions for item inspection.

[0098] Furthermore, in the above embodiments, generative AI 51 and generative AI 81 are respectively set as input text and image components. However, it is also possible to input only prompt text into generative AI 51 and 81 based on the processing content of the inspection processing unit 32 and the image classification model function of generative AI 51 and 81 in a pre-learned state. In this case, the request content of the image classification task based on prompt text can be set in more detail according to the product form. Moreover, it goes without saying that the image generation AI model used as generative AI 51 and generative AI 81 is not limited to a specific model. In addition, in another embodiment, generative AI 81 is described as being programmed and installed in the control unit 70, but similarly to generative AI 51 in one embodiment, it can of course be configured outside the control unit 70.

[0099] As explained above, this invention provides an item inspection apparatus that improves the accuracy of learning or performance verification even without the time-consuming and laborious process of acquiring inspection images used to set inspection conditions for the inspected items. This invention is effective in all item inspection apparatuses that use inspection images obtained by photographing the inspected item and a learning model or inspection algorithm to inspect the quality status of the item.

[0100] Symbol Explanation

[0101] 1, 2 - Item inspection device; 10 - Conveying unit; 11 - Conveyor belt; 11a - Upward section; 12 - Conveying roller; 13 - Conveying roller; 20 - Inspection unit (camera unit); 21 - X-ray generator (X-ray source); 22 - X-ray tube; 23 - X-ray detector; 28 - Item detection sensor; 30 - Control unit; 31 - Inspection image storage unit (inspection image acquisition unit); 32 - Inspection processing unit; 33 - Image processing unit; 34 - Judgment unit; 35 - Learned model (artificial intelligence software, program module); 40 - Display unit (touch panel); 50 - Learning unit; 51, 81 - Generative AI (image generation AI model, artificial intelligence software); 52, 82 - Virtual image storage unit; 53 - Learning processing unit; 70 - Control unit; 75 - Image processing control unit, 83- Algorithm evaluation unit, 84- Algorithm storage unit, 85- Algorithm selection unit, Dpp- Multiple virtual images, Dpp1- Virtual image (first virtual image, dataset of first virtual image) Dpp2- Virtual image (second virtual image, dataset of second virtual image), Dpp3- Virtual image (third virtual image, dataset of third virtual image), Dps1- Sample image (qualified sample image, original image), Dps2- Defective sample image (original image), Dps3- Defective part sample image (original image), Dpx- Inspection image, Lx- X-ray detection signal, P- Item (inspected object), Pgm- Inspection processing algorithm (image processing algorithm, program module), Zx- Defined inspection range.

Claims

1. An article inspection apparatus that has an inspection unit (20) that inspects a quality state of an article (P) using an inspection image (Dpx) obtained by photographing the article, the article inspection apparatus being characterized in that the inspection unit sets an inspection condition of the quality state of the article in accordance with a virtual image (Dpp) of the inspection image generated by a generative AI. 2.The article inspection apparatus according to claim 1, characterized in that the inspection unit sets the inspection condition using a learned model (35) made by learning the virtual image.

3. The article inspection apparatus according to claim 1, characterized by, Further comprising: an image processing algorithm storage unit (84) that stores a plurality of image processing algorithms in advance; and an evaluation unit (83) that respectively evaluates the suitability of inspection of the quality state using the virtual image of the inspection image for the plurality of image processing algorithms and calculates a plurality of evaluation values, the inspection unit inspects the quality state of the article using an image processing algorithm (Pgm) selected in accordance with the plurality of evaluation values.

4. The article inspection apparatus according to claim 2, characterized by, Further comprising: an image processing algorithm storage unit (84) that stores a plurality of image processing algorithms in advance; and an evaluation unit (83) that respectively evaluates the suitability of inspection of the quality state using the virtual image of the inspection image for the plurality of image processing algorithms and calculates a plurality of evaluation values, the inspection unit inspects the quality state of the article using an image processing algorithm (Pgm) selected in accordance with the plurality of evaluation values. 5.The article inspection apparatus according to claim 2, characterized in that a data set used in learning of the learned model contains data in which the virtual images (Dpp1) are made from a good sample image (Dps1) and an OK label is automatically attached. 6.The article inspection apparatus according to claim 2, characterized in that a data set used in learning of the learned model contains data in which the virtual images (Dpp2) are made from a defective sample image (Dps2) and a defective label is automatically attached and data in which the virtual images (Dpp3) are made from a bad portion sample image (Dps3) and a foreign matter label is automatically attached. 7.The article inspection apparatus according to claim 3, characterized in that the evaluation unit evaluates whether or not a good determination can be performed with a probability of a prescribed correct rate or higher and whether or not a defective determination can be performed with a probability of a prescribed correct rate or higher.

Images