Inspection device, program, inspection method and production method

Abstract

To more accurately detect a defect caused in an inspected object.SOLUTION: An inspection device for inspecting a defect caused in an inspected object by using an original image, includes: a restoration unit that converts the original image into an intermediate value having a lower degree of freedom and converts the intermediate value into a restored image; and a detection unit that detects the defect caused in the inspected object, based on a difference image between the original image and the restored image. An inspection method for inspecting a defect caused in an inspection object by using an original image, includes: converting the original image into an intermediate value having a lower degree of freedom; converting the intermediate value into a restored image; and detecting the defect caused in the inspected object, based on the difference image between the original image and the restored image.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present application relates to an inspection apparatus, a program, an inspection method, and a production method.The present application relates to, for example, a technique for determining the state of a defect occurring in an object to be inspected by using image data representing an image of the object to be inspected. [Background technology]

[0002] Glass materials formed by laminating multiple glass layers are also called laminated glass or layered glass. Laminated glass is sometimes used as automobile window glass. Insufficient degassing during the production process (lamination process) can cause quality problems with laminated glass. For example, this can impair visibility or reduce strength or airtightness. Therefore, it is desirable to detect defects that could cause quality problems as much as possible during the production process.

[0003] Previously, attempts have been made to use machine learning models to detect defects in glass materials. For example, Patent Document 1 describes an inspection device that includes an image acquisition unit that acquires an image of a glass plate, and a size detection unit that inputs the image into a trained model that has learned the position and range of a defect in the glass plate so as to infer the position and range of the defect, and detects the size of the range in the inference result as the size of the defect in the imaged glass plate. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-187282 Summary of the Invention [Problem to be solved by the invention]

[0005] However, when using a model that has learned from defective areas in captured images, overdetection tends to occur. Overdetection refers to the detection of normal areas other than defective areas as defective areas. For example, in glass materials, edge areas where the spatial variation of signal values ​​is relatively large or areas where fine components are located tend to be detected as defective areas. Furthermore, in general, a large amount of training data is required to ensure a reasonable level of accuracy in learning machine learning models. In particular, for defects that become more pronounced due to aging, it can be difficult to collect or retrieve defective test pieces after shipment. Ultimately, it can be difficult to secure sufficient training data for learning about defect locations. The present application has been made in consideration of the above points, and one of the objects of the present invention is to provide an inspection device, a program, an inspection method, and a production method that can more accurately detect defects that occur in an object to be inspected. [Means for solving the problem]

[0006] (1) The present application has been made to solve the above-mentioned problems, and one aspect of the present invention is an inspection device that uses an original image to inspect defects that occur in an object to be inspected, the inspection device comprising: a restoration unit that converts the original image into an intermediate value with a lower degree of freedom and converts it into a restored image restored from the intermediate value; and a detection unit that detects defects that occur in the object to be inspected based on a difference image between the original image and the restored image. (2) Another aspect of the present invention may be a program for causing a computer to function as the inspection device described in (1).

[0007] (3) Another aspect of the present invention is an inspection method for inspecting defects occurring in an object to be inspected using an original image, which converts the original image into an intermediate value with a lower degree of freedom, converts the original image into a restored image restored from the intermediate value, and detects defects occurring in the object to be inspected based on a difference image between the original image and the restored image.

[0008] (4) Another aspect of the present invention may be a method for producing an object, comprising: sandwiching an intermediate layer between a plurality of glass layers to form a laminated glass as the object to be inspected; and the steps of: (3) inspecting the object; and (4) rejecting an object having the laminated glass in which a defect is detected. [Effects of the Invention]

[0009] According to the embodiment of the present application, defects occurring in an object to be inspected can be detected more accurately. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a schematic diagram illustrating a configuration example of an inspection system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a schematic block diagram showing an example of the functional configuration of an inspection device according to the present embodiment. [Figure 3] 1 is a schematic block diagram for explaining an overview of an inspection device according to an embodiment of the present invention. [Figure 4] FIG. 1 is a diagram for explaining a model according to the present embodiment. [Figure 5] FIG. 2 is a diagram showing a first example of a captured image of a sample. [Figure 6] 1A and 1B are diagrams showing a first example of a captured image, a restored image, and a difference image. [Figure 7] 10A and 10B are diagrams showing a second example of a captured image, a restored image, and a difference image. [Figure 8] 10 is a flowchart illustrating differential image processing according to the present embodiment. [Figure 9] FIG. 1 is a diagram illustrating the edge of a sample. [Figure 10] FIG. 2 is a diagram illustrating an example of a background portion of a captured image. [Figure 11] FIG. 10 is a diagram illustrating an example of a defect inspection range. [Figure 12] FIG. 10 is a diagram illustrating an example of a binary image within an inspection range. [Figure 13] FIG. 10 is a diagram illustrating an example of a binary image after noise processing. [Figure 14] FIG. 10 is a diagram illustrating an example of a binary image showing a defective area. [Figure 15] FIG. 10 is a diagram showing a first example of defective area detection. [Figure 16] 10 is a histogram showing a second example of defective area detection. [Figure 17] FIG. 10 is a diagram showing overdetection rates for each model. [Figure 18] 10 is a flowchart illustrating a defect detection process according to the present embodiment. [Figure 19] FIG. 10 is a diagram showing a second example of a captured image of a sample. [Figure 20] 10 is a flowchart illustrating a pass / fail determination process according to the present embodiment. [Figure 21] 10 is a histogram showing a third example of defective area detection. [Figure 22] 10 is a histogram showing a fourth example of defective area detection. [Figure 23] FIG. 10 is a diagram showing a third example of a captured image of a sample. [Figure 24] FIG. 10 is a diagram showing a sample image extracted from a captured image. [Figure 25] FIG. 10 is a diagram illustrating an example of a training image in which a sample image and a background image are combined. [Figure 26] 10 is a flowchart illustrating a training data generation process according to the present embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. <System configuration example> First, an outline of a configuration example of the inspection system S1 according to this embodiment will be described. Fig. 1 is a schematic diagram showing a configuration example of the inspection system S1 according to this embodiment. The inspection system S1 according to this embodiment captures an image of the sample Sp supplied from the pre-process 22, and inspects the sample Sp as an object to be inspected using the captured image before the sample Sp is sent to the post-process 26. The inspection system S1 sends the sample Sp that passes the inspection to the post-process 26, and ejects the sample Sp that does not pass the inspection without sending it to the post-process 26.

[0012] In this application, the sample Sp to be inspected is mainly an intermediate product produced in a preliminary bonding process in a production process for a windshield (WS) for a vehicle as a final product. Laminated glass is used for vehicle windows. Laminated glass is constructed by sandwiching an intermediate layer between two glass layers. The intermediate layer is typically a resin layer made of a transparent, thermoplastic synthetic resin. Examples of materials that can be used for the intermediate layer include PVB (Polyvinyl Butyral) and EVA (Ethylene-vinyl Acetate copolymer resin).

[0013] The pre-process 22 includes a preliminary pressure-bonding process. The preliminary pressure-bonding process involves sandwiching an intermediate layer between two glass plates and laminating them while applying heat to form an intermediate product. The preliminary pressure-bonding process is carried out using a nipper roll or the like. The post-process 26 includes a main pressure-bonding process. The main pressure-bonding process is a process in which the intermediate product is heated and pressurized in an environment with a higher temperature and pressure than the preliminary pressure-bonding process. In addition, we will take as an example a case in which the defect that occurs in the sample Sp is a defect that causes a difference in brightness from the surrounding area, such as an air bubble that occurs between the glass layer and the intermediate layer.

[0014] The inspection system S1 includes a conveyor 24, a sample detection unit 32, an illumination unit 34, an imaging unit 36, and a sorting unit 38. The conveyor 24 carries the sample Sp carried out from the equipment related to the previous process 22 (for example, a preliminary crimping device) and transports it at a substantially constant speed toward the equipment related to the next process 26 (for example, a main crimping device). The conveyor 24 may be a belt conveyor that generates power, or a roller conveyor that does not generate power.

[0015] The sample detection unit 32 detects samples Sp transported on the conveyor 24. The sample detection unit 32 is installed at a position lower than the transport surface of the conveyor 24. The sample detection unit 32 includes, for example, a transmission-type optical sensor. The optical sensor includes a light-emitter and a light-receiver, and is installed at positions facing each other across the conveyor 24. The light-receiver includes a light-receiving element that receives the inspection light projected from the light-emitting element of the light-emitter. When the inspection light is attenuated by the passage of the sample Sp, the received light intensity at the light-receiver becomes lower than a predetermined intensity threshold. At this time, the light-receiver can detect the passing sample Sp. When the received light intensity is equal to or higher than the detection threshold, the light-receiver does not detect the passing sample Sp. When the sample detection unit 32 detects the sample Sp, it outputs a detection signal indicating the detection to the imaging unit 36. In addition, one sample detection unit 32 may be installed at a position higher than the conveying surface of the conveyor 24, or two sample detection units 32 may be installed at positions lower and higher than the conveying surface of the conveyor 24, respectively.

[0016] The lighting 34 has a light source and irradiates the samples Sp placed on the conveyor 24 with light. When a detection signal is input from the sample detection unit 32, the imaging unit 36 ​​captures an image of the sample Sp being transported on the conveyor 24 after a preset first waiting time has elapsed from that point. The imaging unit 36 ​​captures an image of the sample Sp obtained by transmitting incident light from the illumination 34. When the incident direction of the transmitted light is included in the field of view of the imaging unit 36, a bright image is captured (bright field). The imaging unit 36 ​​is installed at a position (e.g., the ceiling) sufficiently higher than the conveyor 24. The first waiting time corresponds to the time obtained by subtracting a first processing delay from a first movement time required for the sample Sp to move from the time it passes through the sample detection unit 32 to the time it passes through the center of the field of view of the imaging unit 36. The first processing delay corresponds to the time required for processing from the detection of the sample Sp to the time it is imaged. The imaging unit 36 ​​includes a camera, and is installed with the optical axis of the camera oriented vertically and facing the center of the conveyor 24 in the width direction. The camera may be, for example, a digital still camera that captures still images, or a video camera that captures moving images. The captured image is expressed by the brightness value of each pixel arranged at regular intervals on a two-dimensional plane. The imaging unit 36 ​​outputs image data representing the captured image to the inspection device 10.

[0017] The inspection device 10 detects defects occurring in the sample Sp that appear in the captured image based on image data input from the imaging unit 36. When detecting defects in the sample Sp, the inspection device 10 outputs a discharge control signal to the sorting unit 38 to instruct the sample Sp to be discharged after a preset second waiting time has elapsed since the acquisition of the captured image data. The second waiting time corresponds to the time obtained by subtracting a second processing delay from a second movement time required for the sample Sp to move from the center of the field of view of the imaging unit 36 ​​to the front of the sorting unit 38. The second processing delay corresponds to the time required for the processes of imaging, defect detection, and discharge. Since the second processing delay needs to be shorter than the second movement time, a shorter processing time for defect detection based on the captured image is preferable.

[0018] The sorting unit 38 causes samples Sp in which defects have been detected to drop from the conveyor 24 based on a discharge control signal input from the inspection device 10. The sorting unit 38 passes samples Sp in which no defects have been detected to equipment that performs the subsequent process 26 without dropping them from the conveyor 24. The sorting unit 38 may include, for example, a rod and a pusher, and the pusher pushes the rod in a direction intersecting the direction of travel of the conveyor 24 in response to the input of the discharge control signal. The sorting unit 38 may include, for example, an air injector, and the air injector may inject compressed air in a direction intersecting the direction of travel of the conveyor 24 in response to the input of the discharge control signal.

[0019] <Inspection equipment> Next, an example of the functional configuration of the inspection device 10 according to this embodiment will be described. Fig. 2 is a schematic block diagram showing an example of the functional configuration of the inspection device 10 according to this embodiment. The inspection apparatus 10 according to this embodiment is an inspection apparatus for inspecting defects that may occur in a sample Sp, which is an object to be inspected, using image data that represents an image of the sample Sp. The inspection apparatus 10 acquires image data that represents a captured image of the sample Sp. The inspection apparatus 10 uses the captured image as an original image, converts the original image into intermediate values ​​with a lower degree of freedom, and restores a restored image from the intermediate values ​​obtained by the conversion. The inspection apparatus 10 detects defects that may occur in the object to be inspected based on a difference image between the original image and the restored image.

[0020] The inspection device 10 includes a control unit 110 , an input / output unit 140 , an operation unit 150 , a display unit 160 , and a storage unit 170 . The inspection device 10 may be configured as an information device having general-purpose hardware such as a personal computer (PC), a microcomputer, or a workstation, or may be configured as a dedicated device having dedicated hardware.

[0021] The control unit 110 executes processes for realizing the functions of the inspection device 10 and processes for controlling those functions. The control unit 110 may include general-purpose components such as a processor and be configured as a computer system. The processor reads a program previously stored in the storage unit 170 and performs processes instructed by instructions written in the read program to realize the functions. In this application, performing processes instructed by instructions written in a program may be referred to as executing a program or program execution. Part or all of the control unit 110 is not limited to general-purpose hardware such as a processor, but may also be configured to include dedicated hardware such as an LSI (Large Scale Integration) or an ASIC (Application Specific Integrated Circuit). Functional units that realize the functions of the control unit 110 will be described later.

[0022] The input / output unit 140 is connected to other devices wirelessly or via a wire to input and output various types of data. The input / output unit 140 includes, for example, an input / output interface or a communication interface. The input / output unit 140 is connected to, for example, various control devices, measuring devices, and other devices used in the manufacturing process.

[0023] The operation unit 150 receives a user operation and generates an operation signal corresponding to the received operation. The operation unit 150 may include dedicated components such as buttons, knobs, and dials, or general-purpose components such as a mouse and keyboard. The operation unit 150 may be an input interface that receives an operation signal wirelessly or via a wired connection from another device. The other device may be, for example, a remote controller, a portable device such as a multi-function mobile phone, or the like. The operation unit 150 outputs the acquired operation signal to the control unit 110.

[0024] Display unit 160 displays display information such as images, characters, and symbols based on display data input from control unit 110. Display unit 160 may include, for example, a liquid crystal display (LCD), an organic light emitting diode (OLED) display, or the like.

[0025] In addition to the above programs, the storage unit 170 stores various data used in the processes executed by the control unit 110 and various data acquired by the control unit 110. The storage unit 170 includes, for example, a non-volatile (non-temporary) storage medium such as a ROM (Read Only Memory), a flash memory, or an HDD (Hard Disk Drive). The storage unit 170 also includes a volatile storage medium such as a RAM (Random Access Memory), a register, or the like.

[0026] The control unit 110 includes a restoration unit 114, a detection unit 116, a determination unit 118, a process control unit 120, and a model learning unit 122 as functional units for realizing the functions of the inspection device 10. The restoration unit 114 receives image data representing a captured image from the imaging unit 36. Using a predetermined model, the restoration unit 114 converts the captured image represented by the image data into an intermediate value with a lower degree of freedom than the captured image. In this application, a model refers to a mathematical model, primarily a machine learning model, for performing arithmetic processing on an input value to derive an output value. The restoration unit 114 converts the converted intermediate value into a restored image that simulates the captured image. The restoration unit 114 is realized, for example, by using an autoencoder. The restoration unit 114 outputs restored image data representing the restored image obtained by the conversion to the detection unit 116.

[0027] The detection unit 116 receives image data from the imaging unit 36 ​​and receives restored image data from the restoration unit 114. The detection unit 116 subtracts the restored image indicated in the restored image data from the captured image indicated in the image data to generate a difference image. The detection unit 116 detects defects occurring in the sample Sp based on the generated difference image. The detection unit 116 can identify, for example, a region consisting of pixels in the difference image where the signal value of each pixel is significantly different from zero as a defect region. The detection unit 116 outputs defect detection data indicating the identified defect region to the determination unit 118.

[0028] The determination unit 118 determines whether the sample Sp is a good product or a defective product based on the defect detection data input from the detection unit 116. For example, the determination unit 118 can determine the sample Sp as a good product when the defect detection data does not indicate a defective area, and can determine the sample Sp as a defective product when the defect detection data indicates a defective area. The determination unit 118 outputs determination result data indicating the determination result to the process management unit 120.

[0029] The process control unit 120 refers to the determination result information input from the determination unit 118, and causes the samples Sp determined to be non-defective to proceed to the subsequent process 26. The process control unit 120 ejects the samples Sp determined to be defective, preventing them from proceeding to the subsequent process 26. The process control unit 120 outputs, for example, a discharge control signal instructing the sorting unit 38 to eject the samples Sp. When the discharge control signal is input from the process control unit 120, the sorting unit 38 ejects the samples Sp.

[0030] The model learning unit 122 learns a model used by the restoration unit 114 to convert a captured image into a restored image. The model learning unit 122 acquires a plurality of images of products that are determined to be non-defective in advance, and employs image data representing each of the acquired images as training data. In learning the model, the model learning unit 122 uses training images shown in the training data as input values, and learns a parameter set for the model so that a restored image showing an estimated value derived using the model is as close as possible to the training image. Model learning unit 122 stores the parameter set obtained by learning in storage unit 170. Restoration unit 114 reads the parameter set stored in storage unit 170 and sets the read parameter set in its own unit.

[0031] <Outline of inspection method> Next, an overview of the inspection method according to this embodiment will be described. Fig. 3 is a schematic block diagram for explaining the overview of the inspection device 10 according to this embodiment. The restoration unit 114 uses the trained model to convert the original image to be processed into an intermediate value with a lower degree of freedom, and then converts the intermediate value into a restored image. The degree of freedom refers to the number of independent variables. In other words, the intermediate value is expressed with less information than the original image, which is the input value. An image captured by the imaging unit 36 ​​can be input as the original image. An image representing an image of a sample having the characteristics of a non-defective product is obtained as the restored image.

[0032] The detection unit 116 includes a subtraction unit 116s and a differential image processing unit 116p. The subtraction unit 116s receives an original image and a restored image to be processed. The subtraction unit 116s outputs a difference image obtained by subtracting the restored image from the original image to a difference image processing unit 116p. Here, the subtraction unit 116s subtracts the luminance value representing the restored image from the luminance value representing the original image for each pixel. The difference image is represented by the absolute value of the difference value for each pixel obtained by the subtraction. Components representing the characteristics of a non-defective product have been subtracted from the difference image, leaving mainly components representing the characteristics of defects.

[0033] The differential image processing unit 116p detects regions where the signal value of each pixel representing the differential image input from the subtraction unit 116s is significantly different from zero as defective regions where defects appear. The differential image processing unit 116p, for example, performs binarization processing on the differential image to convert it into a binary image indicating the possibility of a defect. In the binarization processing, the differential image processing unit 116p sets the signal value of the binary image to 1 if the signal value of each pixel is equal to or greater than a predetermined binarization threshold, and sets the signal value of the binary image to 0 if the signal value is less than the predetermined binarization threshold.

[0034] Next, an example configuration of the restoration unit 114 will be described. In the example of FIG. 3, a convolutional auto-encoder (CAE) is applied as a model of the restoration unit 114. The CAE is a type of auto-encoder, and has an encoder to which a convolutional neural network (CNN) is applied. In general, an auto-encoder includes an encoder and a decoder. In the restoration unit 114, the encoder ec01 converts an original image into intermediate values. The decoder dc01 converts the intermediate values ​​into a restored image.

[0035] The encoder ec01 corresponds to a CNN. A CNN is a type of neural network and has at least one convolutional layer. Of the four layers provided in the encoder ec01, layers le01 to le03 are all convolutional layers. Layer le04 is a flattening layer. The decoder dc01 is composed of a neural network that performs approximately the inverse operation of the encoder ec01. In the decoder dc01, layers ld01 to ld03 are de-convolutional layers. Layers ld01 to ld03 correspond to layers le01 to le03 of the encoder ec01, respectively. Layer ld04 is a fully-connected layer.

[0036] Each layer has one or more nodes. Each layer usually has multiple nodes. The nodes are also called nodes or neurons. The nodes calculate input values ​​according to a predetermined function, and output the function value obtained by the calculation as the output value. A convolutional layer has a node for each kernel. A kernel is a set of input values ​​that form a processing unit for calculating one output value at a time. For example, in layer le01, the brightness values ​​of multiple pixels that form part of the original image are assigned as input values ​​to each kernel.

[0037] Each node in a convolutional layer performs a convolution operation on multiple input values ​​belonging to its kernel to calculate a convolution value. The calculated convolution value is then added to a bias value to obtain a correction value, and the resulting function value is then calculated as an output value and output to the next layer. Examples of activation functions that can be used include a rectified linear unit (rectified linear unit) and a sigmoid function. A rectified linear unit is a function that sets a threshold (e.g., 0) as the output value for input values ​​below that threshold, and outputs input values ​​exceeding the threshold as is. In a convolution operation, the sum of the multiplied values ​​obtained by multiplying each input value by an independent convolution coefficient is calculated as the convolutional value. Therefore, the convolution coefficients, bias values, and activation function parameters form part of the parameter set. Layers le01 to le03 each output a value for each sample arranged in three-dimensional space. The number of output value samples decreases with each convolutional layer.

[0038] The flattening layer is a layer that expands multidimensional sample values ​​of two or more dimensions as output values ​​of samples arranged in a lower-dimensional space. Layer le04 outputs the three-dimensional input values ​​input from layer le03 as output values ​​of samples arranged in a one-dimensional array, i.e., as a vector. Layer le04 may also select some output value samples as targets for output from encoder ec01. The samples to be output correspond to the degrees of freedom of the intermediate values ​​output from encoder ec01 and may be preset to be less than the number of samples of the output values ​​from layer le03. In this case, the samples to be output may be fixed to preset samples or may be changeable by being trained as part of a parameter set.

[0039] The fully connected layer is a layer that performs convolution operations on input values ​​input from the previous layer to each of multiple nodes to calculate convolution values, calculates the calculated value by adding the calculated convolution value and a bias value as the output value, and outputs the calculated output value to the next layer. Layer ld04 receives the output value from encoder ec01 as the previous layer as an input value. For samples that are not to be output, each node in layer ld04 considers the output value to be zero.

[0040] The deconvolution layer expands the data size of the input values ​​from the previous layer (increases the number of samples) and performs convolution operations on the expanded input values. The operation in the deconvolution layer is also called transposed convolution. The number of samples of the output values ​​increases with each deconvolution layer added. The number of samples of the input values ​​in layers ld03 to ld01 of the decoder dc01 is set to be equal to the number of samples of the output values ​​from layers le03 to le01 of the encoder ec01. The output value from each node of layer ld01 is the brightness value of the corresponding pixel that represents the restored image.

[0041] A deconvolutional layer consists of nodes that receive output values ​​from the previous layer as input values ​​and nodes that receive a fixed value (e.g., zero) as input values, which are spatially repeated at a fixed period. Similar to a convolutional layer, each node performs a convolution operation on multiple input values ​​belonging to its kernel to calculate a convolution value, and outputs the resulting function value as an output value to the next layer by calculating an activation function on the corrected value obtained by adding the calculated convolution value and a bias value. In a deconvolutional layer, the kernel size is set to be equal to or greater than the node repetition period. The convolution coefficients, bias values, and activation function parameters form part of the parameter set. When input value samples are arrayed multidimensionally, the kernel size for each dimension can be set to be equal to or greater than the node repetition period.

[0042] Next, we will explain the learning of a model used to infer a restored image from an original image in the restoration unit 114. Fig. 4 is a diagram for explaining the model according to this embodiment. Here, we will take as an example a case where an autoencoder is used as a model used for conversion to a restored image. In model training, the model training unit 122 searches for a parameter set of the model that minimizes the magnitude of the difference between an estimated value derived using the model based on training images shown in the training data as input values ​​and the training images used as input values. When searching for a parameter set of the model, the model training unit 122 recursively updates the parameter set so as to reduce an index value of the magnitude of the difference. The model training unit 122 repeats updating the parameter set until the change in the parameter set converges. The model training unit 122 can determine that convergence has occurred when the amount of change in the parameter set before and after the update, or the amount of change in the magnitude of the difference between before and after the update, is less than a predetermined convergence determination threshold. The restoration unit 114 sets the parameter set obtained by training to itself and uses the set parameter set to convert the original image to a restored image.

[0043] In updating the parameter set, for example, a method such as steepest descent, stochastic gradient descent, or conjugate gradient method can be used. As an index value of the magnitude of the difference, for example, a loss function such as the sum of squared differences (SSD), the sum of absolute differences (SAD), or the cross entropy error can be used.

[0044] Next, specific examples of defects that may occur in the sample Sp are described. Figure 5 shows an example of an image of the sample Sp captured while it was being transported on the conveyor 24. The sample Sp is approximately frustum-shaped. The higher the intensity of the transmitted light from the light source 34, the brighter the image. Two vertically extending dark areas in the center of the image represent the conveyor belt. The bright area surrounded by the outer edge of the sample Sp is the area where the laminated glass is placed. A shielding layer, which is an opaque colored ceramic layer, is provided on the edge of the laminated glass. The shielding layer is represented by a dark area that is darker than its surroundings. The sample shown in Figure 5 was obtained through pre-compression bonding using the nipper roll method. (a) shows a good product, and (b) shows a defective product. The area darker than the center near the upper left corner of the defective product is the defective area. This defect is a gap caused by insufficient compression between the glass layer and the interlayer during pre-compression bonding. This defect may occur in a region closer to the shielding layer than the center. A good product will not show such defective areas.

[0045] 6 shows, from left to right, a captured image representing a non-defective product, a restored image converted from the captured image, and a difference image. The difference image shown in FIG. 6 is dark almost entirely. This confirms that the signal values ​​of almost all pixels are not significantly different from zero, and that the restoration unit 114 has reproduced a captured image representing a non-defective product. FIG. 7 shows, from left to right, a captured image representing a defective product, a restored image converted from the captured image, and a difference image. Unlike the captured image, the restored image shown in FIG. 7 does not show any defective areas. This indicates that the restoration unit 114 has reproduced the characteristics of the captured image representing a non-defective product. In the difference image, the defective area appears brighter than its surroundings. That is, the significant difference from zero in the signal value of the pixel included in the defective area confirms that the defect can be easily identified based on the difference image.

[0046] Next, an example of the difference image processing according to this embodiment will be described below. Fig. 8 is a flowchart illustrating the difference image processing according to this embodiment. (Step S112) The differential image processing unit 116p (see FIG. 3) binarizes the captured image used as the original image to generate a brightness reference image, which is a binary image for determining the inspection area. The binary image is expressed as one of two values, 1 or 0, as the signal value for each pixel. The differential image processing unit 116p sets the signal value of a pixel whose brightness value is less than a predetermined brightness reference value to 0, and sets the signal value of a pixel whose brightness value is equal to or greater than the brightness reference value to 1. In this application, a pixel whose signal value is 0 may be referred to as a "dark area," and a pixel whose signal value is 1 may be referred to as a "light area." The differential image processing unit 116p detects a dark area that is spatially continuous and surrounds one or more light areas, and is further surrounded by other light areas, as the edge of the sample Sp (see FIG. 9).

[0047] (Step S114) The differential image processing unit 116p identifies the part of the brightness reference image that surrounds the detected edge as the background (see FIG. 10). (Step S116) The differential image processing unit 116p performs blob processing on the sample portion of the brightness reference image, excluding the background, and identifies a partial region as the inspection area. For each spatially continuous region (blob) constituting the sample portion, the differential image processing unit 116p determines, as the inspection area, a region that meets one or both of predetermined criteria of size and position. In the example of FIG. 11, the differential image processing unit 116p determines, as the inspection area, the region closest to the left end of the sample portion separated by the edge and the conveyor. This is because defects due to insufficient bonding may occur in this area. Note that the differential image processing unit 116p may also determine, as the inspection area, the region closest to the center or the region closest to the right end of the sample portion separated by the edge and the conveyor, depending on factors such as the structure of the object to be inspected and the type of defect to be inspected.

[0048] (Step S118) The differential image processing unit 116p binarizes the differential image within the determined inspection range to generate a binary image for defect inspection (see FIG. 12). The differential image processing unit 116p determines a portion where the signal value of each pixel is less than a predetermined binarization threshold as a dark portion, and a portion where the signal value is equal to or greater than the binarization threshold as a bright portion. (Step S120) The difference image processing unit 116p smooths the boundaries between bright and dark areas in the binary image. The difference image processing unit 116p, for example, performs noise processing on the bright areas. The noise processing includes an expansion step that changes dark areas adjacent to bright areas to bright areas, and a reduction step that changes bright areas adjacent to dark areas to dark areas. The noise processing may be repeated two or more times. As a result, spatially continuous bright areas smaller than a certain size are removed, and the boundaries between bright and dark areas (see FIG. 13) are smoothed. (Step S120) The difference image processing unit 116p identifies, as a defective area, an area having an area equal to or larger than a predetermined reference area among areas including spatially continuous bright areas. In the example of FIG. 14, the bright area that finally appears in the upper left corner is identified as the defective area. Thereafter, the processing of FIG. 8 ends. Note that, among areas including spatially continuous bright areas, the difference image processing unit 116p may identify, as a defective area, an area having a length equal to or larger than a predetermined reference length in the vertical direction of the image, or an area having a width equal to or larger than a predetermined reference width in the horizontal direction.

[0049] Next, an example of defective area detection will be described. FIG. 15 is a diagram showing a first example of defective area detection. FIG. 15 shows the area of ​​the defective area detected for each sample as the detected area (unit: number of pixels), along with its pass / fail judgment. However, for samples in which no defective areas were detected, the detected area was set to zero. Furthermore, for samples in which multiple defective areas were detected, the maximum of those areas was set to the detected area. For all samples judged to be good products, the detected area was almost zero. For all samples judged to be bad products, the detected area exceeded the reference area of ​​20 pixels.

[0050] Figure 16 is a diagram showing a second example of defective area detection, and is a histogram showing the frequency for each detected area. For samples judged to be good, the detected area was below the reference area of ​​20. For samples judged to be defective, the detected area was above the reference area of ​​20. The number 4785 in Figure 16 indicates the frequency of good products with a detected area of ​​0 or more and less than 10. 15 and 16 show that good and bad products can be clearly distinguished based on whether or not they have a defect area whose area is equal to or greater than the reference area. In these examples, no good products were detected as bad (overdetection).

[0051] Note that when restoring a restored image from an original image representing a certain variety of specimen, the restoration unit 114 does not necessarily need to use a model trained using training images representing specimens of that variety but not training images representing specimens of other varieties. The restoration unit 114 may also use a model trained using training images representing specimens of other varieties that share the same model type as the original variety. That is, the restoration unit 114 may convert an original image representing a certain variety of specimen into a restored image using a model obtained by training using training data including training images representing specimens of multiple varieties that share the same model type. Here, the term "variety" refers to differences in the detailed external characteristics of the specimen, such as the type of material and attached components. In the example of a windshield shown in Figures 5 and 6, differences in the shape or size of the components appearing at the base depend on the specimen type. The term "type" refers to differences in the overall external characteristics of the specimen, such as the shape and size of the specimen.

[0052] Furthermore, when restoring a restored image from an original image representing a sample of a certain type, the restoration unit 114 does not necessarily need to use a model trained using training images representing samples of that type and not using training images representing samples of other types. The restoration unit 114 may use a model trained using training images representing samples of multiple types including that type. In other words, the restoration unit 114 may convert an original image representing a sample of that type into a restored image using a model obtained by training using training data including training images representing samples of multiple types including that type.

[0053] Figure 17 shows the overdetection rate for each model. The model column indicates the type of training image used for model training. By type refers to the case where a model trained for each type was used to convert an original image of a sample of that type into a restored image. By type refers to the case where a model trained for each type was used to convert an original image of a sample of that type into a restored image. Across multiple types refers to the case where a model trained across multiple types (eight types) was used to convert an original image of a sample of any of those multiple types into a restored image. In all cases, 5,000 samples were used for inspection. Figure 17 shows that the overdetection rate was 0% for all types, type, and across multiple types. This indicates that even when a pre-trained model is used for inspecting multiple types of products, defect detection can be achieved for those multiple types without degrading performance. This contributes to reducing inspection costs by allowing a single trained model to be used for inspecting multiple types or types of products.

[0054] Next, an example of the defect detection process according to this embodiment will be described with reference to a flowchart shown in FIG. (Step S142) The sample detector 32 monitors whether or not the sample Sp being transported is present. If the sample Sp is detected (YES in step S142), the process proceeds to step S144. If the sample Sp is not detected (NO in step S142), the process of step S142 is repeated. (Step S144) The imaging unit 36 ​​captures an image of the sample Sp.

[0055] (Step S146) The restoration unit 114 of the inspection device 10 acquires a captured image from the imaging unit 36. The restoration unit 114 uses the acquired captured image as an original image, converts it into an intermediate value with a lower degree of freedom using a trained model, and converts the converted intermediate value into a restored image. (Step S148) The detection unit 116 generates a difference image by subtracting the restored image from the original image, and performs difference image processing (FIG. 8) on the difference image to detect defective areas.

[0056] (Step S150) The determination unit 118 determines whether the sample Sp is acceptable or not based on whether it has a defective area whose area is equal to or larger than a predetermined reference area. If the sample Sp is determined to be non-defective (accepted) (YES in step S150), the process proceeds to step S152. If the sample Sp is determined to be defective (rejected) (NO in step S150), the process proceeds to step S154. (Step S152) The process control unit 120 advances the sample Sp toward the subsequent process 26. After that, the process of FIG. 18 ends. (Step S154) The process control unit 120 causes the sorting unit 38 to discharge the samples Sp. After that, the process of FIG. 18 ends.

[0057] Next, other examples of defects will be described. The nature of defects can vary depending on the size and shape of the intermediate product as well as the processing performed in the previous process 22. For example, pre-crimping using a bag member may be performed instead of using nipper rolls. The bag member is made of a material that is heat-resistant, plastic, and elastic. For example, materials such as silicone rubber and fluororubber are used for the bag member. Pre-crimping using a bag member includes the following steps s1 to s3: (s1) An unbonded body consisting of multiple glass layers and an intermediate layer sandwiched between them is carried into the bag member. (s2) Air remaining in the bag member is removed, allowing the bag member to press the unbonded body from the surrounding environment. (s3) The pressed unbonded body is heated. Pre-crimping using a bag member can result in defects in which gaps form between the glass layers and / or the intermediate layer due to distortion of the glass layers and / or the intermediate layer.

[0058] FIG. 19 shows examples of captured images of a sample obtained by preliminary crimping using a bag member. (a) A good product and (b) a defective product are shown. The defect shown in FIG. 19(b) has a vertex at one end of the edge, but it is wider than the insufficient crimping defect shown in FIG. 5(b), and the decrease in brightness from the surrounding area tends to be more gradual. Therefore, to detect the defect shown in FIG. 19(b), a smaller binarization threshold and a larger reference area can be set for the difference image processing unit 116p, and the difference image processing shown in FIG. 8 can be performed.

[0059] Therefore, for samples that may have multiple types of defects, a binarization threshold and a reference area are set in advance for each type of defect in the differential image processing unit 116p of the detection unit 116. The differential image processing unit 116p performs differential image processing (FIG. 8) using the binarization threshold and reference area set in advance for each type of defect to attempt to detect a defective area. The determination unit 118 then determines a sample in which even one type of defective area is detected as a defective product, and determines a sample in which no defective area is detected as a non-defective product.

[0060] For example, for two types of defects A and B, a binarization threshold x A , x B and reference area S A , S B are set in advance. The detection unit 116 and the determination unit 118 may perform a pass / fail determination process, which will be described next, instead of the processes of steps S148 and S150 of the defect detection process (FIG. 18).

[0061] FIG. 20 is a flowchart illustrating the pass / fail determination process according to this embodiment. (Step S148a) The detection unit 116 applies a binarization threshold x related to the defect A to a difference image obtained by subtracting the restored image from the original image. A and reference area S A The defect area is detected by differential image processing (Fig. 8) using (Step S150a) The determination unit 118 determines whether the area is the reference area S A Whether or not defect A has been detected in the sample Sp is determined based on whether or not there is a defective area with the above value. If defect A is detected (YES in step S150a), the sample Sp is determined to be a defective product (rejected), and the processing in FIG. 20 is terminated. When it is determined that the defect A is not detected in the sample Sp (NO in step S150a), the process proceeds to step S148b.

[0062] (Step S148b) The detection unit 116 applies a binarization threshold x related to the defect B to the difference image obtained by subtracting the restored image from the original image. B and reference area SB The defect area is detected by differential image processing (Fig. 8) using (Step S150b) The determination unit 118 determines whether the area is the reference area S B Whether or not a defect B has been detected in the sample Sp is determined based on whether or not there is a defective area with the above value. If a defect B is detected (YES in step S150b), the sample Sp is determined to be a defective product (rejected), and the processing in FIG. 20 is terminated. When it is determined that no defect B is detected in the sample Sp (step S150b NO), the sample Sp is determined to be non-defective (accepted), and the processing of FIG. 20 is terminated.

[0063] 20 may be performed on each of one type of intermediate product as a sample that may have multiple types of defects, or may be performed on each of a group of multiple types of intermediate products. The group of multiple types of intermediate products may be multiple types belonging to one model, or multiple types belonging to multiple models. Therefore, using differential images acquired by a common method regardless of the type of defect, processing to detect defective areas is sequentially performed using different parameters for each type of defect, and a final pass / fail determination is made. Furthermore, although the types of defects that may occur may vary, inspections for multiple types of samples can be consolidated into a single inspection process. This reduces the cost associated with the inspection.

[0064] Next, an example of detecting a defect area for each defect type will be described. Fig. 21 is a histogram showing an example of detecting a defect area related to defect A. Fig. 21 shows the area of ​​the defect area of ​​defect A detected in 5005 samples as the detected area, along with its pass / fail judgment. Note that the signal value for each pixel is expressed as an 8-bit integer value, and the binarization threshold x A is set to 180, and the reference area S A was set to 180. As a result, for the samples judged to be non-defective, the detected area was all less than the reference area S A For samples judged to be defective, the detected area was below the reference area S A exceeded.

[0065] Fig. 22 is a histogram showing an example of the detection of a defect area related to defect B. Fig. 22 shows an example of the area of ​​the defect area of ​​defect B detected in 5005 samples as the detected area, along with the pass / fail judgment. However, the binarization threshold x B is set to 140, and the reference area S B As a result, for the samples judged to be non-defective, the detected area was all less than the reference area S B For samples judged to be defective, the detected area was below the reference area S B exceeded. These detection results show that overdetection does not occur due to the use of a common difference image for defects A and B, and the use of different binarization thresholds and reference areas for each.

[0066] In the inspection system S1, the position or orientation of the sample Sp placed on the conveyor 24 may vary. The outer edge of the image of the sample Sp tends to have a significant difference in brightness from the surrounding background. The inclusion of components in the difference image resulting from differences in brightness distribution due to differences in the position or orientation of each sample Sp can reduce the accuracy of defect detection. Therefore, it is desirable for the model learning unit 122 to learn a model for conversion to a restored image using training data representing a large number of training images representing samples with different positions and / or orientations. However, capturing images of a large number of different samples requires a great deal of effort and time.

[0067] Therefore, the model learning unit 122 acquires a background image representing the background and a sample image representing the sample. For example, the model learning unit 122 performs a known image recognition process on the captured image to separate the portion representing the sample image as a sample image and the background as a background image. FIG. 23 is a captured image illustrating a sample Sp placed on a conveyor 24. FIG. 24 is a sample image extracted from the captured image shown in FIG. 23. In the sample image illustrated in FIG. 24, a known image processing model is used to erase the image of the conveyor belt appearing in the background and synthesize the portion of the sample that is obscured by the conveyor belt. The model learning unit 122 may cause the imaging unit 36 ​​to capture in advance a sample image showing the sample but not the background, and a background image showing the background but not the sample, separately.

[0068] The model learning unit 122 generates training images for each sample by superimposing a sample image on a background image in different positions. Each position is defined by one or both of a displacement from a reference position (e.g., the center of gravity of the training image) and a rotation angle from a reference direction (e.g., the horizontal direction). The model learning unit 122 randomly determines a real number selected from a predetermined range of displacement amounts as a displacement amount for each training image using pseudorandom numbers. The model learning unit 122 positions the sample image at a position displaced from the reference position by the determined displacement amount. The model learning unit 122 randomly determines a real number selected from a predetermined range of rotation angles as a displacement amount for each training image using pseudorandom numbers. The model learning unit 122 positions the sample image in a direction rotated from the reference position by the determined rotation angle. Figure 25 illustrates an example of a training image synthesized by superimposing the sample image shown in Figure 24 on a background image. In the example of Figure 25, the sample image is displaced to the right compared to the captured image illustrated in Figure 23. This reduces the effort and time required to acquire training data representing samples with various positions.

[0069] The range of the displacement amount may be set in advance based on the relationship between the shape and size of the sample image and the size of the training image, within a range in which the image of the sample after displacement does not extend beyond the training image. The range of the rotation amount may also be set in advance based on the relationship between the shape and size of the sample image and the size of the training image, so long as the rotated sample image does not extend beyond the training image.

[0070] Next, an example of the training data generation process according to this embodiment will be described. Fig. 26 is a flowchart illustrating the training data generation process according to this embodiment. Fig. 26 shows an example of generating multiple training images by arranging a sample image representing a displaced sample against a background image representing the background that appears in each frame of captured image. (Step S162) The model learning unit 122 removes the background from the captured image shown in the image data acquired in advance, and extracts the sample image.

[0071] The model learning unit 122 repeats the processing of steps S164 and S164 N times (N is a predetermined integer of 2 or more). (Step S164) The model learning unit 122 randomly determines a displacement amount from a predetermined range of displacement amounts, and randomly determines a rotation angle from a predetermined range of rotation angles. (Step S166) The model learning unit 122 arranges the sample image in the background image according to the determined displacement amount and rotation angle, and generates training images.

[0072] The restoration unit 114 described above is provided with an autoencoder having an encoder and a decoder as a model for converting an original image into a restored image, and a CNN is applied to the encoder, and a model that approximates the inverse processing performed by the encoder is applied to the decoder. The number of layers of the CNN is not limited to that shown in FIG. 3. The number of nodes and kernel size of each layer can be set arbitrarily as long as the number of elements of the output value can be made smaller than the number of pixels of the original image, which corresponds to the number of elements of the input value, for the entire encoder.

[0073] Furthermore, in the encoder, a neural network other than a CNN, for example, a recurrent neural network (RNN), may be applied. The applicable machine learning model is not limited to a neural network, but may be a random forest (RF), a support vector machine (SVM), or the like. Furthermore, the model for converting an original image into a restored image does not necessarily have to be a machine learning model. For example, principal component analysis (PCA) may be used. PCA involves expressing an input value using multiple principal components and deriving the contribution of each principal component as an intermediate value. The degrees of freedom of the intermediate value are lower than the degrees of freedom of the input value. An estimated value that approximates the input value is restored using the derived intermediate value and principal components.

[0074] Furthermore, in the above description, the present embodiment is primarily applied to the inspection of defects occurring in an intermediate product, such as a windshield, which is a final product having laminated glass. However, the present invention is not limited to this. The final product may also be a window glass other than a windshield, for example, a rear window. The sample to be inspected is not limited to an intermediate product, but may also be a final product or a material used in the manufacture of a product. Furthermore, the sample is not limited to a product containing laminated glass, but may be any object in which the presence or absence of a defect causes a significant difference in luminance distribution. The sample may be, for example, an object made of synthetic resin, a metal surface, a semiconductor crystal, or the like. Such an object may be, for example, a display panel, a circuit board, a semiconductor wafer, or the like. In the above description, the processing target is primarily a monochrome original image showing a distribution of shades, but this is not limiting. The original image may also be a color image. It is sufficient that color signal values ​​are applied instead of the luminance values ​​of each pixel of the monochrome image.

[0075] Furthermore, the inspection device 10 does not necessarily have to include the model learning unit 122. It is sufficient that another device separate from the inspection device 10 performs model learning, and the parameter set obtained by the learning is set in the restoration unit 114. The inspection device 10 may be connected to the production line wirelessly or via a wire, and may perform the processing shown in Fig. 18 using image data acquired from the production line. The production line may include a conveyor 24, lighting 34, an imaging unit 36, and a sorting unit 38, or some of these may be omitted. The inspection device 10 may operate independently without being configured as part of the inspection system S1. In this case, the process control unit 120 and the model learning unit 122 may be omitted from the inspection device 10.

[0076] As described above, the inspection device 10 according to this embodiment is an inspection device 10 that uses an original image (e.g., a captured image) to inspect defects occurring in an object to be inspected (e.g., a sample Sp), and is equipped with a restoration unit 114 that converts the original image into an intermediate value with a lower degree of freedom and converts it into a restored image restored from the intermediate value, and a detection unit 116 that detects defects occurring in the object to be inspected based on a difference image between the original image and the restored image. According to this configuration, the characteristics of the object under inspection are represented by intermediate values ​​that have a lower degree of freedom than the original image, and the characteristics of the object under inspection are represented by a restored image reconstructed from the intermediate values, so that in the difference image between the original image and the restored image, abnormal components that cannot explain the characteristics of the object under inspection in the original image are mainly extracted. Therefore, defects occurring in the object under inspection can be accurately detected using the difference image.

[0077] The detection unit 116 may detect defects based on the size of an area where the signal value of the differential image is equal to or greater than a predetermined reference value (for example, a binarization threshold value). This allows an area including pixels whose signal values ​​are equal to or greater than the reference value to be simply detected as a defective area.

[0078] The object under inspection may be an object (e.g., laminated glass) having an opaque shielding layer formed around its periphery and multiple transparent layers (e.g., glass layers, intermediate layers) superimposed on one another. The detection unit 116 may determine at least a portion of the area surrounded by the shielding layers in the original image as the defect inspection range. This limits the area where defects may occur as the inspection range, making it possible to reduce the amount of processing required for defect detection compared to when the inspection range is not defined.

[0079] The defects may be air bubbles or gaps between the layers of material. This allows accurate detection of defects based on significant changes in signal value caused by bubbles or voids.

[0080] The restoration unit 114 may include an autoencoder including an encoder (for example, ec01) that converts an original image into an intermediate value, and a decoder (for example, dc01) that converts the intermediate value into a restored image. As a result, the original image is converted into an intermediate value that indicates the contribution of the principal component that shows its characteristics, and the intermediate value is converted into a restored image that shows its characteristics. By learning in advance so that the restored image approximates the original image without defects, a restored image that has the characteristics of an object to be inspected without defects can be obtained.

[0081] The encoder of the restoration unit 114 may be a neural network (for example, a CNN) having one or more convolutional layers, and the decoder may be a neural network having one or more deconvolutional layers. This allows the characteristics of the original image that constitutes the input value to be expressed as an intermediate value with a degree of freedom significantly lower than that of the input value.

[0082] The inspection device 10 may also include a model learning unit 122 that acquires training data including a plurality of training images in which the image of the object to be inspected appears in different arrangements, and learns a model that converts the original image into the restored image based on the acquired training data. Here, the model learning unit 122 may acquire training data including training images in which images of the object to be inspected are arranged at different positions or orientations in a background image. With this configuration, training data showing training images of test objects with different arrangements is used to learn a model that converts an original image into a restored image. Therefore, training images showing test objects with different arrangements can be obtained without taking images of a large number of test objects. Furthermore, because training images showing test objects with different arrangements are used for learning, it is possible to suppress a decrease in judgment accuracy due to differences in the arrangement of the test objects.

[0083] As described above, the inspection device 10 executes an inspection method including a restoration step of converting an original image into an intermediate value with a lower degree of freedom and converting it into a restored image restored from the intermediate value, and a detection step of detecting defects occurring in the inspected object based on a difference image between the original image and the restored image. The inspection system S1 may be configured to perform a method for producing an object, which includes a step of sandwiching an intermediate layer between multiple glass layers to form a laminated glass as an object to be inspected, the above-mentioned inspection method, and a step of rejecting an object (e.g., a sample Sp) having a laminated glass in which a defect is detected. With this configuration, the characteristics of the object under inspection are represented by intermediate values ​​with a lower degree of freedom than the original image, and the characteristics of the object under inspection are represented by a restored image reconstructed from the intermediate values. Therefore, abnormal components that cannot explain the characteristics of the object under inspection in the original image are mainly extracted in the difference image between the original image and the restored image. Since defects occurring in the object under inspection can be accurately detected using the difference image, it is possible to prevent the shipment of objects with laminated glass that is free of defects.

[0084] Furthermore, part or all of the inspection device 10 in the above-described embodiment may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each functional block of the inspection device 10 may be individually implemented as a processor, or part or all of them may be integrated into a processor. Furthermore, the integrated circuit implementation method is not limited to LSI, and may be implemented using a dedicated circuit or a general-purpose processor. Furthermore, if an integrated circuit implementation technology that can replace LSI emerges due to advances in semiconductor technology, an integrated circuit based on that technology may be used. [Explanation of symbols]

[0085] S1...inspection system, 10...inspection device, 24...conveyor, 32...sample detection unit, 34...illumination, 36...imaging unit, 38...sorting unit, 110...control unit, 114...restoration unit, 116...detection unit, 116s...subtraction unit, 116p...differential image processing unit, 118...determination unit, 120...process control unit, 122...model learning unit, 140...input / output unit, 150...operation unit, 160...display unit, 170...storage unit

Claims

1. An inspection apparatus for inspecting defects occurring in an object to be inspected using an original image, comprising: a restoration unit that converts the original image into an intermediate value having a lower degree of freedom and converts the intermediate value into a restored image; a detection unit that detects defects occurring in the object to be inspected based on a difference image between the original image and the restored image. Inspection equipment.

2. The detection unit The defect is detected based on the size of an area where the signal value of the difference image is equal to or greater than a predetermined reference value. The inspection device according to claim 1 .

3. The object to be inspected is formed by laminating a plurality of transparent layers on a peripheral edge of the object with an opaque shielding layer, The detection unit At least a part of the area surrounded by the shielding layer shown in the original image is defined as an inspection range for the defect. The inspection device according to claim 2 .

4. The defects are air bubbles or gaps between the layers. The inspection device according to claim 3 .

5. The restoration unit is an encoder for converting the original image into the intermediate value; a decoder for converting the intermediate values ​​into the reconstructed image; and an autoencoder comprising The inspection device according to claim 1 .

6. the encoder is a neural network having one or more convolutional layers; The decoder is a neural network with one or more deconvolutional layers. The inspection device according to claim 5 .

7. acquiring training data including a plurality of training images in which the image of the object to be inspected appears in different arrangements; a model learning unit that learns a model for converting the original image into the restored image based on the training data. The inspection device according to claim 1 .

8. The model learning unit Acquire training data including training images in which images of the object to be inspected are arranged at different positions or orientations on a background image. The inspection device according to claim 7.

9. To the computer A program for causing the inspection device according to claim 1 to function.

10. An inspection method for inspecting defects occurring in an object to be inspected using an original image, comprising: Transforming the original image into an intermediate value with a lower degree of freedom, and then transforming the intermediate value into a restored image; Detecting defects occurring in the object to be inspected based on a difference image between the original image and the restored image Testing method.

11. an intermediate layer is sandwiched between a plurality of glass layers to form a laminated glass as the test object; The inspection method according to claim 10; Rejecting an object having the laminated glass in which a defect is detected. Production method.

Images