Solder Inspection Equipment

Abstract

To provide a solder inspection device that can reduce the effort and burden required to obtain an identification means as an AI model, and that can share the identification means even when the sizes of lands are different. [Solution] A predetermined neural network is trained using only image data related to good cream solder as training data to generate an AI model 101, which then receives test image data based on image data acquired by a camera 32d to obtain reconstructed image data, and the quality of the cream solder is determined by comparing the test image data and the reconstructed image data. The training data and test image data are formed by dividing the solder area image into two or four parts, and each part of the divided solder image is set in an image frame, and the image frame related to the training data and the image frame related to the test image data are the same size.

Description

[Technical Field]

[0001] The present invention relates to a solder inspection device for inspecting cream solder applied to a circuit board. [Background technology]

[0002] Generally, in a board manufacturing line where electronic components are mounted on a printed circuit board, cream solder is first printed on the lands of the printed circuit board (solder printing process). Next, the electronic components are temporarily attached to the printed circuit board using the viscosity of the cream solder (mounting process). After that, the printed circuit board is introduced into a reflow furnace, where the cream solder is heated and melted to perform soldering (reflow process). Such board manufacturing lines may be equipped with an inspection device that inspects the printed circuit board.

[0003] Recently, inspection devices that use AI models have been proposed for inspecting printed circuit boards. For example, a known inspection device that uses an AI model is one that inspects the quality of the shape of the upper part of cream solder on a printed circuit board by comparing inspection image data (original image data) of an inspection area on the printed circuit board with reconstructed image data generated by inputting the inspection image data into an AI model (identification means) (see, for example, Patent Document 1, etc.). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-61546 Summary of the Invention [Problem to be solved by the invention]

[0005] Incidentally, to obtain an AI model, it is necessary to use image data relating to cream solder as training data for training. However, since the shape of each individual cream solder is different, if image data relating to the entire cream solder is used as training data, a large amount of image data must be prepared to achieve the desired accuracy, making efficient training difficult.

[0006] Furthermore, in terms of improving inspection accuracy, it is preferable to match the size (width and height) of the training data with the size of the test image data. However, matching the sizes of the training data and test image data by enlarging or reducing the image data may result in a decrease in inspection accuracy. To address this issue, it is possible to prepare various training data of different sizes corresponding to the size of the lands, and to prepare multiple AI models trained only on training data of the same size. Then, during inspection, test image data of a size corresponding to the size of the lands is acquired, and an appropriate AI model that matches the size of the lands (i.e., the test image data) is selected from the multiple AI models and used. However, preparing different AI models for each size of land (test image data) requires complex work and requires a great deal of effort. Furthermore, obtaining multiple AI models requires the preparation of a huge amount of training data.

[0007] The present invention has been made in consideration of the above circumstances, and its purpose is to provide a solder inspection device that can reduce the effort and burden involved in obtaining an identification means as an AI model, and that can use the identification means in common even when the land sizes are different. [Means for solving the problem]

[0008] The following describes each of the means suitable for achieving the above object, with specific effects of the corresponding means added as necessary.

[0009] Means 1. A solder inspection device for inspecting cream solder printed on a printed circuit board, an image data acquisition means for acquiring image data of a predetermined inspection area on the printed circuit board including the printed solder paste; a recognition means for generating the image data by training only image data relating to non-defective cream solders in a neural network having an encoding unit for extracting feature values ​​from input image data and a decoding unit for reconstructing image data from the feature values; an inspection image data acquisition means for acquiring inspection image data including an image of the cream solder to be inspected based on the image data acquired by the image data acquisition means; a reconstructed image data acquisition means for inputting the inspection image data to the identification means and acquiring reconstructed image data as reconstructed image data; a comparison means for comparing the inspection image data and the reconstructed image data; The method is configured to be able to determine whether the cream solder is good or bad based on the comparison result by the comparison means, The learning data is composed of four divided solder images obtained by dividing one solder area image showing cream solder corresponding to one land into four parts by imaginary cross lines that can be divided into four parts of the same shape, and each divided solder image is provided in an image frame that is larger than the size of the divided solder image, The solder inspection device is characterized in that the inspection image data acquisition means acquires the inspection image data in which the solder area image in the image data acquired by the image data acquisition means is divided into four parts by the virtual cross lines, and each of the four divided solder images is placed in an image frame of the same size as the image frame of the learning data.

[0010] The learning data may be generated from image data (actual image data) obtained by capturing an image of a printed circuit board on which good quality cream solder has been printed (i.e., an image relating to the actual cream solder), or may be a virtually generated image relating to good quality cream solder. Examples of the actual image data include image data accumulated in previous inspections and image data of good quality printed circuit boards visually selected by an operator after the cream solder has been printed (the same applies to Means 5 described below).

[0011] Furthermore, the "neural network" includes, for example, a convolutional neural network having multiple convolutional layers. The "learning" includes, for example, deep learning. The "identification means (generative model)" includes, for example, an autoencoder and a convolutional autoencoder (the same applies to Means 5 described below).

[0012] In addition, the "identification means" is generated by learning only image data related to non-defective cream solder. Therefore, the reconstructed image data generated when inspection image data related to defective cream solder is input to the identification means will be approximately identical to the inspection image data in which the defective parts have been corrected (for example, foreign matter has been removed, or the shape and size have been corrected). In other words, when there is a defect in the cream solder, virtual image data related to the cream solder, assuming that there is no defect, is generated as the reconstructed image data related to the cream solder (the same applies to means 5 described below).

[0013] According to the above-mentioned means 1, the inspection image data is formed by placing the divided solder image in an image frame. Therefore, the size (width and height) of the inspection image data matches the size of the image frame and is constant without varying slightly depending on the land size. This eliminates the need to prepare multiple different identification means for each land size, reducing the effort and time required to obtain the identification means. Furthermore, the identification means can be used in common even when the land sizes are different.

[0014] Furthermore, the image frame of the learning data and the image frame of the test image data are the same size, and the sizes of the learning data and the test image data are the same. Therefore, when the test image data is input to the identification means, appropriate reconstructed image data corresponding to the test image data can be more reliably output, and the quality of the cream solder can be more accurately determined. This makes it possible to more reliably obtain good test accuracy.

[0015] Furthermore, the training data is composed of divided solder images, each of which is formed by dividing the solder area image into four parts, and each divided solder image is provided in an image frame. Therefore, compared to using image data of the entire cream solder as training data, a discrimination method with the desired accuracy can be obtained with less training data. Furthermore, since four training data can be obtained from one solder area image, the required number of training data can be easily obtained. These combined features can very effectively reduce the effort and time required to obtain a discrimination method.

[0016] In addition, because the training data and test image data are comprised of divided solder images set in image frames, the size of both data can be made relatively small compared to when an image of the entire cream solder is set in an image frame, which reduces the processing load for training and testing and allows for faster training and testing.

[0017] Means 2: A solder inspection device as described in Means 1, characterized in that the divided solder images in the learning data and the inspection image data are set so that the two sides corresponding to the virtual cross line and the corner formed by those two sides are oriented in a predetermined direction.

[0018] According to the above-mentioned means 2, the divided solder images in the learning data and the inspection image data are set so that the two sides corresponding to the virtual cross line and the corner formed by those two sides are oriented in a predetermined direction. For example, the divided solder images in both image data are set so that one of the two sides corresponding to the virtual cross line faces downward, the other faces left, and the corner formed by those two sides faces downward and left. Therefore, the learning efficiency is improved, and the labor and effort required to obtain the identification means can be further reduced. Furthermore, the accuracy of the reconstructed image data output from the identification means is improved, making it possible to further increase the inspection accuracy.

[0019] Furthermore, according to the above-mentioned means 2, the two sides corresponding to the virtual cross line and the corner formed by the two sides are common to all the learning data and the test image data, and therefore these do not become characteristic parts. Therefore, the accuracy of the reconstructed image data reconstructed by the identification means can be further improved, and the comparison by the comparison means can be performed more accurately. As a result, even better test accuracy can be obtained.

[0020] Means 3. A solder inspection device according to Means 1, characterized in that the virtual cross line comprises a first center line connecting the centers of two parallel sides of a circumscribing rectangle that circumscribes the solder area image and has the smallest area, and a second center line connecting the centers of the other two parallel sides of the circumscribing rectangle.

[0021] According to the above-mentioned means 3, even if the solder area image is rectangular or circular, the solder area image can be appropriately divided into four.

[0022] Means 4: The outer edge of one land to which the solder area image corresponds is rectangular in plan view, The solder inspection device described in means 1 is characterized in that the virtual cross line consists of a first center line connecting the centers of two parallel sides in a circumscribing rectangle that circumscribes the solder area image and extends in a direction that coincides with the extension direction of one land, and a second center line connecting the centers of the other two parallel sides in the circumscribing rectangle.

[0023] In addition, "a circumscribing rectangle extending in the same direction as the extension direction of one land" can be rephrased as a circumscribing rectangle having two sides parallel to two opposing sides that form the outer edge of one land (the same applies to Means 8 described later).

[0024] According to the above-mentioned means 4, the process of determining the circumscribing rectangle becomes relatively easy, which can further reduce the processing load related to learning and inspection, and can further speed up learning and inspection.

[0025] Means 5. A solder inspection device for inspecting cream solder printed on a printed circuit board, an image data acquisition means for acquiring image data of a predetermined inspection area on the printed circuit board including the printed solder paste; a recognition means for generating the image data by training only image data relating to non-defective cream solders in a neural network having an encoding unit for extracting feature values ​​from input image data and a decoding unit for reconstructing image data from the feature values; an inspection image data acquisition means for acquiring inspection image data including an image of the cream solder to be inspected based on the image data acquired by the image data acquisition means; a reconstructed image data acquisition means for inputting the inspection image data to the identification means and acquiring reconstructed image data as reconstructed image data; a comparison means for comparing the inspection image data and the reconstructed image data; The method is configured to be able to determine whether the cream solder is good or bad based on the comparison result by the comparison means, The learning data is composed of two divided solder images obtained by dividing one solder area image showing a cream solder corresponding to one land by a virtual line that can divide the image into two identical shapes, and each divided solder image is provided in an image frame that is larger than the size of the divided solder image, The solder inspection device is characterized in that the inspection image data acquisition means acquires the inspection image data in which the solder area image in the image data acquired by the image data acquisition means is divided into two by the virtual line, and each of the two divided solder images is placed in an image frame of the same size as the image frame of the learning data.

[0026] According to the above-mentioned means 5, the size (width and height) of the inspection image data matches the size of the image frame and is constant without minute variations depending on the land size. This eliminates the need to prepare a large number of different identification means for each land size, reducing the labor and time required to obtain the identification means. Furthermore, the identification means can be commonly used even when the land sizes are different.

[0027] Furthermore, since the sizes of the learning data and the test image data are the same, when the test image data is input to the identification means, appropriate reconstructed image data corresponding to the test image data can be output more reliably. This makes it possible to more accurately determine whether the cream solder is good or bad, and more reliably obtain good test accuracy.

[0028] Furthermore, the training data is formed by dividing the solder area image into two parts, each of which is provided with an image frame. Therefore, compared to using image data of the entire cream solder as training data, a discrimination method with the desired accuracy can be obtained with less training data. Furthermore, since two training data can be obtained from one solder area image, the required number of training data can be easily obtained. These combined features effectively reduce the effort and time required to obtain a discrimination method.

[0029] In addition, because the training data and test image data are comprised of divided solder images set in image frames, the size of both data can be made relatively small compared to when an image of the entire cream solder is set in an image frame, which reduces the processing load for training and testing and allows for faster training and testing.

[0030] Means 6. A solder inspection device according to Means 5, characterized in that the divided solder images in the learning data and the inspection image data are set so that the side corresponding to the virtual line is oriented in a predetermined direction.

[0031] According to the above-mentioned means 6, the divided solder images in the learning data and the inspection image data are set so that the sides corresponding to the virtual lines are oriented in a predetermined direction. For example, the divided solder images in both image data are set so that the sides corresponding to the virtual lines are oriented downward. This improves learning efficiency and reduces the labor and time required to obtain an identification means. Furthermore, the accuracy of the reconstructed image data output from the identification means is improved, making it possible to further increase inspection accuracy.

[0032] Furthermore, according to the above-mentioned means 6, the side corresponding to the virtual line is common to all the learning data and the test image data, so this side does not become a feature. Therefore, the accuracy of the reconstructed image data reconstructed by the identification means can be further improved, and the comparison by the comparison means can be performed more accurately. As a result, better test accuracy can be obtained.

[0033] Means 7: The solder inspection device according to Means 5, wherein the virtual line is a center line connecting the centers of two parallel sides of a circumscribing rectangle that circumscribes the solder area image and has the smallest area.

[0034] According to the above-mentioned means 7, even if the solder area image is rectangular or circular, the solder area image can be appropriately divided into two.

[0035] Means 8. The outer edge of one land to which the solder area image corresponds is rectangular in plan view, The solder inspection device described in means 5 is characterized in that the virtual line consists of a center line connecting the centers of two parallel sides of a circumscribing rectangle that circumscribes the solder area image and extends in a direction that coincides with the extension direction of one land.

[0036] According to the above-mentioned means 8, the process of determining the circumscribing rectangle becomes relatively easy, and the processing load for learning and inspection can be further reduced, thereby more effectively speeding up learning and inspection.

[0037] The technical matters relating to the above means may be combined as appropriate. For example, the technical matters relating to the above means 2 may be combined with the technical matters relating to the above means 3 or 4. [Brief explanation of the drawings]

[0038] [Figure 1] FIG. 2 is a partially enlarged plan view of a portion of the printed circuit board. [Figure 2] FIG. 2 is a partially enlarged cross-sectional view of a portion of the printed circuit board. [Figure 3] FIG. 1 is a block diagram showing the configuration of a manufacturing line for printed circuit boards. [Figure 4] FIG. 1 is a schematic diagram illustrating a solder inspection device. [Figure 5] FIG. 2 is a block diagram showing the functional configuration of the solder inspection device. [Figure 6] FIG. 1 is a schematic diagram for explaining the structure of a neural network. [Figure 7] 10 is a flowchart showing the flow of a learning process of a neural network. [Figure 8] 10 is a flowchart showing the flow of an inspection process. [Figure 9] FIG. 2 is a schematic diagram showing original image data for learning; [Figure 10] 3 is a schematic diagram showing a solder region image Hg1 in the first embodiment. FIG. [Figure 11]FIG. 2 is a schematic diagram showing a solder region image Hg2 in the first embodiment. [Figure 12] 10A to 10C are schematic diagrams for explaining the process of acquiring divided solder images Bg1, Bg2, Bg3, and Bg4, and the image processing of the divided solder images Bg1, Bg2, and Bg3 in the first embodiment. [Figure 13] 10A and 10B are schematic diagrams for explaining the process of acquiring divided solder images Bg5, Bg6, Bg7, and Bg8, and the image processing of divided solder images Bg5, Bg6, and Bg7 in the first embodiment. [Figure 14] 3A to 3C are schematic diagrams showing image frames and learning data G1 in the first embodiment. [Figure 15] 3A to 3C are schematic diagrams showing image frames and learning data G2 in the first embodiment. [Figure 16] 10A to 10C are schematic diagrams showing image frames and learning data G3 in the first embodiment. [Figure 17] 10A and 10B are schematic diagrams showing image frames and learning data G4 in the first embodiment. [Figure 18] 10A and 10B are schematic diagrams showing image frames and learning data G5 in the first embodiment. [Figure 19] 10A and 10B are schematic diagrams showing image frames and learning data G6 in the first embodiment. [Figure 20] 10A and 10B are schematic diagrams showing image frames and learning data G7 in the first embodiment. [Figure 21] 10A and 10B are schematic diagrams showing image frames and learning data G8 in the first embodiment. [Figure 22] FIG. 2 is a schematic diagram showing original image data for inspection. [Figure 23] FIG. 2 is a schematic diagram showing a solder region image Hk1 in the first embodiment. [Figure 24] FIG. 2 is a schematic diagram showing a solder region image Hk2 in the first embodiment. [Figure 25] 10A to 10C are schematic diagrams for explaining the process of acquiring divided solder images Bk1, Bk2, Bk3, and Bk4, and the image processing of divided solder images Bk1, Bk2, and Bk3 in the first embodiment. [Figure 26]10A and 10B are schematic diagrams for explaining the process of acquiring divided solder images Bk5, Bk6, Bk7, and Bk8, and the image processing of divided solder images Bk5, Bk6, and Bk7 in the first embodiment. [Figure 27] 3A to 3C are schematic diagrams showing an image frame and test image data K1 in the first embodiment. [Figure 28] 3A to 3C are schematic diagrams showing an image frame and test image data K2 in the first embodiment. [Figure 29] 3A to 3C are schematic diagrams showing an image frame and test image data K3 in the first embodiment. [Figure 30] 10A and 10B are schematic diagrams showing an image frame and test image data K4 in the first embodiment. [Figure 31] 10A and 10B are schematic diagrams showing an image frame and test image data K5 in the first embodiment. [Figure 32] 10A and 10B are schematic diagrams showing an image frame and test image data K6 in the first embodiment. [Figure 33] 10A and 10B are schematic diagrams showing an image frame and test image data K7 in the first embodiment. [Figure 34] 10A and 10B are schematic diagrams showing an image frame and test image data K8 in the first embodiment. [Figure 35] FIG. 2 is a schematic diagram showing reconstructed image data output from an AI model when test image data K1 is input in the first embodiment. [Figure 36] FIG. 10 is a schematic diagram showing reconstructed image data output from an AI model when inspection image data K5 is input in the first embodiment. [Figure 37] 10A and 10B are schematic diagrams for explaining the process of obtaining divided solder images Bg1 and Bg2, and the image processing of divided solder image Bg1 in the second embodiment. [Figure 38] 10A and 10B are schematic diagrams for explaining the process of obtaining divided solder images Bg3 and Bg4, and the image processing of divided solder image Bg3 in the second embodiment. [Figure 39] 10A and 10B are schematic diagrams showing image frames and learning data G1 in the second embodiment. [Figure 40] 10A and 10B are schematic diagrams showing image frames and learning data G2 in the second embodiment. [Figure 41] 10A and 10B are schematic diagrams showing image frames and learning data G3 in the second embodiment. [Figure 42] 10A and 10B are schematic diagrams showing image frames and learning data G4 in the second embodiment. [Figure 43] 10A and 10B are schematic diagrams for explaining the process of acquiring divided solder images Bk1 and Bk2, image processing of divided solder image Bk1, and the like in the second embodiment. [Figure 44] 10A and 10B are schematic diagrams for explaining the process of acquiring divided solder images Bk5 and Bk6, and the image processing of divided solder image Bk5 in the second embodiment. [Figure 45] 10A and 10B are schematic diagrams showing an image frame and test image data K1 in the second embodiment. [Figure 46] 10A and 10B are schematic diagrams showing an image frame and test image data K2 in the second embodiment. [Figure 47] 10A and 10B are schematic diagrams showing an image frame and test image data K3 in the second embodiment. [Figure 48] 10A and 10B are schematic diagrams showing an image frame and test image data K4 in the second embodiment. [Figure 49] FIG. 10 is a schematic diagram showing circumscribing rectangles, imaginary cross lines, and the like when dividing a solder region image into four parts in another embodiment. [Figure 50] FIG. 10 is a schematic diagram showing a circumscribing rectangle, imaginary lines, etc. when dividing a solder region image into two in another embodiment. [Figure 51] FIG. 10 is a schematic diagram showing a circumscribing rectangle or the like having a shape corresponding to the outer edge of a land in another embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0039] Hereinafter, embodiments will be described with reference to the drawings. [First embodiment] First, the configuration of the printed circuit board will be described. Fig. 1 is a partially enlarged plan view of a portion of the printed circuit board, and Fig. 2 is a partially enlarged cross-sectional view of a portion of the printed circuit board.

[0040] 1 and 2, the printed circuit board 1 has a wiring pattern (not shown) made of copper foil and a plurality of lands 3 that are rectangular or circular in plan view formed on the surface of a flat base substrate 2 made of glass epoxy resin or the like. A resist film 4 is coated on the surface of the base substrate 2 except for the lands 3.

[0041] Furthermore, cream solder 5 made by kneading solder particles with flux is printed on the lands 3. In this embodiment, at least the cream solder 5 that is rectangular in plan view and printed on the rectangular lands 3, and the cream solder 5 that is circular in plan view and printed on the circular lands 3 are present. For convenience, in FIG. 1 and other figures, the cream solder 5 is indicated by a scattered dot pattern. Furthermore, in FIG. 1 and other figures, only a small portion of the multiple lands 3 and cream solder 5 that the printed circuit board 1 has is shown.

[0042] Next, a manufacturing line (manufacturing process) for manufacturing the printed circuit board 1 will be described with reference to Fig. 3. As shown in Fig. 3, a manufacturing line 10 is provided with, in order from the upstream side (upper side in Fig. 3), a solder printer 12, a solder inspection device 13, a component mounter 14, a reflow device 15, and a post-reflow inspection device 16.

[0043] The solder printer 12 performs a solder printing process for printing cream solder 5 on each land 3 of the printed circuit board 1. In the solder printing process, the cream solder 5 is printed by, for example, screen printing. In screen printing, first, the bottom surface of a screen mask is brought into contact with the printed circuit board 1, and cream solder 5 is supplied to the top surface of the screen mask. The screen mask has a plurality of openings formed therein, each corresponding to each land 3 of the printed circuit board 1. Next, a predetermined squeegee is brought into contact with the top surface of the screen mask and moved, thereby filling the openings with the cream solder 5. Thereafter, the printed circuit board 1 is separated from the bottom surface of the screen mask, and the cream solder 5 is printed on each land 3 of the printed circuit board 1.

[0044] The solder inspection device 13 inspects the cream solder 5, and judges the quality of the cream solder 5, for example, regarding the shape, the presence or absence of foreign matter, etc. The solder inspection device 13 will be described in detail later.

[0045] The component mounter 14 mounts an electronic component 25 (see FIGS. 1 and 2) on the land 3 on which the cream solder 5 is printed. The electronic component 25 has a plurality of electrodes (not shown), and each of the electrodes is temporarily fixed to a predetermined cream solder 5.

[0046] The reflow device 15 performs a reflow process in which the cream solder 5 is heated and melted to solder the lands 3 and the electrodes of the electronic component 25 together.

[0047] The post-reflow inspection device 16 checks whether the solder joints have been properly made in the reflow process by checking for the presence or absence of misalignment in the electronic components 25 using, for example, brightness image data.

[0048] In addition, although not shown, the manufacturing line 10 is equipped with conveyors or the like for transporting the printed circuit board 1 between the above-mentioned devices, such as between the solder printer 12 and the solder inspection device 13. Branching devices are also provided between the solder inspection device 13 and the component mounter 14 and downstream of the post-reflow inspection device 16. The printed circuit boards 1 that have been determined to be non-defective by the solder inspection device 13 or the post-reflow inspection device 16 are guided directly downstream, while the printed circuit boards 1 that have been determined to be defective are discharged by the branching device to a defective product storage area.

[0049] Next, the configuration of the solder inspection device 13 will be described in detail with reference to Figures 4 and 5. Figure 4 is a schematic diagram showing the configuration of the solder inspection device 13. Figure 5 is a block diagram showing the functional configuration of the solder inspection device 13.

[0050] The solder inspection device 13 includes a transport mechanism 31 that transports and positions the printed circuit board 1, an inspection unit 32 that obtains image data of the printed circuit board 1, and a control device 33 (see Figure 5) that controls the drive of the transport mechanism 31 and the inspection unit 32, as well as performs various controls, image processing, and calculation processing in the solder inspection device 13.

[0051] The transport mechanism 31 includes a pair of transport rails 31a arranged along the direction in which the printed circuit board 1 is carried in and out, and an endless conveyor belt 31b rotatably disposed relative to each of the transport rails 31a. Although not shown, the transport mechanism 31 also includes a driving means such as a motor for driving the conveyor belt 31b, and a chuck mechanism for positioning the printed circuit board 1 at a predetermined position. The transport mechanism 31 is driven and controlled by a control device 33 (a transport mechanism control unit 79, which will be described later).

[0052] With the above configuration, the printed circuit board 1 is carried into the solder inspection device 13. Both side edges in the width direction perpendicular to the carrying-in / out direction are inserted into the conveyor rails 31a, and the printed circuit board 1 is placed on the conveyor belt 31b. The conveyor belt 31b then starts operating, transporting the printed circuit board 1 to a predetermined inspection position. When the printed circuit board 1 reaches the inspection position, the conveyor belt 31b stops and the chucking mechanism is activated. The operation of this chucking mechanism pushes up the conveyor belt 31b, and both side edges of the printed circuit board 1 are clamped between the conveyor belt 31b and the upper edge of the conveyor rail 31a. This positions and fixes the printed circuit board 1 at the inspection position. When the inspection is completed, the chucking mechanism releases the fixation, and the conveyor belt 31b starts operating. The printed circuit board 1 is then carried out of the solder inspection device 13. Of course, the configuration of the transport mechanism 31 is not limited to the above, and other configurations may be adopted.

[0053] The inspection unit 32 is disposed above the transport rail 31a (the transport path of the printed circuit board 1). The inspection unit 32 includes a first lighting device 32a, a second lighting device 32b, a third lighting device 32c, and a camera 32d. In this embodiment, the camera 32d constitutes the "image data acquisition means."

[0054] The inspection unit 32 also includes an X-axis movement mechanism 32e (see FIG. 5) that allows movement in the X-axis direction (left-right direction in FIG. 4), and a Y-axis movement mechanism 32f (see FIG. 5) that allows movement in the Y-axis direction (front-back direction in FIG. 4). These movement mechanisms 32e and 32f are driven and controlled by the control device 33 (a movement mechanism control unit 76, which will be described later).

[0055] When performing three-dimensional measurement of the printed circuit board 1, the first lighting device 32a and the second lighting device 32b each irradiate a predetermined inspection area on the printed circuit board 1 from diagonally above with predetermined light for three-dimensional measurement (pattern light having a striped light intensity distribution).

[0056] Specifically, the first lighting device 32a includes a first light source 32a1 that emits a predetermined light, and a first liquid crystal shutter 32a2 that forms a first grating that converts the light from the first light source 32a1 into a first pattern light having a striped light intensity distribution, and is driven and controlled by the control device 33 (the lighting control unit 72 described later).

[0057] The second lighting device 32b includes a second light source 32b1 that emits a predetermined light, and a second liquid crystal shutter 32b2 that forms a second grating that converts the light from the second light source 32b1 into a second pattern light having a striped light intensity distribution, and is driven and controlled by the control device 33 (the lighting control unit 72 described later).

[0058] With the above configuration, the light emitted from each of the light sources 32a1 and 32b1 is guided to a condenser lens (not shown), where it is converted into parallel light, and then guided to a projection lens (not shown) via the liquid crystal shutters 32a2 and 32b2, and projected as patterned light onto the printed circuit board 1. In this embodiment, the liquid crystal shutters 32a2 and 32b2 are switched and controlled so that the phase of each patterned light is shifted by a quarter pitch.

[0059] Furthermore, by using liquid crystal shutters 32a2 and 32b2 as the grating, it is possible to irradiate pattern light that is close to an ideal sine wave. This improves the measurement resolution of three-dimensional measurement. In addition, the phase shift of the pattern light can be electrically controlled, allowing for a more compact device.

[0060] When performing two-dimensional measurement of the printed circuit board 1, the third illumination device 32c irradiates a predetermined inspection area on the printed circuit board 1 with predetermined light for two-dimensional measurement (for example, uniform light). The third illumination device 32c is equipped with a ring light capable of irradiating blue light, a ring light capable of irradiating green light, and a ring light capable of irradiating red light. Note that the third illumination device 32c has a configuration similar to that of known technology, and therefore detailed description thereof will be omitted.

[0061] The camera 32d captures an image of a predetermined inspection area of ​​the printed circuit board 1 from directly above. The camera 32d has an imaging element such as a CCD (Charge Coupled Device) image sensor or a CMOS (Complementary Metal Oxide Semiconductor) image sensor, and an optical system (lens unit, diaphragm, etc.) that forms an image of the printed circuit board 1 on the imaging element, and is arranged so that its optical axis is aligned in the vertical direction (Z-axis direction). Of course, the imaging element is not limited to these, and other imaging elements may be used.

[0062] Camera 32d is driven and controlled by control device 33 (camera control unit 73, described later). More specifically, control device 33 executes imaging processing by camera 32d in synchronization with the irradiation processing by each of lighting devices 32a, 32b, and 32c. As a result, light irradiated from any of lighting devices 32a, 32b, and 32c and reflected by printed circuit board 1 is imaged by camera 32d. As a result, image data of the inspection area of ​​printed circuit board 1, including cream solder 5 printed on lands 3, is acquired. Note that the "inspection area" of printed circuit board 1 is one of multiple areas preset on printed circuit board 1, with the size of the imaging field of view (imaging range) of camera 32d being one unit.

[0063] Furthermore, the camera 32d in this embodiment is configured as a color camera, which makes it possible to simultaneously capture images of the light of each color that is simultaneously irradiated from the ring lights of each color of the third illumination device 32c and reflected by the printed circuit board 1.

[0064] The image data captured and generated by the camera 32d is converted into a digital signal inside the camera 32d and then transferred in the form of a digital signal to the control device 33 (an image acquisition unit 74 described below). The control device 33 then stores the transferred image data and performs various image processing, arithmetic processing, and the like based on the image data.

[0065] The control device 33 consists of a computer including a CPU (Central Processing Unit) that executes predetermined arithmetic processing, a ROM (Read Only Memory) that stores various programs and fixed value data, a RAM (Random Access Memory) that temporarily stores various data when executing various arithmetic processing, and peripheral circuits for these.

[0066] The control device 33 functions as various functional units such as a main control unit 71, a lighting control unit 72, a camera control unit 73, an image acquisition unit 74, a data processing unit 75, a movement mechanism control unit 76, a learning unit 77, an inspection unit 78, and a transport mechanism control unit 79, as the CPU operates in accordance with various programs.

[0067] However, the various functional units are realized by the cooperation of various hardware such as the CPU, ROM, RAM, etc., and there is no need to clearly distinguish between functions realized by hardware and functions realized by software, and some or all of these functions may be realized by hardware circuits such as ICs.

[0068] Furthermore, the control device 33 is provided with an input unit 55 consisting of a keyboard, mouse, touch panel, etc., a display unit 56 with a display screen consisting of an LCD display, etc., a memory unit 57 capable of storing various data, programs, calculation results, test results, etc., and a communication unit 58 capable of sending and receiving various data to and from the outside.

[0069] Here, the above-mentioned various functional units that constitute the control device 33 will be described in detail.

[0070] The main control unit 71 is a functional unit that controls the entire solder inspection device 13, and is configured to be able to send and receive various signals to and from other functional units such as the illumination control unit 72 and camera control unit 73.

[0071] The illumination control unit 72 is a functional unit that controls the driving of the illumination devices 32a, 32b, and 32c, and performs switching control of the illumination light based on a command signal from the main control unit 71.

[0072] The camera control unit 73 is a functional unit that controls the driving of the camera 32d, and controls the timing of image capture and the like based on a command signal from the main control unit 71.

[0073] The image acquisition unit 74 is a functional unit for taking in image data captured and acquired by the camera 32d.

[0074] The data processing unit 75 is a functional unit that performs predetermined image processing on the image data captured by the image acquisition unit 74, and performs two-dimensional measurement processing, three-dimensional measurement processing, and the like using the image data.

[0075] The movement mechanism control unit 76 is a functional unit that drives and controls the X-axis movement mechanism 32e and the Y-axis movement mechanism 32f, and controls the position of the inspection unit 32 based on a command signal from the main control unit 71. By controlling and driving the X-axis movement mechanism 32e and the Y-axis movement mechanism 32f, the movement mechanism control unit 76 can move the inspection unit 32 to a position above any inspection area of ​​the printed circuit board 1 that is positioned and fixed at the inspection position. Then, the inspection unit 32 is moved sequentially to a plurality of inspection areas set on the printed circuit board 1, and inspections of the inspection areas are performed, thereby inspecting the entire area of ​​the printed circuit board 1.

[0076] The learning unit 77 is a functional unit that uses learning data to learn a deep neural network 90 (hereinafter simply referred to as "neural network 90"; see FIG. 6) and constructs an AI (Artificial Intelligence) model 101 as an "identification means."

[0077] As will be described later, the AI ​​model 101 in this embodiment is a generative model constructed by deep learning a neural network 90 using only image data relating to good cream solder 5 as training data, and has the structure of a so-called autoencoder.

[0078] Here, the structure of the neural network 90 will be described with reference to Fig. 6. Fig. 6 is a schematic diagram conceptually showing the structure of the neural network 90. ​​As shown in Fig. 6, the neural network 90 has a convolutional auto-encoder (CAE) structure that includes an encoder unit 91 as an "encoding unit" that extracts a feature (latent variable) TA from input image data GA, and a decoder unit 92 as a "decoding unit" that reconstructs image data GB from the feature TA.

[0079] The structure of a convolutional autoencoder is well known, and therefore a detailed description will be omitted. However, the encoder unit 91 has a plurality of convolution layers 93, and in each convolution layer 93, a convolution operation is performed on input data using a plurality of filters (kernels) 94, and the result is output as input data for the next layer. Similarly, the decoder unit 92 has a plurality of deconvolution layers 95, and in each deconvolution layer 95, a deconvolution operation is performed on input data using a plurality of filters (kernels) 96, and the result is output as input data for the next layer. Then, in the learning process described below, the weights (parameters) of each filter 94, 96 are updated.

[0080] The inspection unit 78 is a functional unit that inspects the quality of the cream solder 5. In this embodiment, the inspection unit 78 inspects whether the cream solder 5 is properly printed in terms of the presence, size, and shape of foreign matter.

[0081] The transfer mechanism control unit 79 is a functional unit that controls the drive of the transfer mechanism 31 , and controls the position of the printed circuit board 1 based on a command signal from the main control unit 71 .

[0082] The memory unit 57 is composed of a hard disk drive (HDD) or a solid state drive (SSD), and has a predetermined memory area for storing, for example, an AI model 101 (neural network 90 and learning information acquired by its learning).

[0083] The communication unit 58 is equipped with a wireless communication interface conforming to communication standards such as a wired LAN (Local Area Network) or a wireless LAN, and is configured to be able to transmit and receive various data to and from the outside. For example, the results of the inspection performed by the inspection unit 78 are output to the outside via the communication unit 58, and the results of the inspection performed by the post-reflow inspection device 16 are input via the communication unit 58.

[0084] Next, the learning process of the neural network 90 performed by the solder inspection device 13 will be described with reference to the flowchart of FIG.

[0085] When the learning process is started based on the execution of a predetermined learning program, the main control unit 71 first performs pre-processing for learning the neural network 90 in step S101.

[0086] In this pre-processing, first, inspection information of a large number of printed circuit boards 1 stored in the post-reflow inspection device 16 is obtained via the communication unit 58. Next, based on the inspection information, learning original image data Ig, which is image data relating to non-defective cream solder 5 that has passed the post-reflow inspection, is obtained from the storage unit 57 (see, for example, FIG. 9). FIG. 9 shows only a portion of the learning original image data Ig, and the actual learning original image data Ig contains many more cream solders 5 and lands 3.

[0087] The original training image data Ig pertains to the printed circuit board 1 after the printing of the cream solder 5 and before the mounting of the electronic components 25, and is used to obtain the later-described training data G1, G2, G3, G4, G5, G6, G7, and G8 (hereinafter referred to as "training data G1 to G8") used for training the neural network 90. ​​The original training image data Ig includes three-dimensional data, which is image data obtained by imaging the printed circuit board 1 with the camera 32d while irradiating it with patterned light from the first illumination device 32a or the second illumination device 32b, and two-dimensional data, which is image data obtained by imaging the printed circuit board 1 with the camera 32d while irradiating it with uniform light from the third illumination device 32c.

[0088] The original image data for learning Ig may be image data obtained by the camera 32d without any special processing (for example, monochrome luminance image data or RGB luminance image data), or image data obtained by performing predetermined processing on the image data obtained by the camera 32d (for example, HLS image data obtained by converting RGB image data, or height image data obtained by converting image data) (the same applies to the original image data for inspection Ik described later).

[0089] Next, the area occupied by the cream solder 5 in the acquired learning original image data Ig is identified. If the learning original image data Ig is two-dimensional data, the area occupied by the cream solder 5 is identified using, for example, brightness, hue, saturation, etc. If the learning original image data Ig is three-dimensional data, the area occupied by the cream solder 5 is identified using, for example, height information, etc.

[0090] Next, images of connected components (lump portions) in the area occupied by the identified cream solder 5 are extracted as solder area images Hg1 and Hg2 (see, for example, FIGS. 10 and 11). The solder area image Hg1 relates to a rectangular cream solder 5 printed on a rectangular land 3. The solder area image Hg2 relates to a circular cream solder 5 printed on a circular land 3. FIGS. 10 and 11 show some of the solder area images Hg1 and Hg2 as examples.

[0091] The solder region images Hg1 and Hg2 correspond to the land 3, and in this embodiment, the connected components (lump portions) located on the land 3 in the design within the area occupied by the cream solder 5 are extracted as the solder region images Hg1 and Hg2. The connected components (lump portions) may be simply extracted as the solder region images Hg1 and Hg2 without considering the position of the land 3. Furthermore, to eliminate the influence of minute shape differences in the outer edge (foot portion) of the cream solder 5, the solder region images Hg1 and Hg2 may be contracted to obtain the solder region images Hg1 and Hg2 from which the outer edge (foot portion) has been removed. Furthermore, the solder region images Hg1 and Hg2 from which the outer edge (foot portion) has been removed may be obtained by cutting portions below a predetermined height.

[0092] Next, divided solder images Bg1, Bg2, Bg3, Bg4, Bg5, Bg6, Bg7, and Bg8 (see FIGS. 12 and 13; hereinafter referred to as "divided solder images Bg1 to Bg8") are obtained from the solder area images Hg1 and Hg2. The divided solder images Bg1, Bg2, Bg3, and Bg4 (hereinafter referred to as "divided solder images Bg1 to Bg4") are images obtained by dividing one solder area image Hg1 into four by imaginary cross lines Xga (see FIG. 12) that can be divided into four equal-shaped images. The divided solder images Bg5, Bg6, Bg7, and Bg8 (hereinafter referred to as "divided solder images Bg5 to Bg8") are images obtained by dividing one solder area image Hg2 into four by imaginary cross lines Xgb (see FIG. 13) that can be divided into four equal-shaped images. Therefore, the divided solder images Bg1 to Bg4 relate to rectangular cream solder 5 printed on a rectangular land 3, and the divided solder images Bg5 to Bg8 relate to circular cream solder 5 printed on a circular land 3.

[0093] The imaginary cross line Xga consists of a first center line Cga1 connecting the centers of two parallel sides of the circumscribing rectangle Sga, and a second center line Cga2 perpendicular to the first center line Cga1 and connecting the centers of the other two parallel sides of the circumscribing rectangle Sga. The imaginary cross line Xgb consists of a first center line Cgb1 connecting the centers of two parallel sides of the circumscribing rectangle Sgb, and a second center line Cgb2 perpendicular to the first center line Cgb1 and connecting the centers of the other two parallel sides of the circumscribing rectangle Sgb. As described above, the imaginary cross lines Xga and Xgb are capable of dividing the solder region images Hg1 and Hg2 into four equal-shaped portions. However, the phrase "capable of being divided into four equal-shaped portions" more precisely means that when the cream solder 5 is printed in an ideal shape (i.e., a designed shape), the solder region image corresponding to the cream solder 5 can be divided into four equal-shaped portions (the same applies to the imaginary cross lines Xka and Xkb described below). The circumscribing rectangle Sga is the rectangle with the smallest area among the rectangles circumscribing the solder region image Hg1, and the circumscribing rectangle Sgb is the rectangle with the smallest area among the rectangles circumscribing the solder region image Hg2.

[0094] Next, image processing (rotation and flipping) is performed on those of the divided solder images Bg1 to Bg8 that require it (in this embodiment, divided solder images Bg1, Bg2, Bg3, Bg5, Bg6, and Bg7), so that the two sides corresponding to the imaginary cross lines Xga and Xgb in the divided solder images Bg1 to Bg8 and the corner formed by these two sides are set to have a predetermined orientation (see the lower diagrams of FIGS. 12 and 13). In this embodiment, the divided solder images Bg1 to Bg8 are each set so that the side corresponding to the first center lines Cga1 and Cgb1 faces downward, the side corresponding to the second center lines Cga2 and Cgb2 faces left, and the corner formed by these two sides faces downward and left.

[0095] Then, by pasting the divided solder images Bg1 to Bg4 respectively into the image frame W1, learning data G1, G2, G3, G4 (see FIGS. 14 to 17, hereinafter referred to as "learning data G1 to G4") are obtained in which the divided solder images Bg1 to Bg4 are arranged in the image frame W1. Also, by pasting the divided solder images Bg5 to Bg8 respectively into the image frame W1, learning data G5, G6, G7, G8 (see FIGS. 18 to 21, hereinafter referred to as "learning data G5 to G8") are obtained in which the divided solder images Bg5 to Bg8 are arranged in the image frame W1.

[0096] The image frame W1 is a rectangular image with a height (width in the vertical direction of the paper surface of FIG. 14, etc.) of m (pixels) and a width (width in the horizontal direction of the paper surface of FIG. 14, etc.) of n (pixels), and its size (width and height) is set based on design data, etc. so as to be larger than the size of each of the divided solder images Bg1 to Bg8. Note that m and n are both natural numbers, and in this embodiment, m = n. Furthermore, by adjusting the pasting position, each of the learning data G1 to G8 is made to coincide with the center or center of gravity of the divided solder images Bg1 to Bg8 and the center of the image frame W1.

[0097] Then, by repeating the above processes, such as extracting solder region images Hg1 and Hg2, acquiring divided solder images Bg1 to Bg8, performing image processing on the divided solder images Bg1 to Bg8 as needed, and pasting the divided solder images Bg1 to Bg8 into the image frame W1, multiple pieces of training data G1 to G8 are acquired from one training original image data Ig. At this time, four pieces of training data G1 to G4 are acquired for each rectangular land 3 (cream solder 5) included in the training original image data Ig. Also, four pieces of training data G5 to G8 are acquired for each circular land 3 (cream solder 5) included in the training original image data Ig. Furthermore, by using multiple training original image data Ig, the required number of pieces of training data G1 to G8 are finally acquired. In this embodiment, the training data G1 to G8 include data acquired based on two-dimensional data and data acquired based on three-dimensional data.

[0098] In step S101, the number of pieces of learning data G1 to G8 required for learning is acquired. In the following step S102, the learning unit 77 prepares an unlearned neural network 90 based on a command from the main control unit 71. For example, the learning unit 77 reads out the neural network 90 stored in advance in the storage unit 57 or the like. Alternatively, the neural network 90 is constructed based on network configuration information (for example, the number of layers of the neural network and the number of nodes in each layer) stored in the storage unit 57 or the like.

[0099] In this embodiment, two neural networks 90 are constructed: one for learning using learning data G1 to G8 acquired based on two-dimensional data, and the other for learning using learning data G1 to G8 acquired based on three-dimensional data. Therefore, in this embodiment, two neural networks 90 are constructed.

[0100] In step S103, reconstructed image data is acquired. That is, based on a command from the main control unit 71, the learning unit 77 provides the training data G1 to G8 acquired in step S102 as input data to the input layer of the neural network 90, thereby acquiring reconstructed image data output from the output layer of the neural network 90. ​​More specifically, the learning unit 77 provides the training data G1 to G8 acquired in step S102 that corresponds to the neural network 90 as input data to the input layer of the neural network 90, thereby acquiring reconstructed image data output from the output layer of the neural network 90. ​​For example, the learning unit 77 provides the training data G1 to G8 obtained from two-dimensional data as input data to the input layer of the neural network 90 that performs training using the training data G1 to G8, and acquires the reconstructed image data output from the neural network 90. ​​That is, the learning unit 77 inputs appropriate training data G1 to G8 to each of the two neural networks 90 and acquires the output reconstructed image data.

[0101] In the following step S104, the learning unit 77 compares the input learning data G1 to G8 with the reconstructed image data output by the neural network 90, and determines whether the error is sufficiently small (whether it is equal to or smaller than a predetermined threshold).

[0102] If the error is sufficiently small, then in step S106, the learning unit 77 determines whether or not a learning termination condition is met. For example, if a positive determination is made in step S104 a predetermined number of times in succession without going through the processing of step S105 (described later), or if learning using all of the prepared learning data G1 to G8 is repeated a predetermined number of times, then it is determined that the termination condition is met. If the termination condition is met, then the neural network 90 and its learning information (such as updated parameters (described later)) are stored in the storage unit 57 as the AI ​​model 101, and the learning process is terminated.

[0103] In this embodiment, the AI ​​model 101 ultimately stores an AI model obtained by learning the learning data G1 to G8 acquired from two-dimensional data and an AI model obtained by learning the learning data G1 to G8 acquired from three-dimensional data.

[0104] On the other hand, if the termination condition is not met in step S106, the process returns to step S102, and the neural network 90 is trained again.

[0105] If the error is not sufficiently small in step S104, the network update process (learning of the neural network 90) is performed in step S105, and then the process returns to step S103 to repeat the above series of processes.

[0106] Specifically, in the network update process of step S105, a known learning algorithm such as backpropagation is used to update the weights (parameters) of the filters 94, 96 in the neural network 90 to more appropriate ones so that the loss function representing the difference between the training data G1 to G8 and the reconstructed image data is minimized. Note that, for example, BCE (Binary Cross-entropy) can be used as the loss function.

[0107] By repeating the processes of steps S103 to S105 many times, the neural network 90 minimizes the error between the training data G1 to G8 and the reconstructed image data, and outputs more accurate reconstructed image data.

[0108] When image data relating to a good cream solder 5 is input, the finally obtained AI model 101 generates reconstructed image data that substantially matches the image data. Furthermore, when image data relating to a defective cream solder 5 in terms of shape, size, and the presence or absence of foreign matter is input, the AI ​​model 101 generates reconstructed image data that substantially matches the image data after correcting the shape and size of the cream solder 5 and removing noise portions (portions corresponding to foreign matter). In other words, when the cream solder 5 is defective, virtual image data relating to the cream solder 5, assuming that there are no defective portions, is generated as the reconstructed image data relating to the cream solder 5.

[0109] Next, the inspection process performed by the solder inspection device 13 will be described with reference to the flowchart of Fig. 8. This inspection process is executed for each area to be inspected on the printed circuit board 1.

[0110] When the printed circuit board 1 is carried into the solder inspection device 13 and positioned at a predetermined inspection position, inspection processing is started based on the execution of a predetermined inspection program.

[0111] When the inspection process starts, first, in step S301, an image data acquisition process is performed. In the image data acquisition process, original inspection image data Ik (see, for example, FIG. 22) related to the printed circuit board 1 to be inspected is acquired. The original inspection image data Ik is image data for obtaining inspection image data K1, K2, K3, K4, K5, K6, K7, and K8 (hereinafter referred to as "inspection image data K1 to K8") described below. In this embodiment, an example of the printed circuit board 1 to be inspected is one having cream solder 5 with foreign matter F1 attached and cream solder 5 with an abnormal shape (chipped).

[0112] The original inspection image data Ik includes three-dimensional data, which is image data obtained by imaging the printed circuit board 1 with the camera 32d while irradiating it with patterned light from the first illumination device 32a or the second illumination device 32b, and two-dimensional data, which is image data obtained by imaging the printed circuit board 1 with the camera 32d while irradiating it with uniform light from the third illumination device 32c. In the image data acquisition process, a three-dimensional data acquisition process and a two-dimensional data acquisition process are performed.

[0113] First, the three-dimensional data acquisition process will be described. In this process, the phase of the first pattern light irradiated from the first illumination device 32a is changed, and four imaging processes are performed under the first pattern light with different phases. Then, the phase of the second pattern light irradiated from the second illumination device 32b is changed, and four imaging processes are performed under the second pattern light with different phases, and a total of eight sets of three-dimensional data are acquired. This will be described in detail below.

[0114] As described above, when the printed circuit board 1 carried into the solder inspection device 13 is positioned and fixed at a predetermined inspection position, based on instructions from the main control unit 71, the movement mechanism control unit 76 first drives and controls the X-axis movement mechanism 32e and the Y-axis movement mechanism 32f to move the inspection unit 32, and aligns the imaging field of view (imaging range) of the camera 32d with the predetermined inspection area of ​​the printed circuit board 1.

[0115] Additionally, the illumination control unit 72 controls the switching of the liquid crystal shutters 32a2, 32b2 of the illumination devices 32a, 32b, and sets the positions of the first grating and second grating formed on the liquid crystal shutters 32a2, 32b2 to predetermined reference positions.

[0116] When the switching between the first and second gratings is complete, the illumination control unit 72 causes the first light source 32a1 of the first illumination device 32a to emit light and irradiate the first pattern light, and the camera control unit 73 drives and controls the camera 32d to perform a first imaging process under the first pattern light. The image data generated by the imaging process is continuously imported into the image acquisition unit 74 (and so on). This allows three-dimensional data of the inspection area including the multiple lands 3 and multiple cream solders 5 to be acquired.

[0117] Thereafter, upon completion of the first imaging process under the first pattern light, the illumination control unit 72 turns off the first light source 32a1 of the first illumination device 32a and executes a switching process for the first liquid crystal shutter 32a2. Specifically, the illumination control unit 72 switches the position of the first grating formed on the first liquid crystal shutter 32a2 from the reference position to a second position where the phase of the first pattern light is shifted by a quarter pitch (90°).

[0118] When the switching setting of the first grating is completed, the illumination control unit 72 causes the light source 32a1 of the first illumination device 32a to emit light to irradiate the first pattern light, and the camera control unit 73 drives and controls the camera 32d to perform a second imaging process under the first pattern light. Thereafter, by repeating the same process, four sets of three-dimensional data are acquired under the first pattern light, each with a phase difference of 90°.

[0119] Next, the lighting control unit 72 causes the second light source 32b1 of the second lighting device 32b to emit light and irradiate the second pattern light, and the camera control unit 73 drives and controls the camera 32d to perform the first imaging process under the second pattern light.

[0120] Thereafter, at the same time as the end of the first imaging process under the second pattern light, the illumination control unit 72 turns off the second light source 32b1 of the second illumination device 32b and executes a switching process for the second liquid crystal shutter 32b2. Specifically, the position of the second grating formed on the second liquid crystal shutter 32b2 is switched from the reference position to a second position where the phase of the second pattern light is shifted by a quarter pitch (90°).

[0121] When the second grating setting is complete, the illumination control unit 72 causes the light source 32b1 of the second illumination device 32b to emit light to irradiate the second pattern light, and the camera control unit 73 drives and controls the camera 32d to perform a second imaging process under the second pattern light. Thereafter, by repeating the same process, four sets of three-dimensional data are acquired under the second pattern light, each with a phase difference of 90°.

[0122] Next, the two-dimensional data acquisition process will be described. In this process, based on a command from the main control unit 71, the illumination control unit 72 causes the third illumination device 32c to emit light, irradiating a predetermined inspection area with uniform light, while the camera control unit 73 drives and controls the camera 32d to perform imaging processing under the uniform light. As a result, the predetermined inspection area on the printed circuit board 1 is imaged, and two-dimensional data related to the inspection area is acquired.

[0123] The acquired original inspection image data Ik (three-dimensional data and two-dimensional data) is stored in the storage unit 57.

[0124] Next, in step S302, a test image data acquisition process is executed. In the test image data acquisition process, test image data K1 to K8 (see FIGS. 27 to 34) are acquired based on the original test image data Ik obtained in the image data acquisition process. The method for acquiring the test image data K1 to K8 is the same as the method for acquiring the learning data G1 to G8 described above.

[0125] That is, first, solder area images Hk1 and Hk2 (see FIGS. 23 and 24) are extracted from the acquired original image data for inspection Ik. The extraction of solder area images Hk1 and Hk2 is performed in the same manner as the extraction of solder area images Hg1 and Hg2 from the original image data for learning Ig. Note that solder area image Hk1 relates to a rectangular cream solder 5 printed on a rectangular land 3. Solder area image Hk2 relates to a circular cream solder 5 printed on a circular land 3. In FIG. 23 etc., some of the solder area images Hk1 and Hk2 are shown as examples.

[0126] Next, divided solder images Bk1, Bk2, Bk3, Bk4, Bk5, Bk6, Bk7, and Bk8 (see FIGS. 25 and 26; hereinafter referred to as "divided solder images Bk1 to Bk8") are obtained from the solder area images Hk1 and Hk2 in the same manner as the divided solder images Bg1 to Bg8 obtained in the learning process. That is, by dividing the solder area image Hk1 into four by virtual cross lines Xka that can divide it into four equal-shaped areas, four divided solder images Bk1, Bk2, Bk3, and Bk4 (see FIG. 25; hereinafter referred to as "divided solder images Bk1 to Bk4") are obtained from the solder area image Hk1. Furthermore, by dividing the solder area image Hk2 into four parts by a virtual cross line Xkb that can divide the image into four parts of the same shape, four divided solder images Bk5, Bk6, Bk7, and Bk8 (see FIG. 26; hereinafter referred to as "divided solder images Bk5 to Bk8") are obtained from the solder area image Hk2.

[0127] The imaginary cross line Xka is made up of a first center line Cka1 connecting the centers of two parallel sides of the circumscribing rectangle Ska, and a second center line Cka2 perpendicular to the first center line Cka1 and connecting the centers of the other two parallel sides of the circumscribing rectangle Ska. The imaginary cross line Xkb is made up of a first center line Ckb1 connecting the centers of the two parallel sides of the circumscribing rectangle Skb, and a second center line Ckb2 perpendicular to the first center line Ckb1 and connecting the centers of the other two parallel sides of the circumscribing rectangle Skb. The circumscribing rectangle Ska is the rectangle with the smallest area among the rectangles circumscribing the solder area image Hk1. The circumscribing rectangle Skb is the rectangle with the smallest area among the rectangles circumscribing the solder area image Hk2.

[0128] Next, image processing (rotation and flipping) is performed on the divided solder images Bk1 to Bk8 that require image processing (in this embodiment, divided solder images Bk1, Bk2, Bk3, Bk5, Bk6, and Bk7), so that the two sides corresponding to the imaginary cross lines Xka and Xkb in the divided solder images Bk1 to Bk8 and the corner formed by these two sides are set to have a predetermined orientation (see the lower diagrams of FIGS. 25 and 26). In this embodiment, the divided solder images Bk1 to Bk8 are set, similar to the divided solder images Bg1 to Bg8 in the learning data G1 to G8, so that the side corresponding to the first center lines Cka1 and Ckb1 faces downward, the side corresponding to the second center lines Cka2 and Ckb2 faces left, and the corner formed by these two sides faces downward and left.

[0129] Next, the divided solder images Bk1 to Bk8 are pasted into the image frame W1, respectively, to obtain inspection image data K1 to K8 in which the divided solder images Bk1 to Bk8 are provided in the image frame W1 (see FIGS. 27 to 34). The image frame W1 of the inspection image data K1 to K8 has the same size as the image frame W1 of the learning data G1 to G8, so the size (width and height) of the inspection image data K1 to K8 is the same as the size of the learning data G1 to G1.

[0130] Then, by repeating the above processes, such as extracting solder region images Hk1 and Hk2, acquiring divided solder images Bk1 to Bk8, performing image processing on the divided solder images Bk1 to Bk8 as needed, and pasting the divided solder images Bk1 to Bk8 into the image frame W1, multiple test image data K1 to K8 are acquired from one test original image data Ik. At this time, four test image data K1, K2, K3, and K4 (hereinafter referred to as "test image data K1 to K4") are acquired for each rectangular land 3 (cream solder 5) included in the test original image data Ik. In addition, four test image data K5, K6, K7, and K8 (hereinafter referred to as "test image data K5 to K8") are acquired for each circular land 3 (cream solder 5) included in the learning original image data Ig. In this embodiment, the test image data K1 to K8 include data acquired based on two-dimensional data and data acquired based on three-dimensional data. In this embodiment, the inspection unit 78 that acquires the test image data K1 to K8 constitutes the "test image data acquisition means."

[0131] In the following step S303, a reconstructed image data acquisition step is executed. Specifically, based on a command from the main control unit 71, the inspection unit 78 inputs the inspection image data K1-K8 acquired in step S302 to the input layer of the AI ​​model 101 corresponding to the type of the inspection image data K1-K8. Therefore, the inspection image data K1-K8 acquired based on two-dimensional data is input to the AI ​​model 101 corresponding to the two-dimensional data, and the inspection image data K1-K8 acquired based on three-dimensional data is input to the AI ​​model 101 corresponding to the three-dimensional data. Then, the image data reconstructed by the AI ​​model 101 and output from the output layer is acquired as reconstructed image data. The acquired reconstructed image data is stored in association with the inspection image data K1-K8 from which the reconstructed image data was derived.

[0132] Here, when the AI ​​model 101 receives, for example, inspection image data K1 (see FIG. 27) relating to cream solder 5 having a shape abnormality, it outputs, as reconstructed image data S, image data relating to a good cream solder 5 from which the shape abnormality has been corrected (see FIG. 35, for example), based on the learning described above. Also, when the AI ​​model 101 receives, for example, inspection image data K5 (see FIG. 31) relating to cream solder 5 having a foreign matter F1 attached thereto, it outputs, as reconstructed image data S, image data relating to a good cream solder 5 from which the foreign matter F1 has been removed (see FIG. 36, for example), based on the learning described above.

[0133] On the other hand, when the AI ​​model 101 receives, for example, inspection image data K2, K6 (see FIGS. 28 and 32) relating to a non-defective cream solder 5, it outputs image data relating to the non-defective cream solder 5 that is substantially identical to the inspection image data K2, K6 as reconstructed image data S. The size (width and height) of the reconstructed image data S is the same as the size of the original inspection image data K1 to K8. In this embodiment, the inspection unit 78 that acquires the reconstructed image data S constitutes the "reconstructed image data acquisition means."

[0134] In step S304, a quality determination process is performed based on the acquired reconstructed image data S. In the quality determination process, based on a command from the main control unit 71, the inspection unit 78 compares the entire test image data K1-K8 acquired in step S302 with the entire reconstructed image data S acquired in step S303 using the test image data K1-K8, and calculates the difference between the two sets of image data K1-K8, S. For example, dots (pixels) at the same coordinates in the two sets of image data K1-K8, S are compared, and the area (number of dots) of dot clusters where the difference in brightness is equal to or greater than a predetermined value is calculated. Note that there is a relatively large difference between the test image data K1, K5 containing an abnormality and the reconstructed image data S constructed from the test image data K1, K5. In this embodiment, the inspection unit 78, which compares the test image data K1-K8 and the reconstructed image data S, constitutes the "comparison means."

[0135] Next, the inspection unit 78 determines whether the calculated difference is smaller than a predetermined threshold value. If the calculated difference is smaller than the predetermined threshold value, the inspection unit 78 determines the product as a "good product," whereas if the difference is larger than the predetermined threshold value, the inspection unit 78 determines the product as a "defective product."

[0136] Furthermore, the inspection unit 78 makes the above-mentioned judgment for all the test image data K1 to K8 relating to the inspection area of ​​the printed circuit board 1, and if the inspection unit 78 judges all the test image data K1 to K8 to be "good", it judges the inspection area to be "good" and stores this result in the storage unit 57. On the other hand, if the inspection unit 78 makes the above-mentioned judgment for all the test image data K1 to K8 relating to the inspection area and judges at least one of the test image data K1 to K8 to be "defective", it judges the inspection area to be "defective" and stores this result in the storage unit 57.

[0137] Then, if the above inspection process is performed on all inspected areas of the printed circuit board 1 and all inspected areas are judged to be "good," the solder inspection device 13 judges that the printed circuit board 1 has no abnormalities in the cream solder 5 (pass judgment) and stores this result in the memory unit 57.

[0138] On the other hand, if there is even one inspected area that is judged to be "defective," the solder inspection device 13 judges that the printed circuit board 1 has an abnormality in the cream solder 5 (failure judgment), stores this result in the memory unit 57, and notifies the outside world of this fact via the display unit 56, communication unit 58, etc.

[0139] As described above in detail, according to this embodiment, the inspection image data K1 to K8 are formed by providing divided solder images Bk1 to Bk8 in the image frame W1. Therefore, the size (width and height) of the inspection image data K1 to K8 matches the size of the image frame W1 and is constant without varying slightly depending on the size of the land 3. This eliminates the need to prepare multiple AI models 101 that differ for each size of land 3, thereby reducing the effort and time required to obtain the AI ​​model 101. Furthermore, the AI ​​model 101 can be commonly used even when the size of the land 3 is different.

[0140] Furthermore, the image frames W1 of the learning data G1-G8 and the image frames W1 of the test image data K1-K8 are the same size, and the sizes of the learning data G1-G8 and the test image data K1-K8 are the same. Therefore, when the test image data K1-K8 are input to the AI ​​model 101, appropriate reconstructed image data S corresponding to the test image data K1-K8 can be output more reliably, and the quality of the cream solder 5 can be more accurately determined. This makes it possible to more reliably obtain good test accuracy.

[0141] Additionally, the training data G1-G8 are each formed by dividing the solder region images Hg1 and Hg2 into four parts, resulting in divided solder images Bg1-Bg8, each of which is placed in the image frame W1. Therefore, compared to using image data of the entire cream solder 5 as the training data G1-G8, an AI model 101 with desired accuracy can be obtained with less training data. Furthermore, because four pieces of training data G1-G4 and G5-G8 can be obtained from the solder region images Hg1 and Hg2, respectively, the required number of pieces of training data G1-G8 can be easily obtained. These combined advantages significantly reduce the effort and time required to obtain the AI ​​model 101.

[0142] In addition, since the learning data G1-G8 and the inspection image data K1-K8 are formed by placing divided solder images Bg1-Bg8, Bk1-Bk8 in the image frame W1, the size of both data G1-G8, K1-K8 can be made relatively small compared to when the image of the entire cream solder 5 is placed in the image frame W1. This reduces the processing load related to learning and inspection, and enables learning and inspection to be performed more quickly.

[0143] Furthermore, the divided solder images Bg1-Bg8 and Bk1-Bk8 in the learning data G1-G8 and the inspection image data K1-K8 are set so that the two sides corresponding to the virtual cross lines Xga, Xgb, Xka, and Xkb and the angle formed by these two sides are oriented in a predetermined direction. This improves learning efficiency and reduces the effort and time required to obtain the AI ​​model 101. Furthermore, the accuracy of the reconstructed image data S output from the AI ​​model 101 is improved, making it possible to further increase inspection accuracy.

[0144] Furthermore, the two sides corresponding to the virtual cross lines Xga, Xgb, Xka, and Xkb and the corners formed by these two sides are common to all of the learning data G1 to G8 and the inspection image data K1 to K8, and therefore these do not become feature parts. Therefore, the accuracy of the reconstructed image data S reconstructed by the AI ​​model 101 can be further improved, and the comparison (good / bad judgment process) by the inspection unit 78 can be performed more accurately. As a result, even better inspection accuracy can be obtained.

[0145] In addition, the rectangular images Sga, Sgb, Ska, and Ska are circumscribing the solder region images Hg1, Hg2, Hk1, and Hk2 and are rectangles with the smallest area, so that even if the solder region images Hg1, Hg2, Hk1, and Hk2 are rectangular or circular, the solder region images Hg1, Hg2, Hk1, and Hk2 can be appropriately divided into four. Second Embodiment Next, the second embodiment will be described, focusing on the differences from the first embodiment. In the first embodiment, the solder region images Hg1 and Hg2 are each divided into four parts to obtain the learning data G1 to G8. In contrast, in the second embodiment, the solder region images Hg1 and Hg2 are each divided into two parts to obtain the learning data G1 to G4 (see FIGS. 39 to 42).

[0146] To explain the acquisition of the learning data G1 to G4 in more detail, first, solder area images Hg1 and Hg2 are extracted from the learning original image data Ig, as in the first embodiment. Next, divided solder images Bg1 to Bg4 (see FIGS. 37 and 38) are acquired from the solder area images Hg1 and Hg2. The divided solder images Bg1 and Bg2 are images obtained by dividing one solder area image Hg1 into two by a virtual line Lga (see FIG. 37) that can divide the image into two identical shapes. Furthermore, the divided solder images Bg3 and Bg4 are images obtained by dividing one solder area image Hg2 into two by a virtual line Lgb (see FIG. 38) that can divide the image into two identical shapes.

[0147] The imaginary line Lga is made up of a center line Cga that connects the centers of two parallel sides of the circumscribing rectangle Sga. The imaginary line Lgb is made up of a center line Cgb that connects the centers of two parallel sides of the circumscribing rectangle Sgb. As described above, the imaginary lines Lga and Lgb are capable of dividing the solder area images Hg1 and Hg2 into two identically shaped parts. However, the phrase "capable of dividing into two identically shaped parts" more precisely means that when the cream solder 5 is printed in an ideal shape (i.e., a designed shape), the solder area image relating to the cream solder 5 can be divided into two identically shaped parts (the same applies to the imaginary lines Lka and Lkb described below).

[0148] Next, by performing image processing (rotation and / or flipping) on ​​the divided solder images Bg1 to Bg4 that require image processing (in the second embodiment, divided solder images Bg2 and Bg4), the sides of the divided solder images Bg1 to Bg4 that correspond to the virtual lines Lga and Lgb are set to face in a predetermined direction (see the lower diagrams of FIGS. 37 and 38). For example, the divided solder images Bg1 to Bg4 are set so that the sides that correspond to the center lines Cga and Cgb face downward.

[0149] Then, by pasting the divided solder images Bg1 to Bg4 into the image frame W1, training data G1 to G4 (see FIGS. 39 to 42) in which the divided solder images Bg1 to Bg4 are provided in the image frame W1 are obtained. Note that training of the AI ​​model 101 using the training data G1 to G4 is performed in the same manner as in the first embodiment.

[0150] Furthermore, since the learning data G1 to G4 are obtained based on the solder area images Hg1 and Hg2 divided into two as described above, in this second embodiment, the inspection image data K1 to K4 (see Figures 45 to 48) are obtained based on the solder area images Hk1 and Hk2 divided into two.

[0151] To explain in more detail how the inspection image data K1 to K4 are obtained, first, solder area images Hk1 and Hk2 are extracted from the original inspection image data Ik, as in the first embodiment. Next, divided solder images Bk1 to Bk4 (see FIGS. 43 and 44) ​​are obtained from the solder area images Hk1 and Hk2. The divided solder images Bk1 and Bk2 are images obtained by dividing one solder area image Hk1 into two by a virtual line Lka (see FIG. 43) that can divide the image into two identical shapes. The divided solder images Bk3 and Bk4 are images obtained by dividing one solder area image Hk2 into two by a virtual line Lkb (see FIG. 44) that can divide the image into two identical shapes.

[0152] The imaginary line Lka is made up of a center line Cka that connects the centers of two parallel sides of the circumscribing rectangle Ska, and the imaginary line Lkb is made up of a center line Ckb that connects the centers of two parallel sides of the circumscribing rectangle Skb.

[0153] Next, by performing image processing (rotation and / or flipping) on ​​the images (divided solder images Bk2 and Bk4 in the second embodiment) that require image processing among the divided solder images Bk1 to Bk4, the sides corresponding to the virtual lines Lka and Lkb in the divided solder images Bk1 to Bk4 are set to face in a predetermined direction (see the lower diagrams of FIGS. 43 and 44). For example, the divided solder images Bk1 to Bk4 are set so that the sides corresponding to the center lines Cka and Ckb face downward.

[0154] Then, by pasting the divided solder images Bk1 to Bk4 respectively onto the image frame W1, inspection image data K1 to K4 (see FIGS. 45 to 48) are obtained in which the divided solder images Bk1 to Bk4 are arranged in the image frame W1. Note that the inspection of the cream solder 5 using the inspection image data K1 to K4 is carried out in the same manner as in the first embodiment.

[0155] As described above, according to the second embodiment, the size (width and height) of the inspection image data K1 to K4 matches the size of the image frame W1 and is constant without minute variations depending on the size of the land 3. This eliminates the need to prepare a large number of different AI models 101 for each size of the land 3, thereby reducing the effort and time required to obtain the AI ​​model 101. Furthermore, even if the size of the land 3 is different, the AI ​​model 101 can be commonly used.

[0156] Furthermore, since the learning data G1 to G4 and the test image data K1 to K4 have the same size, when the test image data K1 to K4 are input to the AI ​​model 101, appropriate reconstructed image data S corresponding to the test image data K1 to K4 can be output more reliably. This makes it possible to more accurately determine whether the cream solder 5 is good or bad, and more reliably obtain good test accuracy.

[0157] Furthermore, the training data G1-G4 are each formed by dividing the solder region images Hg1 and Hg2 into two parts, resulting in divided solder images Bg1-Bg4, each of which is provided in an image frame W1. Therefore, compared to using image data of the entire cream solder 5 as training data, the AI ​​model 101 can be obtained with fewer pieces of training data G1-G4. Furthermore, because two pieces of training data can be obtained from each of the solder region images Hg1 and Hg2, the required number of pieces of training data G1-G4 can be easily obtained. These combined advantages effectively reduce the effort and time required to obtain the AI ​​model 101.

[0158] In addition, since the learning data G1-G4 and the inspection image data K1-K4 are formed by placing divided solder images Bg1-Bg4, Bk1-Bk4 in the image frame W1, the size of both data G1-G4, K1-K4 can be made relatively small compared to when the image of the entire cream solder 5 is placed in the image frame. This reduces the processing load related to learning and inspection, and enables learning and inspection to be performed more quickly.

[0159] Furthermore, the divided solder images Bg1-Bg4 and Bk1-Bk4 in the learning data G1-G4 and the inspection image data K1-K4 are set so that the sides corresponding to the virtual lines Lga, Lgb, Lka, and Lkb are oriented in a predetermined direction. This improves learning efficiency and reduces the effort and time required to obtain the AI ​​model 101. Furthermore, the accuracy of the reconstructed image data output from the AI ​​model 101 is improved, making it possible to further increase inspection accuracy.

[0160] In addition, since the edges corresponding to the virtual lines Lga, Lgb, Lka, and Lkb are common to all of the learning data G1 to K4 and the test image data K1 to K4, these edges are no longer characteristic features. This further improves the accuracy of the reconstructed image data reconstructed by the AI ​​model 101, and also enables more accurate comparison (pass / fail determination processing) by the test unit 78. As a result, even better test accuracy can be achieved.

[0161] (a) In the first embodiment, the imaginary cross line Xka is associated with a circumscribing rectangle Sk a that circumscribes the solder region image Hk1 and has the smallest area. Alternatively, as shown in FIG. 49, the imaginary cross line Xkc may be associated with a circumscribing rectangle Sk c that circumscribes the solder region image Hk1 and extends in the same direction as the extension direction of one rectangular land 3 corresponding to the solder region image Hk1. In this case, the imaginary cross line Xkc is composed of a first center line Ckc1 connecting the centers of two parallel sides of the circumscribing rectangle Skc and a second center line Ckc2 connecting the centers of the other two parallel sides of the circumscribing rectangle Skc. Note that the "circumscribing rectangle Sk c extending in the same direction as the extension direction of one land 3" can be rephrased as a circumscribing rectangle Sk c having two sides parallel to the two opposing sides that form the outer edge of one land 3. The "direction in which the land extends" can be identified from, for example, design data or manufacturing data (such as Gerber data).

[0162] In the second embodiment, the virtual line Lka is related to the circumscribing rectangle Sk a that circumscribes the solder region image Hk1 and has the smallest area. Alternatively, as shown in Fig. 50, the virtual line Lkc may be related to the circumscribing rectangle Sk c that circumscribes the solder region image Hk1 and extends in the same direction as the extension direction of one rectangular land 3 corresponding to the solder region image Hk1. In this case, the virtual line Lkc is formed by a center line Ckc that connects the centers of two parallel sides of the circumscribing rectangle Sk c.

[0163] By adopting the above-described circumscribing rectangle Skc, the process of determining the circumscribing rectangle Skc becomes relatively easy, which further reduces the processing load related to training and inspection, and more effectively speeds up training and inspection.

[0164] Furthermore, as shown in FIG. 51, the circumscribing rectangle Skd may have a shape that matches the outer edge of the rectangular land 3 corresponding to the solder area image Hk1. In this case, the imaginary cross line Xkd is composed of a first center line Ckd1 connecting the centers of two parallel sides of the circumscribing rectangle Skd, and a second center line Ckd2 connecting the centers of the other two parallel sides of the circumscribing rectangle Skd. Note that the "outer edge of the land 3" can be identified, for example, from design data or manufacturing data (e.g., Gerber data). Furthermore, a virtual line for dividing the solder area image into two parts may be set based on the circumscribing rectangle Skd described above.

[0165] Furthermore, the same circumscribing rectangles as the above-mentioned circumscribing rectangles Skc and Skd may be used as circumscribing rectangles for obtaining training data.

[0166] (b) In the above embodiment, the AI ​​model 101 is configured to collectively learn the learning data G1 to G4 relating to rectangular cream solder 5 printed on rectangular lands 3 and the learning data G5 to G8 relating to circular cream solder 5 printed on circular lands 3. That is, the AI ​​model 101 corresponds to each of rectangular cream solder 5 and circular cream solder 5. Alternatively, an AI model that learns the learning data G1 to G4 and corresponds to rectangular cream solder 5, and an AI model that learns the learning data G5 to G8 and corresponds to circular cream solder 5 may be separately provided. In this case, learning efficiency and inspection accuracy can be further improved.

[0167] (c) In the above embodiment, separate AI models 101 are provided, one corresponding to two-dimensional data and one corresponding to three-dimensional data, but it is also possible to provide a common AI model that corresponds to both two-dimensional data and three-dimensional data.

[0168] (d) The configuration of the AI ​​model 101 (neural network 90) and its learning method are not limited to those described in the above embodiment. For example, the neural network 90 may be configured to perform normalization or other processing on various data as needed during the learning process or the reconstructed image data acquisition process. Furthermore, the structure of the neural network 90 is not limited to that shown in FIG. 6 , and may include, for example, a pooling layer after the convolution layer 93. Of course, the number of layers of the neural network 90, the number of nodes in each layer, and the connection structure of each node may be different.

[0169] Furthermore, in the above embodiment, the AI ​​model 101 (neural network 90) is a generative model having the structure of a convolutional autoencoder (CAE), but it is not limited to this and may be a generative model having the structure of a different type of autoencoder, such as a variational autoencoder (VAE).

[0170] Furthermore, in the above embodiment, the neural network 90 is configured to learn using the error backpropagation method, but this is not limiting, and the neural network 90 may be configured to learn using various other learning algorithms.

[0171] Additionally, neural network 90 may be configured by a dedicated AI processing circuit such as an AI chip. In this case, only learning information such as parameters may be stored in storage unit 57, and the dedicated AI processing circuit may read this information and set it in neural network 90, thereby configuring AI model 101.

[0172] Additionally, in the above embodiment, the control device 33 is provided with the learning unit 77, and the neural network 90 is configured to be learned within the control device 33, but this is not limiting. For example, the learning unit 77 may be omitted, and the neural network 90 may be configured to be learned outside the control device 33, and the AI ​​model 101 (trained neural network 90) that has been trained outside may be stored in the storage unit 57.

[0173] (e) In the above embodiment, two-dimensional data and three-dimensional data are acquired as the original inspection image data Ik, but only one of the two-dimensional data and the three-dimensional data may be acquired. Also, depending on the data to be acquired, the AI ​​model 101 may be provided to correspond to only one of the two-dimensional data and the three-dimensional data.

[0174] (f) In the above embodiment, the image frame W1 is a rectangle (square) that satisfies m=n, but it may also be a rectangle that satisfies m≠n. [Explanation of symbols]

[0175] 1...printed circuit board, 3...land, 5...cream solder, 13...solder inspection device, 32d...camera (image data acquisition means), 78...inspection unit (inspection image data acquisition means, reconstructed image data acquisition means, comparison means), 101...AI model (identification means), Bg1 to Bg8, Bk1 to Bk8...divided solder images, Cga, Cgb, Cka, Ckb, Ckc...center lines, Cga1, Cgb1, Cka1, Ckb1, Ckc1, Ckd1...th First center line, Cga2, Cgb2, Cka2, Ckb2, Ckc2, Ckd2...second center line, G1 to G8...learning data, Hg1, Hg2, Hk1, Hk2...solder area image, K1 to K8...inspection image data, Lga, Lgb, Lka, Lkb, Lkc...virtual line, Sga, Sgb, Ska, Skb, Skc, Skd...circumscribing rectangle, W1...image frame, Xga, Xgb, Xka, Xkb, Xkc, Xkd...virtual cross line.

Claims

1. A solder inspection device for inspecting cream solder printed on a printed circuit board, comprising: an image data acquisition means for acquiring image data of a predetermined inspection area on the printed circuit board including the printed solder paste; a recognition means for generating the image data by training only image data relating to non-defective cream solders in a neural network having an encoding unit for extracting feature values ​​from input image data and a decoding unit for reconstructing image data from the feature values; an inspection image data acquisition means for acquiring inspection image data including an image of the cream solder to be inspected based on the image data acquired by the image data acquisition means; a reconstructed image data acquisition means for inputting the inspection image data to the identification means and acquiring reconstructed image data as reconstructed image data; a comparison means for comparing the inspection image data and the reconstructed image data; The method is configured to be able to determine whether the cream solder is good or bad based on the comparison result by the comparison means, The learning data is composed of four divided solder images obtained by dividing one solder region image showing cream solder corresponding to one land into four parts by imaginary cross lines that can divide the image into four parts of the same shape, and each divided solder image is provided in an image frame that is larger than the size of the divided solder image, The solder inspection device is characterized in that the inspection image data acquisition means acquires the inspection image data in which the solder area image in the image data acquired by the image data acquisition means is divided into four parts by the virtual cross line, and each of the four divided solder images is placed in an image frame of the same size as the image frame of the learning data.

2. The solder inspection device according to claim 1, characterized in that the divided solder images in the learning data and the inspection image data are set so that the two sides corresponding to the virtual cross line and the corner formed by the two sides are oriented in a predetermined direction.

3. The solder inspection device of claim 1, characterized in that the virtual cross line consists of a first center line connecting the centers of two parallel sides of a circumscribing rectangle that circumscribes the solder area image and has the smallest area, and a second center line connecting the centers of the other two parallel sides of the circumscribing rectangle.

4. The outer edge of one land corresponding to the solder region image has a rectangular shape in a plan view, The solder inspection device described in claim 1, characterized in that the virtual cross line consists of a first center line connecting the centers of two parallel sides in a circumscribing rectangle that circumscribes the solder area image and extends in a direction that coincides with the extension direction of one land, and a second center line connecting the centers of the other two parallel sides in the circumscribing rectangle.

5. A solder inspection device for inspecting cream solder printed on a printed circuit board, comprising: an image data acquisition means for acquiring image data of a predetermined inspection area on the printed circuit board including the printed solder paste; a recognition means for generating the image data by training only image data relating to non-defective cream solders in a neural network having an encoding unit for extracting feature values ​​from input image data and a decoding unit for reconstructing image data from the feature values; an inspection image data acquisition means for acquiring inspection image data including an image of the cream solder to be inspected based on the image data acquired by the image data acquisition means; a reconstructed image data acquisition means for inputting the inspection image data to the identification means and acquiring reconstructed image data as reconstructed image data; a comparison means for comparing the inspection image data and the reconstructed image data; The method is configured to be able to determine whether the cream solder is good or bad based on the comparison result by the comparison means, The learning data is composed of two divided solder images obtained by dividing one solder area image showing a cream solder corresponding to one land by a virtual line that can divide the image into two identical shapes, and each divided solder image is provided in an image frame that is larger than the size of the divided solder image, The solder inspection device is characterized in that the inspection image data acquisition means acquires the inspection image data in which the solder area image in the image data acquired by the image data acquisition means is divided into two by the virtual line, and each of the two divided solder images is placed in an image frame of the same size as the image frame of the learning data.

6. 6. The solder inspection device according to claim 5, wherein the divided solder images in the learning data and the inspection image data are set so that the sides corresponding to the virtual lines are oriented in a predetermined direction.

7. 6. The solder inspection device according to claim 5, wherein the virtual line is a center line connecting the centers of two parallel sides of a circumscribing rectangle that circumscribes the solder area image and has the smallest area.

8. The outer edge of one land corresponding to the solder region image has a rectangular shape in a plan view, 6. The solder inspection device according to claim 5, wherein the virtual line is a center line connecting the centers of two parallel sides of a circumscribing rectangle that circumscribes the solder area image and extends in a direction that coincides with the extension direction of one land.

Images