WO2023113773A1

WO2023113773A1 - Manufacturing defect detection augmented training sets

Info

Publication number: WO2023113773A1
Application number: PCT/US2021/063213
Authority: WO
Inventors: Xing Liu; Qian Lin
Original assignee: Hewlett Packard Development Co LP
Current assignee: Hewlett Packard Development Co LP
Priority date: 2021-12-14
Filing date: 2021-12-14
Publication date: 2023-06-22
Anticipated expiration: 2024-06-14

Abstract

In an example in accordance with the present disclosure, a computing device is described. The computing device includes a database of a training set of images. The training set includes images of defects from a manufacturing process. The computing device also includes a processor which generates an augmented training set. The processor does so by 1) annotating images of the training set to categorize certain defects of the training set, 2) isolating an identified defect depicted in an image of the training set, and 3) replicating an isolated defect to an additional image of the training set. The computing device also includes a neural network that identifies a defect in a target image by analyzing the target image against the augmented training set.

Description

MANUFACTURING DEFECT DETECTION AUGMENTED TRAINING SETS

BACKGROUND

[0001] There are numberless types of electronic devices used in the world today and each electronic device includes hundreds of hardware components. For example, a computing device includes processors, memory devices, and other integrated circuits. As a specific example, printers include printheads which include nozzles for ejecting fluid onto a substrate and the circuitry for effectuating this ejection.

BRIEF DESCRIPTION OF THE DRAWINGS

[0002] The accompanying drawings illustrate various examples of the principles described herein and are part of the specification. The illustrated examples are given merely for illustration, and do not limit the scope of the claims.

[0003] Fig. 1 is a block diagram of a computing device for detecting manufacturing defects with an augmented training set, according to an example of the principles described herein.

[0004] Fig. 2 is a flowchart of a method for detecting manufacturing defects with an augmented training set, according to an example of the principles described herein.

[0005] Fig. 3 is a flowchart of a method for detecting manufacturing defects with an augmented training set, according to an example of the principles described herein. [0006] Figs. 4A - 4D depict the generation of an augmented training set, according to an example of the principles described herein.

[0007] Fig. 5 depicts the generation of annotations for an augmented training set, according to an example of the principles described herein.

[0008] Fig. 6 depicts a non-transitory machine-readable storage medium for detecting manufacturing defects with an augmented training set, according to an example of the principles described herein.

[0009] Throughout the drawings, identical reference numbers designate similar, but not necessarily identical, elements. The figures are not necessarily to scale, and the size of some parts may be exaggerated to more clearly illustrate the example shown. Moreover, the drawings provide examples and/or implementations consistent with the description; however, the description is not limited to the examples and/or implementations provided in the drawings.

DETAILED DESCRIPTION

[0010] Electronic devices have become widespread in today’s society and it is not uncommon for an individual to interact with multiple electronic devices on a daily basis. Electronic devices include integrated circuits to carry out their intended functionality. For example, a printer may include a printhead which includes nozzles to eject fluid. These nozzles are connected via electrical traces to a computing device which triggers the sequential activation of nozzles to eject fluid in a pattern on a substrate. While particular reference is made to particular electronic device components, millions of other types of electronic devices are manufactured, with great quantities produced daily.

[0011] Historically, in factories in the world, a human operator monitors the production line to determine when a manufactured component has a defect. Some machine vision systems have been developed which rely on photos of defects which are analyzed to identify a defect. However, such machine vision systems generally rely on manually-identified features to detect a defect in the manufacturing process. [0012] Inspection and defect detection in manufacturing enhances the production quality and detects potential issues with the production line. Given this relevance, developments to inspection and defect detection may provide more efficient and effective defect detection results. For example, manual inspection relies on human examination and requires considerable experience in defect classification. As such, a manual operator may go through months of training at the production line before an operator can inspect independently. Moreover, manual methods are prone to human error, mis-identification, or a lack of identification of a manufacturing defect.

[0013] Computer vision methods combine feature extraction and machine learning to analyze the inspection images and make decisions. However, handcrafted feature extraction still relies on manual intervention for entry of the features by which a defect is identified.

[0014] Accordingly, the present specification describes a system and method that automatically learns the features of interest from a training set with initially untagged data. Rather than including manually-identified specific features, the training set of the present specification includes criteria by which a detected defect may be categorized and includes defects that have been replicated from one image to another to build the data set by which a defect in a target image is identified. As such, the manual input of features indicative of a defect and data manipulation are reduced. Accordingly, the present specification applies deep learning to industrial inspection by training a model to detect defects on electronic devices. Specifically, the present specification improves the accuracy of defect detection and does so with less human interaction.

[0015] The method may detect defects in a variety of electronic devices such as integrated circuits, printhead components, and packaging boxes.

[0016] In general, prior to operation of a neural network, the data in a training set is augmented by adding samples of certain defects such that the neural network is properly trained to detect these certain defects. Still further, in many products, the location and size of a defect are relevant. For example, a contamination defect in a printhead, i.e., where a foreign substance is found on the printhead, may be permissible if not in a particular area of the printhead which connects the printhead to a circuit. The present specification describes a computing device that allows the neural network to categorize certain defects as permissible within the predetermined area so that a time-consuming alignment step is avoided.

[0017] Specifically, the present specification describes a computing device. The computing device includes a database of a training set of images. The training set includes images of defects from a manufacturing process. The computing device also includes a processor to generate an augmented training set. The processor does so by 1) annotating images of the training set to categorize certain defects of the training set, 2) isolating an identified defect depicted in an image of the training set, and 3) replicating an isolated defect to an additional image of the training set. The computing device also includes a neural network to identify a defect in a target image by analyzing the target image against the augmented training set.

[0018] The present specification also describes a method. According to the method, a training set of images of defects from a manufacturing process are annotated to categorize certain defects depicted in the images as acceptable. An identified defect from an image is isolated and replicated in an image of the training set. The neural network is trained on this augmented data set and then identifies a defect in a target image by analyzing the target image.

[0019] The present specification also describes a non-transitory machine- readable storage medium encoded with instructions executable by a processor of a computing device. When executed by the processor, the instructions cause the processor to generate an augmented training set by annotating a training set of images of defects from an electronic circuit manufacturing process. This annotation includes categorizing certain defects depicted in the images as acceptable and categorizing certain regions of the image as non-analyzed regions excluded from defect detection by a neural network. The instructions to cause the processor to generate the augmented training set also cause the processor to 1) isolate an identified defect from an image of the training set and 2) replicate an isolated defect depicted in an image of the training set. When executed by the processor, the instructions cause the processor to train a neural network to identify a defect in a target image based on the augmented training set.

[0020] In summary, such a computing device, method, and machine- readable storage medium may, for example 1) efficiently identify defects in a variety of manufactured components; 2) avoid user error in defect detection; 3) automate defect detection; 4) create additional images of real defects to identify unique defects; and 5) categorize certain defects as permissible to train the neural network to learn characteristics of permissible defects. However, it is contemplated that the devices disclosed herein may address other matters and deficiencies in a number of technical areas, for example.

[0021] As used in the present specification and in the appended claims, the term “a number of” or similar language is meant to be understood broadly as any positive number including 1 to infinity.

[0022] Turning now to the figures, Fig. 1 is a block diagram of a computing device (100) for detecting manufacturing defects with an augmented training set (108), according to an example of the principles described herein. As described above, defect detection in a manufacturing process ensures that a resulting product meets certain manufacturing metrics and is thus satisfactory to perform an intended function. However, human-based defect identification may be susceptible to user-error and mis-identification of certain defects or a total lack of detection altogether. Some automated methods, while automating some aspects of defect detection, still largely rely on human identified features to detect a particular defect. Accordingly, the present computing device (100) includes a neural network (110) which may identify a defect with less to no user interaction. Moreover, the present computing device (100) removes certain post processing operations, thus reducing the load on the computing device (100) to post process an output of the neural network (110).

[0023] Specifically, the computing device (100) includes a database (102) which includes a training set (104) of images. The images may be of defects that resulted during a manufacturing process. These images may be acquired during the manufacturing process. That is, as a product such as a printhead is being fabricated, images of the product may be captured. These images may form the training set (104) by which defects are subsequently detected. In some examples, the defects in the training set (104) images may be tagged as such. That is, in some examples, the computing device (100) via user input or automatically, may identify a particular defect in an image and may tag the image in the training set (104) accordingly. It is this training set (104) that is relied on by the neural network (110) to detect a particular defect.

[0024] This training set (104) is passed to a processor (106) of the computing device (100). The processor (106) may extract instructions from a memory of the computing device (100) an execute the instructions to perform an intended function.

[0025] The memory may include a machine-readable storage medium, which machine-readable storage medium may contain, or store computer-usable program code for use by or in connection with an instruction execution system, apparatus, or device. The memory may include many types of memory including volatile and non-volatile memory. For example, the memory may include Random Access Memory (RAM), Read Only Memory (ROM), optical memory disks, and magnetic disks, among others. The executable code may, when executed by the respective component, cause the component to implement the functionality described herein.

[0026] As described above, there may be insufficient images in the training set (104) to allow the neural network (110) to accurately detect a defect. That is, the accuracy and reliability of the neural network (110) to detect a defect is based on the number of instances of the defect in the training set (104). As such, if there is an insufficient number of images of defects, or an insufficient number of images of a particular defect, the neural network (110) may misidentify or fail to identify a particular defect. Accordingly, the processor (106) generates an augmented training set (108), which augmented training set (108) provides more samples on which the neural network (110) may rely in identifying and marking a defect in a particular target image.

[0027] Accordingly, the processor (106) may annotate images of the training set (104) to categorize certain defects of the training set (104). In one particular example, under certain conditions some manufacturing defects may be acceptable or permissible. For example, electrical signals are routed from a computing device to a printhead, where the nozzles are located, to sequentially activate the nozzles to eject fluid in a pattern to be formed on a substrate. Accordingly, connections between electrical traces on the printhead are coupled to electrical traces on a circuit. To protect against mechanical damage to these electrical traces and the connection between the two, an encapsulant, or encap, may be formed over the connection. The encap may include polymer liquid that is poured over the connection point and cured to form a hardened structure. Defects in the encap formation may negatively impact the electrical connection and/or may expose the underlying connection to mechanical damage either during manufacturing, transport or use. One particular example of a defect in an encap is a contamination defect where some foreign matter, such as dust or a small piece of metal, becomes lodged in the encap. A contamination defect in the encap structure may be acceptable if it is under a particular size. Currently, to identify these permissible small defects, a user engages in a post-processing step to compute the sizes of all detected defects and drops the defects with sizes that are smaller than the particular size. Rather than determining an acceptable size of defect as a post-neural network operation, the present computing device (100) provides annotations such that an augmented training set (108) enables the neural network (110) to learn the characteristics of the small-sized defects. Specifically, the computing device (100) may provide an annotation, or metadata, associated with an image of the training set (104). The annotation may indicate that the particular defect is acceptable, for example based on its size and/or location on the product being produced. As such, the neural network (110) upon detecting a defect, may rely on this annotation to determine that similar defects are permissible.

[0028] By annotating, or classifying some of the defects in the training set (104) of images as permissible, the operation of the neural network (110) is improved as less than all of the defects are analyzed. That is, as certain defects are identified as permissible, when a defect in a target image is identified as being similar to a permissible defect, the neural network (110) may terminate subsequent analysis and classification of this permissible defect. [0029] While particular reference is made to an annotation which indicates permissibility based on a size of the defect, other annotations may be implemented as well. For example, a defect may be acceptable based on a location of the defect. Accordingly, in this example, the processor (106) may annotate defects as permissible based on the location of the defect. Returning to the printhead example, as long as the defect does not touch a certain area, such as a nozzle area, the defect may not impact printhead performance and as such may be permissible. In either example, by providing these annotations, the neural network (110) may be trained to not identify permissible defects as defects that require attention such as discarding the printhead and/or repairing the printhead.

[0030] In addition to annotating pass criteria for the training set (104), the augmented training set (108) may include other metadata. For example, it may be that there is inadequate data in a manufacturing facility. That is, some defects may occur infrequently enough that a neural network (110) may not be able to identify the defect. That is, the neural network (110) may identify a defect in a target image based on an analysis of thousands, if not more, images of defects in the training set (104). Accordingly, the present computing device (100) provides for the increased accuracy and reliability in the identification of certain defects, including those that occur infrequently.

[0031] Accordingly, the processor (106) isolates an identified defect depicted in an image of the training set (104) and replicates the isolated defect to an additional image of the training set (104). For example, within an image of the training set (104), a defect may be identified. This detected defect may be cropped from the image and inserted either into 1) a new image that is added to the augmented training set (108) or 2) an existing image of the training set with a second identified object. Such isolation and replication provide another sample on which the neural network (110) may rely on in identifying a defect in a target image. While particular reference is made to isolating and replicating a single identified defect, the isolation and replication may be performed for multiple identified defects in the training set (104) such that more data points are relied on in detecting defects in target images. As will be described below, in some examples, in addition to isolating and replicating the identified defect, the identified defect may be altered so as to provide variation to the defect which aids the neural network (110) to more accurately and reliably identify defects in target images.

[0032] In some examples, such isolation and replication may be performed for a subset of the images in the training set (104). That is, as described above, some defects occur infrequently such that there are less images of these particular defects. As the number of occurrences of a defect in the training set (104) impacts the reliability with which the neural network (110) can identify a subsequent occurrence of the defect in a target image, the low quantity of instances of infrequently occurring defects may impact the reliability of the neural network (110) to identify such. Accordingly, the subset of images which are isolated and replicated may be those images of the training set (104) that depict infrequently occurring defects. For example, if a particular defect occurs on the order of 2-10 times less often than another defect, it may be deemed as an infrequently occurring defect for which isolation and annotation are performed. By comparison, images which depict frequently occurring defects, i.e. , 2-10 times more than another defect, may not be isolated and replicated, even though they may be annotated.

[0033] The computing device (100) also includes a neural network (110) which as described above identifies a defect in a target image by analyzing the target image against the augmented training set (108). As described above, the neural network (110) analyzes characteristics of the target image and compares it against characteristics of the images in the augmented training set (108) to identify those features which are indicative of a defect. By relying on the augmented training set (108) and associated annotations, features indicative of a defect are identified, while those features that are associated with an acceptable defect are not flagged as indicating a defect.

[0034] In general, a neural network (110) includes a mapping from a given input to underlying output. It is an end-to-end structure which has multiple layers. Each layer conducts certain operation on the received inputs. Example layers include a convolutional layer, a max pooling layer, and an activation layer. By stacking multiple layers of operations, a deep neural network becomes non-linear and is able to simulate complicated mapping functions between the input and output.

[0035] Fig. 2 is a flowchart of a method (200) for detecting manufacturing defects with an augmented training set (Fig. 1 , 108), according to an example of the principles described herein. According to the method (200), a training set (Fig. 1 , 104) of images of defects from a manufacturing process are annotated (block 201). The annotations may include a variety of information. For example, the annotations may indicate a type of defect. As specific examples, the annotations may indicate whether a defect in an encap is a contamination defect, a blowhole defect, or a microcrack defect, among a variety of other types. In one particular example, the annotations indicate whether an identified defect is acceptable. For example, certain defects may occur in a manufacturing process, but may not impact the performance of the component, or may impact the performance of the component in a negligible way. This may be the case when a defect is smaller than a predetermined size or located outside a particular region. In this example, the annotations (block 201) prevent the neural network (Fig. 1 , 110) from identifying these defects as ones that require additional attention or remedial action.

[0036] An identified defect, which may be an infrequently occurring defect, may be isolated (block 202) from an image in the training set (Fig. 1 , 104) where it is found and may be replicated (block 203) to another image in the training set (Fig. 1 , 104). As described above, doing so may provide more data of known defects to a neural network (Fig. 1 , 110) which can then more accurately and reliably identify a defect in a subsequent target image.

[0037] Relying on the augmented training set (Fig. 1 , 108), the neural network (Fig. 1 , 110) identifies (block 204) a defect in a target image by analyzing the target image against the augmented training set (Fig. 1 , 108). That is, the neural network (Fig. 1 , 110) may be trained on the augmented training set (Fig. 1 , 108) and the neural network (Fig. 1 , 110) based on this training may identify a defect in a target image. As described above, the neural network (Fig. 1 , 110) may receive an image of a product that is generated as the product is being manufactured. Such an image may be referred to as a target image. By comparing features of the target image against features of the augmented training set (Fig. 1 , 108), the neural network (Fig. 1 , 110) may be able to identify and classify a defect captured in the target image. As the augmented training set (Fig. 1 , 108) includes an increased quantity of images of at least some defects, certain defects may be more reliably identified.

Moreover, as the augmented training set (Fig. 1 , 108) includes annotations regarding the acceptability of certain defects, the neural network (Fig. 1 , 110) may properly classify a defect in a target image as one that needs remedial action, as opposed to drawing attention to a defect that does not impact product performance.

[0038] Fig. 3 is a flowchart of a method (300) for detecting manufacturing defects with an augmented training set (Fig. 1 , 108), according to an example of the principles described herein. According to the method (300), a training set (Fig. 1 , 104) of defects are annotated (block 301) to identify certain defects as acceptable. Specifically, the processor (Fig. 1 , 106) may categorize certain defects depicted in the images as acceptable as described above.

[0039] In another example, the processor (Fig. 1 , 106) may categorize certain regions of the images of the training set (Fig. 1 , 104) as non-analyzed regions that are excluded from defect detection by the neural network (Fig. 1 , 110). For example, in some manufacturing process, a defect may be permissible if found outside a particular region. Returning to the printhead encap example, a contamination defect may be permissible if it is outside of the encap region. However, an image of the encap may also capture other structural components. The neural network (Fig. 1 , 110) may identify defects in these other structural components, which defects do not impact the performance of the product. In another example, the neural network (Fig. 1 , 110) may identify elements in these other structural components that are not defects, but that may have characteristics similar to a defect such that the neural network (Fig. 1 , 110) may improperly identify these other non-defect elements as a defect. As such, the neural network (Fig. 1 , 110) may be improperly trained to accurate identify defects, which improper training may lead to misidentification of target defects.

[0040] In some examples, a user may manually align the images in the training set (Fig. 1 , 104) and crop out the other structural components. However, such manual cropping and exclusion is time-consuming and may not be possible with certain images of the training set (Fig. 1 , 104). Still further, in some cases such as with encap, the area to be analyzed may not be regularshaped and its boundary may vary from image to image.

[0041] Accordingly, rather than relying on manual identification of a region to be operated on by the neural network (Fig. 1 , 110), the processor may categorize the region to be analyzed so that the neural network (Fig. 1 , 110) can recognize the region to be analyzed at a pixel level. As a result, the exact location and boundary of the region of the image to be analyzed is obtained. A post processing step may then eliminate false alarms outside of the to-be- analyzed region by comparing the location of each detected defect to the location of the non-analyzed region. With this method, cropping out the region to be analyzed is conducted at the same time as when defect recognition is conducted, thus saving computational time over an otherwise lengthy alignment step.

[0042] As described above, an identified defect depicted in an image of the training set (Fig. 1 , 104) may be isolated and replicated. Specifically, isolating the identified defect includes cropping (block 302) the identified defect from the image in the training set (Fig. 1 , 104) and applying (block 303) a defect mask over the identified defect. The cropped defect mask may then be altered (block 304). That is, the depiction of the defect may be altered such that not only are there more instances of a particular defect in the augmented training set (Fig. 1 , 108), but the appearance of the defect is adjusted such that similar defects, but with difference appearances may still be identified.

[0043] The alterations that may be performed may have a variety of forms. In some examples, defects in the production line may be specific. Accordingly, when there are not sufficient samples of one defect type, the computing device (Fig. 1 , 100) may apply certain kinds of alterations while avoiding others. For example, changing brightness may turn one defect type into another. As another example, stretching the defect may distort the appearance of the defect. Accordingly, the type of alteration that is performed may be based on the type of defect. Examples of alterations include altering a size of the identified defect, altering a position of the identified defect, or rotating the identified defect. In an example, low-frequency noise may be added to the identified defect. For example, certain modifications such as changing the brightness of certain areas of the defect may be made. Doing so may increase the scope of the augmented training set (Fig. 1 , 108) that the neural network (Fig. 1 , 110) operates on. While particular reference is made to particular alterations, other alterations may be implemented, which other alterations may be based on the type of identified defect. Once altered (block 304), the defect may be replicated (block 305) in another image of the training set as described above in connection with Fig. 2.

[0044] A particular example of defect isolation, alteration, and replication is now provided. In an example, an identified defect is cropped from a sample image from the training set (Fig. 1 , 104) and a mask applied over the defect. Distortions such as resizing, rotating, and adding low-frequency noise to the defect mask are then applied. The altered defect may then be blended into an additional image of the training set (Fig. 1 , 104), which additional image may be a newly generated image with the altered defect or may be an existing image. In this later example, the altered defect may be present in the image alongside a second identified defect which may or may not itself be altered. Following generation of the augmented training set (Fig. 1 , 108), a defect in a target image is identified (block 306) as described above in connection with Fig. 2.

[0045] Figs. 4A - 4D depict the generation of an augmented training set (Fig. 1 , 108), according to an example of the principles described herein. As described above, a printhead may include an encap (412) which protects electrical connections between electrical traces (416) of the printhead die and electrical traces (414) of a connecting circuit. The electrical traces (414, 416) are depicted in Fig. 4A as dashed lines due to their positioning underneath the encap (412). As depicted in Fig. 4B, a defect (418) may occur in the encap (412). Such a defect may include a microcrack in the encap surface or a depression in the encap which may be a result of the manufacturing process. Accordingly, the electronic circuit that is produced and for which defects are identified by the neural network (Fig. 1 , 110) may be a printhead component and specifically an encap (412) formed over electrical contacts.

[0046] As described above and as depicted in Fig. 4C, the defect (418) may be isolated, i.e. , cropped from the image of the encap (412) and may be altered. For example, as depicted in Fig. 4C, the defect (418) is enlarged and rotated. Changing the size and orientation of the defect (418) may provide more features that can be analyzed by the neural network (Fig. 1 , 110) to identify a defect in a target image.

[0047] As depicted in Fig. 4D, the defect (418) may be replicated on an image of another encap (412). In the example depicted in Fig. 4D, the enlarged and rotated defect (418) is re-positioned on the second image of the encap (412).

[0048] Fig. 5 depicts the generation of annotations for an augmented training set (Fig. 1 , 108), according to an example of the principles described herein. As described above, certain defects may be acceptable based on any number of criteria such as a size and/or location of the defect (418). Fig. 5 depicts a variety of defects (418) some of which may be classified as permissible or acceptable.

[0049] Specifically, Fig. 5 depicts a nozzle region (520) of a printhead. It may be the case that defects (418) within the nozzle region (520) that are smaller than a predetermined size may be acceptable in that they do not impact performance of the nozzles while other defects (418) in the nozzle region (520) may be flagged for remedial measure. For example, a first defect (418-1) may be identified as a defect for which remedial action is proscribed. By comparison, a second defect (418-2) may include an annotation (522-1) categorizing the second defect (418-2) as acceptable based on the size of the second defect (418-2) being less than a predetermined size. Similarly, a third defect (418-3) may include an annotation (522-2) categorizing the third defect (418-3) as acceptable based on the location of the third defect (418-3) being outside a predetermined region of the image. Accordingly, categorization of a defect as acceptable may be based on a location of the defect and/or a size of the defect. While particular reference is made to defect categorization based on size and/or location, other criteria may be used and other annotations provided which indicate acceptable defects.

[0050] In a test, the computing device (Fig. 1 , 100) implementing these methods had high precision, recall, and F1 scores regarding the identification and detection of defects in the encap, die, and cover layer. Further in a test, the process time for each image was a fraction of a second. As such, the method (200) may be performed in a production line during production of a particular product.

[0051] Fig. 6 depicts a non-transitory machine-readable storage medium (624) for detecting manufacturing defects (Fig. 4, 418) with an augmented training set (Fig. 1 , 108), according to an example of the principles described herein. To achieve its desired functionality, the computing device (Fig. 1 , 100) includes various hardware components. Specifically, the computing device (Fig. 1 , 100) includes a processor (Fig. 1 , 106) and a machine-readable storage medium (624). The machine-readable storage medium (624) is communicatively coupled to the processor (Fig. 1 , 106). The machine-readable storage medium (624) includes a number of instructions (626, 628, 630, 632) for performing a designated function. In some examples, the instructions may be machine code and/or script code.

[0052] The machine-readable storage medium (624) causes the processor to execute the designated function of the instructions (626, 628, 630, 632). The machine-readable storage medium (624) can store data, programs, instructions, or any other machine-readable data that can be utilized to operate the computing device (Fig. 1 , 100). Machine-readable storage medium (624) can store machine readable instructions that the processor (Fig. 1 , 106) of the computing device (Fig. 1 , 100) can process, or execute. The machine-readable storage medium (624) can be an electronic, magnetic, optical, or other physical storage device that contains or stores executable instructions. Machine- readable storage medium (624) may be, for example, Random-Access Memory (RAM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), a storage device, an optical disc, etc. The machine-readable storage medium (624) may be a non-transitory machine-readable storage medium (624).

[0053] Referring to Fig. 6, annotate instructions (626), when executed by the processor (Fig. 1 , 106), cause the processor (Fig. 1 , 106) to, annotate a training set (Fig. 1 , 104) of images of defects from an electronic circuit manufacturing process by 1 ) categorizing certain defects depicted in the images as acceptable and 2) categorizing certain regions of the image as non-analyzed regions excluded from defect detection by a neural network (Fig. 1 , 100). Isolate instructions (628), when executed by the processor (Fig. 1 , 106), cause the processor (Fig. 1 , 106) to, isolate an identified defect from an image in the training set (Fig. 1 , 104). Replicate instructions (630), when executed by the processor (Fig. 1 , 106), cause the processor (Fig. 1 , 106) to replicate an isolated defect depicted in an image of the training set (Fig. 1 , 104). Train instructions (632), when executed by the processor (Fig. 1 , 106), also cause the processor (Fig. 1 , 106) to train a neural network (Fig. 1 , 110) to identify a defect in a target image based on the augmented training set.

[0054] In summary, such a computing device, method, and machine- readable storage medium may, for example 1) efficiently identify defects in a variety of manufactured components; 2) avoid user error in defect detection; 3) automate defect detection; 4) creates additional images of real defects to identify unique defects; and 5) categorize certain defects as permissible to train the neural network to learn characteristics of permissible defects. However, it is contemplated that the devices disclosed herein may address other matters and deficiencies in a number of technical areas, for example.

Claims

CLAIMS What is claimed is:

1 . A computing device, comprising: a database comprising a training set of images, the training set comprising images of defects from a manufacturing process; a processor to generate an augmented training set by: annotating images of the training set to categorize certain defects of the training set; isolating an identified defect depicted in an image of the training set; and replicating an isolated defect to an additional image of the training set; and a neural network to identify a defect in a target image by analyzing the target image against the augmented training set.

2. The computing device of claim 1 , wherein the neural network is to identify from the augmented training set and associated annotations, features indicative of a defect.

3. The computing device of claim 1 , wherein the additional image is a new image added to the training set.

4. The computing device of claim 1 , wherein the additional image is an existing image of the training set with a second identified defect.

5. The computing device of claim 1 , wherein the processor is to annotate a subset of the images of the training set.

6. The computing device of claim 5, wherein the subset comprises images of the training set with infrequently occurring defects.

7. A method, comprising: generating an augmented training set by: annotating a training set of images of defects from a manufacturing process to categorize certain defects depicted in the images as acceptable; isolating an identified defect from an image in the training set; replicating an isolated defect depicted in an image of the training set; and identifying a defect in a target image by analyzing the target image against the augmented training set.

8. The method of claim 7, wherein isolating the identified defect comprises: cropping the identified defect from the image in the training set; and applying a defect mask over the identified defect.

9. The method of claim 8, further comprising altering a cropped identified defect.

10. The method of claim 9, wherein altering the cropped identified defect comprises: altering a size of the identified defect; altering a position of the identified defect; rotating the identified defect; adding a low-frequency variation to the identified defect; or a combination thereof.

11. A non-transitory machine-readable storage medium encoded with instructions executable by a processor of a computing device to, when executed by the processor, cause the processor to: generate an augmented training set by: annotating a training set of images of defects from an electronic circuit manufacturing process by: categorizing certain defects depicted in the images as acceptable; and categorizing certain regions of the image as non-analyzed regions excluded from defect detection by a neural network; isolate an identified defect from an image of the training set; replicate an isolated defect depicted in an image of the training set; and train a neural network to identify a defect in a target image based on the augmented training set.

12. The non-transitory machine-readable storage medium of claim 11 , wherein a categorization of a certain defect is acceptable is based on a location of the certain defect.

13. The non-transitory machine-readable storage medium of claim 11 , wherein a categorization of a certain defect is acceptable is based on a size of the certain defect.

14. The non-transitory machine-readable storage medium of claim 11 , wherein the electronic circuit is a printhead component.

15. The non-transitory machine-readable storage medium of claim 14, wherein the printhead component is an encap formed over electrical contacts.

19