Detecting non-uniform defects in the main panel of flat panel displays during manufacturing

Abstract

A method for detecting non-uniform defects in a master panel during manufacturing is provided, the master panel comprising a plurality of flat screen displays. The method comprises: preparing a combined image from image data of the master panel; enhancing the quality of the combined image, including removing artifacts from the combined image; filtering the quality-enhanced combined image to detect local non-uniform defects, the local non-uniform defects comprising at least one structured pattern of a defined geometric shape; applying different candidate patterns to the filtered combined image; selecting one of the candidate patterns as a defect detection pattern, the defect detection pattern being closest to the structured pattern of the defined geometric shape of the detected local non-uniform defect; and displaying at least a portion of the defect detection pattern and the quality-enhanced combined image on a display.

Description

Technical Field

[0001] The present application relates to a method for detecting non-uniform defects in a main panel during manufacturing and a method for detecting non-uniform defects in a flat panel display before separation from a main panel (105) during a manufacturing process. Background Art

[0002] Flat panel displays such as liquid crystal displays (LCDs) are used in many electronic devices such as cell phones, televisions, and computer monitors. They are typically manufactured by forming a thin film transistor (TFT) array on a master panel (or substrate) (where the TFT array corresponds to a plurality of flat panel displays) and eventually separating the master panel into individual flat panel displays. In a production line, a large number of master panels are processed every day.

[0003] Flat panel displays are manufactured in multiple processing steps. In each step, chemical and mechanical surface changes occur, some of which are considered defects, including mura defects. In order to ensure product quality (panel yield), the surface of the main panel must be repeatedly (e.g., pixel by pixel) checked between processing steps. Inspection can be performed using an inspection system (such as the Keysight 88000 HS-100 series array test system available from, for example, Keysight Technologies, Inc). The inspection generates data in the form of a matrix corresponding to the pixels on the main panel. The data matrix typically includes noise from multiple sources within both the environment and the inspection hardware. Noise should be suppressed to provide a clear image of the measurement data to the user and allow automatic detection of panel defects.

[0004] Non-uniform defects are particularly likely to appear in the master panel before the individual flat panel displays are separated. For example, in the case of LCDs, non-uniform defects are often caused by process defects associated with cell assembly, which affect light transmission through the flat panel display and are generally annoying to the observer. An inspection system can be put in place to provide measurements of the flat panel display that produce information about defects that may have been introduced in individual manufacturing steps. For example, the Keysight 88000 HS-100 series array test system produces a single-pixel resolution charge map of a flat panel display that can be used to identify defects. Given a larger data set, certain defect patterns can be attributed to specific production equipment. However, the resulting gigapixel data set poses a challenge to visual inspection and annotation. Therefore, there is a need for an efficient automated system and method for detecting, classifying and quantifying certain defect types (including non-uniform defects), for example based on Fourier transforms. Automated inspection and quality control can reduce labor costs, thereby achieving greater standardization. Summary of the invention

[0005] This application involves the following technical solutions:

[0006] 1. A method of detecting non-uniform defects in a main panel during manufacturing, the main panel comprising a plurality of flat screen displays, the method comprising:

[0007] preparing a combined image from the image data of the primary panel;

[0008] enhancing a quality of the combined image, wherein enhancing the quality comprises removing artifacts from the combined image;

[0009] filtering the quality enhanced combined image to detect localized inhomogeneity defects comprising at least one structured pattern of a defined geometry;

[0010] applying a plurality of different candidate patterns to the filtered combined image;

[0011] selecting one of the plurality of candidate patterns as a defect detection pattern, selecting a defect detection pattern from the plurality of candidate patterns that is closest to the structured pattern of the defined geometric shape of the detected local non-uniform defect; and

[0012] At least a portion of the defect detection pattern and the quality enhanced combined image are displayed on a display to show the location of the detected local non-uniform defect in the structured pattern of defined geometry.

[0013] 2. The method of claim 1, wherein filtering the quality enhanced combined image to detect local non-uniform defects comprises:

[0014] filtering out relevant spatial frequencies corresponding to the length scale of the detected localized inhomogeneity defects to provide a first filtered image; and

[0015] The first filtered image is convolved using a set of templates corresponding to the defined geometric shapes respectively.

[0016] 3. The method of claim 1 , wherein the defined geometric shape comprises a spot and / or a plurality of rings.

[0017] 4. A method for detecting non-uniform defects in a flat panel display before separation from a main panel during a manufacturing process, the method comprising:

[0018] obtaining a plurality of images of the master panel and storing image data corresponding to at least some of the plurality of images, wherein the image data comprises pixel signals received from a matrix of pixels on the master panel;

[0019] combining the stored image data of the plurality of images to provide a combined image;

[0020] filtering the combined image to detect inhomogeneous defects using at least one defect pattern specific template to detect a structured pattern of defined geometry corresponding to the inhomogeneous defect;

[0021] Creating candidate patterns, each of which has a two-dimensional coordinate system of different sizes, the two-dimensional coordinate system attempting to demonstrate the location of the detected non-uniform defect;

[0022] selecting a defect detection pattern and a scanning offset that most closely corresponds to the location of the detected non-uniform defect from the created candidate patterns, the selected defect detection pattern providing the strongest signal from the total filter response of the filtered combined image;

[0023] fine-tuning the selected defect detection pattern; and

[0024] The detected non-uniform defects are quantified using the fine-tuned defect detection pattern.

[0025] 5. The method of item 4, wherein each of the candidate patterns includes horizontal and vertical periodicity of the detected non-uniform defect and horizontal and vertical scanning offsets, thereby creating a rectangular pattern.

[0026] 6. The method of item 5, wherein at least one of the candidate patterns is rotated by introducing a rotation angle parameter.

[0027] 7. The method of item 4, wherein obtaining the multiple images of the main panel includes scanning the main panel using a scanning head, wherein electrical signals or optical signals from the scanning head are respectively connected to pixels on the main panel in a predetermined order to provide the pixel signals.

[0028] 8. The method of claim 4, wherein the primary panel comprises a glass substrate.

[0029] 9. The method of claim 4, wherein combining the stored image data of the plurality of images to provide the combined image comprises:

[0030] downsampling the plurality of images to reduce storage requirements for storing the image data;

[0031] Arranging the downsampled multiple images into a two-dimensional pattern, and combining the two-dimensional pattern into a single larger image;

[0032] Correcting the contrast and background levels of the downsampled multiple images of the single larger image to provide the combined image to appear uniform; and

[0033] The marking is performed in a separate binary matrix region of the combined image consisting of actual data, rather than in an unknown region, wherein the binary matrix region is used to weight a filter response from filtering the combined image.

[0034] 10. The method of claim 9, further comprising cleaning the combined image before filtering the combined image to detect the non-uniform defect, the cleaning the combined image comprising:

[0035] removing measurement artifacts from the stored image data before combining the two-dimensional patterns into the single larger image; and

[0036] Artifacts are suppressed in individual frames of this single larger image before correcting contrast and background levels.

[0037] 11. The method of claim 10, wherein suppressing artifacts in individual frames of the single larger image comprises removing local overshoot and / or undershoot of signals corresponding to at least some of the multiple images in corners of the stored image data.

[0038] 12. The method of claim 10, wherein removing the measurement artifact from the stored image data comprises:

[0039] Remove noise originating from electronic amplifier components;

[0040] Use a low-pass filter to remove white noise; and

[0041] The image is flattened by subtracting the local mean extracted by convolving the image with a Gaussian kernel.

[0042] 13. The method of claim 12, wherein removing the noise originating from the electronic amplifier components comprises:

[0043] Remove drift by subtracting the sliding mean;

[0044] Gain variations are removed by normalizing the signal strength to unity;

[0045] Remove crosstalk between amplifier channels; and

[0046] Narrowband noise is removed by applying a narrowband spatial frequency filter.

[0047] 14. The method of claim 4, wherein before filtering with the at least one defect pattern specific template, the method further comprises:

[0048] filtering the combined image with a kernel that selects the relevant spatial frequencies; and

[0049] A histogram of the filtered combined image is compiled, and the filtered combined image is renormalized pixel-by-pixel to a probability of not being noise by comparing the pixel signals of the filtered combined image at least in part to the histogram.

[0050] 15. The method of claim 14, wherein after filtering with the at least one defect pattern specific template, the method further comprises:

[0051] convolving the renormalized filtered combined image with a set of templates corresponding to the defined geometric shape;

[0052] The convolution result is weighted by the number of pixel signals that are actual data rather than the number of zero pixel signals corresponding to pixels located in an area where no image data is collected; and the weighted result is smoothed with the kernel to provide a filtered combined image, the smoothing including penalizing isolated signals and enforcing signals that match a predetermined characteristic length scale.

[0053] 16. The method of claim 14, wherein the kernel used to extract the relevant spatial frequencies comprises a simple Gaussian function having a width based on visual inspection of a predetermined characteristic length scale.

[0054] 17. The method of claim 14, further comprising:

[0055] The selected relevant spatial frequencies are downsampled to reduce processing time.

[0056] 18. The method of item 14, wherein the histogram comprises an empirical histogram fitted with a Gaussian distribution, and the difference between the empirical histogram and the Gaussian fit is larger for strong signals, and wherein the probability that the pixel is background is calculated by comparing the corresponding pixel signal with both the empirical histogram and the Gaussian fit.

[0057] 19. The method of claim 15, wherein creating the candidate pattern comprises:

[0058] Determine a reasonable set of two-dimensional coordinate systems of varying sizes; and

[0059] The two-dimensional coordinate system is scanned on the smoothed filtered combined image, local pixel signals at the intersection of each of the two-dimensional coordinate systems are removed, and combined signal strength in the removed local pixel signals within a tolerance region at each of the intersections of each of the candidate patterns is measured.

[0060] 20. The method of claim 19, wherein selecting the defect detection pattern from the created candidate patterns comprises:

[0061] summing the filter responses at regularly spaced grid points for each candidate pattern; and

[0062] The candidate pattern with the highest sum filter response is selected as the best candidate pattern,

[0063] The grid points of each candidate pattern mark the location of a strong filter signal, thereby capturing the location of the detected non-uniform defect. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The exemplary embodiments will be best understood from the detailed description that follows when read in conjunction with the accompanying drawings. It should be emphasized that the various features are not necessarily drawn to scale. In fact, for clarity of discussion, the dimensions may be increased or decreased arbitrarily. Where applicable and practical, the same reference numerals represent the same elements.

[0065] Figure 1 is a simplified block diagram of a defect detection system for detecting non-uniform defects in a master panel including a plurality of flat panel displays during manufacturing according to a representative embodiment.

[0066] Figure 2A is a flow chart illustrating an automated method for detecting non-uniform defects in a primary panel during manufacturing, according to a representative embodiment.

[0067] Figure 2B is a flow chart illustrating an automated method of filtering a primary panel image to detect non-uniform defects in the primary panel during manufacturing, according to a representative embodiment.

[0068] Figure 3 is a screenshot of an illustrative combined image obtained according to a representative embodiment.

[0069] Figure 4 Screen shots of illustrative combined images before and after downsampling, cleaning, and alignment of image data are shown according to a representative embodiment.

[0070] Figure 5 Screen shots of illustrative combined images before and after filtering with a kernel are shown according to a representative embodiment.

[0071] Figure 6 shows before and after kernel filtering according to a representative embodiment Figure 5 An enlarged portion of a screenshot of an illustrative combined image in .

[0072] Figure 7 Screen shots of an illustrative combined image before and after renormalizing the kernel filtered combined image are shown according to a representative embodiment.

[0073] Figure 8A screenshot of an illustrative renormalized combined image convolved with a defect pattern specific template is shown, according to a representative embodiment.

[0074] Fig. 9 A screenshot of an illustrative smoothed filtered combined image showing a non-uniform defect template is shown, according to a representative embodiment.

[0075] Fig.10 A screenshot showing an illustrative filtered combined image of a ring-shaped non-uniformity defect template and a plurality of different candidate patterns is shown, according to a representative embodiment.

[0076] Fig.11 A screenshot of an illustrative filtered combined image showing a ring-shaped non-uniformity defect template and candidate patterns with tolerance regions is shown, according to a representative embodiment.

[0077] Fig.12 A screenshot of an illustrative filtered combined image showing a selected defect detection pattern is shown, according to a representative embodiment. DETAILED DESCRIPTION

[0078] In the detailed description below, exemplary embodiments that disclose specific details are set forth for the purpose of explanation rather than limitation, so as to more fully understand the embodiments according to the present teachings. However, it is apparent to those of ordinary skill in the art who benefit from the present disclosure that other embodiments according to the present teachings that deviate from the specific details disclosed herein are still within the scope of the appended claims. In addition, known devices and methods are not described here in detail so as not to affect the description of the exemplary embodiments. Such methods and devices are obviously within the scope of the present teachings.

[0079] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. The defined terms are to be given the technical and scientific meanings of the defined terms as commonly understood and accepted in the technical field of the present teachings.

[0080] As used in this specification and the appended claims, the terms "a / an" and "the" include singular and plural referents unless the context clearly indicates otherwise. Thus, for example, "a device" includes both one device and plural devices. As used in the specification and the appended claims, and in addition to their ordinary meanings, the terms "substantially" or "substantially" mean within acceptable limits or degrees. As used in the specification and the appended claims, and in addition to their ordinary meanings, the term "approximately" means within acceptable limits or amounts for those skilled in the art. For example, "approximately the same" means that those skilled in the art would consider the items being compared to be the same.

[0081] Relational terms such as "above," "below," "top," "bottom" may be used to describe the relationship between different elements as shown in the accompanying drawings. These relational terms are intended to encompass different orientations of their elements in addition to the orientation depicted in the accompanying drawings. For example, if a device (e.g., a semiconductor package or a coaxial cable) depicted in a drawing is inverted relative to the view in the drawing, then, for example, an element described as being "above" another element would now be below the element. Similarly, if the device is rotated 90° relative to the view in the drawing, an element described as being "above" or "below" another element would now be "adjacent" to the other element; where "adjacent" means adjacent to the other element, or having one or more layers, one or more materials, structures, etc. between the elements.

[0082] In general, according to representative embodiments, a method for detecting non-uniform defects in a master panel during manufacturing is provided, wherein the master panel includes a plurality of flat screen displays. The method includes: preparing a combined image from image data of the master panel; enhancing the quality of the combined image, including removing artifacts from the combined image; filtering the quality-enhanced combined image to detect local non-uniform defects, including the local non-uniform defect having at least one structured pattern of a defined geometry; applying different candidate patterns to the filtered combined image; selecting one of the different candidate patterns as a defect detection pattern, the defect detection pattern being closest to the structured pattern of the defined geometry of the detected local non-uniform defect; and displaying at least a portion of the defect detection pattern and the quality-enhanced combined image on a display to show the location of the detected local non-uniform defect in the structured pattern of the defined geometry. Filtering the quality-enhanced combined image to detect local non-uniform defects may include filtering out relevant spatial frequencies corresponding to the length scale of the detected local non-uniform defect to provide a first filtered image; and convolving the first filtered image with a set of templates corresponding to the defined geometry, respectively.

[0083] Figure 1 is a simplified block diagram of a defect detection system for detecting non-uniform defects in a master panel including a plurality of flat panel displays during manufacturing according to a representative embodiment.

[0084] refer to Figure 1 , the defect detection system 100 includes a controller 110 that controls the operation of a scan head 120 and processes image data provided by the scan head 120. Image processing includes, for example, Figure 2A and Figure 2B 1. The controller 110 includes a processor 112, a memory 114, and an interface (I / F) 116. The defect detection system 100 also includes a display 130 to display images and data / image processing results provided by the processor 112 via the I / F 116 (which includes a display interface). The display 130 may include a graphical user interface (GUI) through which a user may interact with the controller 110 using information displayed on the display 130 and user input devices (not shown) such as a keyboard, mouse, touch pad, and / or touch-sensitive screen. Any other compatible device that provides input and receives output information may be included in the defect detection system 100 without departing from the scope of the present teachings.

[0085] Without departing from the scope of the present teachings, the scan head 120 may include different types of scanning devices to collect data from the main panel 105 being scanned in order to provide corresponding image data. For example, the scan head 120 may be an electrical contact scanner that collects data by making electrical contact with the main panel 105. Address and data signals indicating the top surface of the main panel 105 are transmitted by the metal probe of the scan head. More specifically, the electrical signals from the scan head 120 are connected to the pixels on the main panel 105 in a predetermined order to provide pixel signals (image data). The scan head 120 may also be an optical contact scanner that functions in a manner similar to an electro-optical scanner. In an alternative configuration, the scan head 120 may provide data by providing a non-contact scanning method that indicates that the main panel 105 is being scanned, including, for example, using one or more cameras, scanning sensors (such as a charge coupled device (CCD)) or laser excitation. Moreover, in an alternative configuration, the scanning head 120 may include a laser source, a lens, and electrical contacts through a probe to stimulate photoconductive currents in the TFT cells of the main panel 105 for analysis through electrical contact terminals, which may be part of the scanning head 120.

[0086] The main panel 105 includes pixels arranged in an array corresponding to a flat panel display. As mentioned above, the scanning head 120 can be electrically connected to the pixels on the main panel 105 in order to obtain image data. In one embodiment, the scanning head 120 is mobile and can be moved substantially parallel to the main panel 105, as indicated by arrow 122. This enables the scanning head 120 to obtain image data from different parts of the main panel 105, as long as the main panel 105 is larger than the scanning area of ​​the scanning head 120. The movement of the scanning head 120 can be performed manually, or can be automatically performed under the control of the controller 110. The scanning head 120 can be used to scan the entire main panel 105 or any one or more parts thereof. For example, the scanning of the scanning head 120 and the uneven detection of the controller 110 can be performed on a selected and / or randomly identified portion of the main panel 105 to sample the presence and nature of uneven defects without scanning the entire main panel 105.

[0087] According to various embodiments, the processor 112 of the controller 110 is programmed to perform a defect detection process, such as described below with reference to Figure 2A and Figure 2B The method steps described. Processor 112 can be implemented by one or more computer processors, application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or a combination thereof using software, firmware, hard-wired logic circuits, or a combination thereof. In one embodiment, processor 112 can be, for example, a digital signal processor (DSP). In particular, a computer processor can be constructed by a combination of hardware, firmware, or software architectures and can include its own memory (e.g., non-volatile memory) for storing executable software / firmware executable code that allows it to perform corresponding functions. In an embodiment, a computer processor can include, for example, a central processing unit (CPU) that executes an operating system.

[0088] The memory 114 stores at least a portion of the image data obtained using the scan head 120, as well as the processing results from the processor 112. The memory 114 may also be a database to store various types of templates, structured patterns of defined geometries, and candidate patterns of potential defect detection patterns applied by the processor 112 when processing image data and identifying non-uniform defects as discussed below. The memory 114 may be implemented by, for example, any number, type, and combination of random access memory (RAM) and read-only memory (ROM) and may, for example, store various types of information, such as computer programs and software algorithms executable by the processor 112 (and / or other components), as well as image data and / or test and measurement data storage and templates and patterns (discussed above). The various types of ROM and RAM may include any number, type, and combination of computer-readable storage media, such as disk drives, disk storage devices, electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), CDs, DVDs, universal serial bus (USB) drives, etc., which are tangible, non-transitory storage media (e.g., compared to transient propagating signals).

[0089] Figure 2A is a flow chart illustrating an automated method for detecting non-uniform defects in a primary panel during manufacturing, according to a representative embodiment.

[0090] refer to Figure 2A , in block S211, a combined image of a master panel (e.g., master panel 105) is obtained, wherein the master panel may be, for example, a glass substrate (e.g., a glass motherboard (MGS)) comprising a plurality of flat panel displays, but without departing from the scope of the present teachings, the master panel may also be a substrate comprising one or more other materials. Obtaining the combined image may include, for example, scanning the master panel using a scanning head (e.g., scanning head 120) to obtain a plurality of images of portions of the master panel, respectively. The scanning head includes one or more imaging devices that obtain image data of the master panel when the master panel passes through a defect detection system during a manufacturing process. In one embodiment, the image includes pixel signals received from a pixel matrix on the master panel in response to scanning. That is, the scanning signals from the scanning head may be connected to pixels on the master panel, respectively, in a predetermined order to provide pixel signals for obtaining the plurality of images. The plurality of images include corresponding image data, at least a portion of which may be stored in a memory (e.g., memory 114).

[0091] Figure 3 is a screenshot of an illustrative composite image obtained according to a representative embodiment. Figure 3, for example, the combined image 300 of the master panel can be larger than 10,000×10,000 pixels. As shown, the combined image 300 includes images of multiple flat panel displays that have not been separated from the master panel during the manufacturing process. As discussed below, the combined image of the master panel has had artifacts removed or suppressed and has been flattened. The images of the flat panel displays are indicated by an 8×6 rectangular array, but the number of flat panel display images in each combined image can vary without departing from the scope of the present teachings. For convenience, only the first row of the combined image and a portion of the second row are marked: the first row includes flat panel display images 311 to 318, and the second row includes flat panel display images 321 to 323. The combined image 300 also shows examples of non-uniform defects, such as annular non-uniform defects 331 and 332 of annular arrays partially shown in flat panel display images 312, 313 and 316, 317 (although the rings of the annular arrays also appear in other flat panel display images). The rings can be traces of suction cups used to fix the master panel during the manufacturing process. Another example of a non-uniformity defect is a spot array including representative spot non-uniformity defects 341, 342, and 343 in corresponding flat panel display images 321, 322, and 323, respectively (although not labeled, additional spot non-uniformity defects of the spot array also appear in other flat panel display images.) For example, the spot non-uniformity defect may result from handling equipment used to manipulate and / or move the primary panel.

[0092] Combining the image data of the multiple images (e.g., stored in memory 114) to provide a combined image can include: downsampling the multiple images to reduce storage requirements for storing the image data, removing measurement artifacts, arranging the downsampled images into a two-dimensional pattern, combining the two-dimensional pattern into a single larger image, suppressing processing artifacts, and correcting the contrast and background levels of the downsampled images of the single larger image to provide a combined image of the master panel. The combined image thus appears uniform. Areas of the combined image consisting of actual data (rather than unknown areas) can be marked in a separate binary matrix. The binary matrix areas are used to weight filter responses from filtering the combined image, for example, as discussed below with respect to block S213.

[0093] By cleaning the image data before and after combining the image data into a single larger image, removing and suppressing artifacts enhances the quality of the combined image. For example, assuming that the scan head 120 is an electrical contact scanner as discussed above, removing measurement artifacts from the image data in the individual frames (flat panel display images) before combining the two-dimensional patterns into a single larger image may include removing noise originating from electronic amplifier components, removing white noise using a low-pass filter, locally flattening the image, and suppressing pure vertical and pure horizontal spatial frequencies. Removing noise originating from electronic amplifier components may include removing drift by subtracting a sliding average, removing gain variations by normalizing the signal intensity, removing crosstalk between amplifier channels, and removing narrowband noise by applying a narrowband spatial frequency filter. Removing measurement artifacts may also include removing lines caused by, for example, laser annealing and other manufacturing processes. Suppressing artifacts after combining the two-dimensional patterns into a single larger image may include, for example, removing signal overshoots and undershoots in the corners of the stored image data of the individual frames by local flattening of the image signal, and performed before correcting contrast and background levels.

[0094] Figure 4 Screen shots of illustrative combined images before and after downsampling, cleaning, and aligning image data are shown according to a representative embodiment. Comparing the master image 401 and the combined image 402, it is apparent that the combined image 402 is flattened. Moreover, signal overshoots and undershoots that appear as bright spots (such as representative bright spots 411 and 421) in some corners of the individual flat panel display images in the master image 401 do not appear in the quality enhanced combined image 402.

[0095] In block S212 , local non-uniform defects are detected by filtering the quality enhanced combined image. Figure 2B is a flow chart illustrating an automated method of filtering a quality enhanced combined image to detect localized non-uniform defects (as indicated by block S212) according to a representative embodiment. Variations of the method of filtering a quality enhanced combined image may be included without departing from the scope of the present teachings.

[0096] The detected local non-uniform defects include at least one structured pattern of a corresponding defined geometry. For example, the structured pattern of a defined geometry may include a structured pattern of rings and / or spots on the quality enhanced combined image, such as Figure 3331 and 332 and representative spot non-uniform defects 341, 342 and 343 in the image. Thus, filtering the quality enhanced combined image to detect local non-uniform defects may include detecting structured patterns of defined geometric shapes using defect pattern specific templates. In general, filtering the quality enhanced combined image may include extracting relevant spatial frequencies corresponding to a length scale (e.g., a characteristic length) of the local non-uniform defects to provide a first filtered image, and then convolving the first filtered image with a set of defect pattern specific templates corresponding to defined geometric shapes (e.g., ring and spot non-uniform defects, respectively).

[0097] Figure 2B Depicted by Figure 2A The exemplary process for filtering the quality enhanced combined image is indicated by block S212 of FIG. Figure 2B According to one embodiment, the quality enhanced combined image from frame S211 is filtered using a kernel that selects (or extracts) relevant spatial frequencies, to compile a histogram of the combined image in frame S221 filtered by the kernel, and in frame S222, the filtered combined image is renormalized on the basis of each pixel signal by comparing the pixel signal of the filtered combined image at least partially with the histogram, as a probability of not being random noise. The histogram may include an empirical histogram fitted with a Gaussian distribution, wherein the difference between the empirical histogram and the Gaussian fit is larger for a strong signal. The probability that a pixel is background (random noise) is calculated by comparing the corresponding pixel signal with both the empirical histogram and the Gaussian fit. The kernel for extracting the relevant spatial frequency may include a simple Gaussian function, the width of which is based on a visual inspection of a predetermined characteristic length scale. For example, if time is a limiting factor, the selected relevant spatial frequency may be downsampled to reduce processing time. Irrelevant spatial frequencies (e.g., too high frequencies, too low frequencies) may be discarded.

[0098] Figure 5Screen shots of illustrative combined images before and after kernel filtering are shown according to a representative embodiment. Comparing the quality enhanced combined image 501 and the kernel filtered combined image 502, it is apparent that the frequency characteristics of the signal of interest have been emphasized. This is illustrated by the enhanced contrast of uneven defects and similar signals with reference to the background of the kernel filtered combined image 502. Moreover, high spatial frequencies are discarded by filtering with a kernel as mentioned above, as indicated by the scale of the vertical and horizontal axes of the combined images 501 and 502. That is, the vertical and horizontal axes are shorter because the image resolution has been reduced. High frequencies are rejected, and the remaining low frequencies can be equally represented by larger (and therefore fewer) pixels. For efficiency reasons, the image is therefore downscaled. The combined image 501 may correspond to Figure 4 The flattened combined image 402 in (discussed above).

[0099] Figure 6 shows before and after kernel filtering according to a representative embodiment Figure 5 . Each of the enlarged portions 501 'and 502' corresponds to the location of the annular non-uniformity defect 530 of the quality enhanced combined image 501 and the kernel filtered combined image 502, respectively. For the purpose of explanation, the annular non-uniformity defect 530 may correspond to Figure 3 One of the annular non-uniform defects 331 or 332 in the image 501 is also shown. A Gaussian filter 505 is shown, which indicates the kernel filtering performed on the combined image 501. The structure of interest in the depicted example has a "line width" of about 10 pixels to 20 pixels. Therefore, the filter width of the Gaussian filter 505 is selected to have a similar size for convolution (e.g., 20x 20 pixels), but other similar filter widths may also be included without departing from the scope of the present teachings. Comparing the enlarged portions 501' and 502' of the quality enhanced combined image 501 and the kernel filtered combined image 502, it can be seen that the signals from adjacent pixels are combined and averaged. This prevents individual pixels from having an undue impact on subsequent steps in the non-uniform defect detection process. It is assumed that there are a large number of random background responses from feature detection and small areas where actual signals exist in the feature detection. The absolute strength of the signal is not as important as whether the signal is considered to be real (rather than arising from particularly large background changes).

[0100] Figure 7Screen shots of an illustrative combined image before and after renormalizing the kernel filtered combined image in block S222 are shown according to a representative embodiment. The kernel filtered combined image 701 shows a significant amount of variation in the signal, indicated by spots throughout the combined image. As shown, the signal intensity of the background fluctuations is close to the actual signal intensity, making it more difficult to detect the actual signal. However, in contrast, the renormalized combined image 702 cuts out the background signal and sets the actual signal to a uniform signal intensity. This is shown by the sharp contrast of potential non-uniform defects (bright areas) with the substantially uniform and uniform background of the renormalized combined image 702. Notably, Figure 7 The kernel filtered combined image 701 in may correspond to the above Figure 5 Furthermore, the re-normalized combined image 702 may be the first filtered image mentioned above.

[0101] Next, in box S223, filtering the quality enhanced combined image to detect local non-uniform defects includes: convolving the renormalized combined image (first filtered image) with a set of defect pattern specific templates corresponding to defined geometric shapes (e.g., rings and spots) to identify defect patterns in the renormalized filtered combined image. The defect pattern does not have to match the shape and / or size of the template exactly, and multiple iterations can be performed if necessary. In box S224, the result of convolving the renormalized combined image with the set of defect pattern specific templates is weighted by the number of pixel signals that are actual data (rather than the number of zero pixel signals corresponding to pixels in areas where no image data is collected). The weighted result is smoothed with a kernel in box S225 to provide a filtered combined image. Smoothing includes penalizing isolated signals and enforcing signals that match a predetermined feature length scale, such as penalizing longer and shorter feature lengths.

[0102] Figure 8 A screenshot of an illustrative renormalized combined image convolved with a defect pattern specific template is shown according to a representative embodiment. Figure 8, the convolved image 801 shows a convolution using a template corresponding to a ring-shaped inhomogeneity defect, such as an illustrative ring template 811. The convolution image 802 shows a convolution using a template corresponding to a spot inhomogeneity defect, such as an illustrative spot template 812. For example, the ring template 811 and the spot template 812 can be implemented as a software algorithm or a small application. In the depicted embodiment, the ring-shaped inhomogeneity defect is significantly larger than the spot inhomogeneity defect, for example, the ring template 811 can be 200×200 pixels, and the spot template can be only 20×20 pixels, for example, but other sizes and relative sizes can also be included without departing from the scope of the present teachings. Of course, if there is uncertainty about the size and aspect ratio of the ring-shaped inhomogeneity defect and / or the spot inhomogeneity defect, multiple ring and / or spot templates can be used. Moreover, the various embodiments are not limited to ring-shaped inhomogeneity defects and spot inhomogeneity defects. That is, templates can be created as needed for any kind of shape that is expected to appear as an inhomogeneity defect in the renormalized combined image.

[0103] The convolution performed in block S223 is Figure 7 The cross-correlation between the ring template 811 and each of the potential non-uniform defects, and the cross-correlation between the spot template 812 and each of the potential non-uniform defects are detected in the re-normalized combined image 702 of FIG. 8 . In the case where the cross-correlation indicates that the shapes are sufficiently similar, the initial pattern of the non-uniform defects is detected. As mentioned above, the convolution result is weighted by the number of pixel signals that are actual data, and the weighted result is smoothed with a kernel to provide a filtered combined image of the detected pattern including ring and spot non-uniform defects. Smoothing effectively penalizes isolated signals (e.g., false responses) and enforces signals that match the correlation length scale.

[0104] Fig. 9 Screen shots of illustrative smoothed filtered combined images showing templates for non-uniform defects according to representative embodiments are shown. Specifically, combined image 901 shows a template corresponding to a ring-shaped non-uniform defect, and combined image 902 shows a template corresponding to a spot non-uniform defect. As shown above with reference to Figure 2B As discussed in block S225 of FIG. 1 , smoothing includes penalizing isolated signals and enforcing (or enhancing) a predetermined characteristic length scale (such as Fig. 9 For example, characteristic length scale 915 is slightly smaller than spot template 812, thereby enhancing spot non-uniformity defects and ring non-uniformity defects while not enhancing signals having different characteristic lengths (at least not to the same extent).

[0105] Reference again Figure 2A, in box S213, different candidate patterns are created and applied to the filtered combined image. The candidate patterns have two-dimensional coordinate systems of different sizes, respectively, which attempt to show the positioning of the detected local non-uniform defects. Each of the candidate patterns may include horizontal and vertical periodicity of the detected local non-uniform defects, as well as horizontal and vertical positioning, thereby creating a rectangular pattern. Of course, without departing from the scope of the present teachings, patterns with shapes other than rectangles may be included, for example, depending on the shape of the substrate and / or the shape of the scanning head. For example, the candidate pattern may be circular or hexagonal. The horizontal and vertical positioning of each of the candidate patterns varies, for example, according to a search algorithm that may be referred to as a scan offset. That is, the scan offset corresponds to the positioning of the candidate pattern relative to the position of a single display on the glass panel. Moreover, at least one of the candidate patterns may be rotated by introducing a rotation angle parameter to better fit the orientation of one or more of the detected local non-uniform defects.

[0106] To create candidate patterns, the user can estimate the range of plausible horizontal and vertical periodicities. Alternatively, the maximum value of the cross-correlation of the template and potential non-uniform defects discussed above can be analyzed and applied to determine plausible horizontal and vertical periodicities. Furthermore, an exhaustive scan of all possible horizontal and vertical periodicity values ​​can be performed and the results used to constrain the range of plausible horizontal and vertical periodicities in subsequent detection attempts repeated multiple times.

[0107] The candidate pattern is scanned on a data matrix (array) corresponding to the filtered combined image showing the non-uniform defect template. The local data at the intersection of the candidate pattern is cut off, and the signal intensity in the pixel corresponding to the cut-off local data is measured. In order to allow a certain error range, a region of a finite size corresponding to the allowable error range is cut off around the intersection of the candidate pattern, and the maximum value of the combined signal therein is measured, rather than the average value. For data matrices with many rows and / or columns, points in the data matrix that have no chance of contributing to the signal because they happen to be located at a position where no data is recorded should not be penalized. However, if there are only 2x 2 positions in the data matrix, all positions must be at least partially visible, otherwise there is no defined solution to the positioning problem.

[0108] Fig.10 A screenshot of an illustrative filtered combined image showing a circular non-uniform defect template and a plurality of different candidate patterns according to a representative embodiment is shown. Specifically, a first candidate pattern 1030, a second candidate pattern 1040, and a third candidate pattern 1050 of different sizes and dimensions are shown on the combined image 901, where the template corresponds to a circular non-uniform defect. (It should be understood that the candidate patterns of different sizes and dimensions will be similarly applied to Fig. 9, where the template corresponds to a spot non-uniformity defect, and thus the discussion also applies to the combined image 902). Moreover, double arrows 1041 and 1042 indicate the horizontal and vertical scan offsets of the second candidate pattern 1040, and double arrows 1051 and 1052 indicate the horizontal and vertical scan offsets of the third candidate pattern 1050. The first candidate pattern 1030 also has horizontal and vertical scan offsets, but is not in Fig.10 . The first candidate pattern 1030, the second candidate pattern 1040, and the third candidate pattern 1050 are arranged in the horizontal and vertical directions (i.e., not rotated), but one or more of the first candidate pattern 1030, the second candidate pattern 1040, and the third candidate pattern 1050 may be rotated to better fit the template pattern corresponding to the annular non-uniform defect. A signal is detected at the corners of the first candidate pattern 1030, the second candidate pattern 1040, and the third candidate pattern 1050, and the signal is part of the total filter response from the combined image 901. As discussed above, a certain limited size area corresponding to the allowable error range is cut off around the intersection of the first candidate pattern 1030, the second candidate pattern 1040, and the third candidate pattern 1050 (e.g., the corner), and the maximum combined signal is measured within the rest of the first candidate pattern 1030, the second candidate pattern 1040, and the third candidate pattern 1050, rather than the average combined signal. The signal is encoded in pixel values.

[0109] In one embodiment, the signal detected at the corners of the first candidate pattern 1030, the second candidate pattern 1040, and the third candidate pattern 1050 is the maximum signal obtained within the tolerance region. Fig.11 A screenshot of an illustrative filtered combined image showing a ring-shaped non-uniformity defect template and candidate patterns with tolerance regions is shown according to a representative embodiment. Fig.11, for convenience, tolerance regions are shown on the corners of the first candidate pattern 1030, but it should be understood that other candidate patterns (e.g., the second candidate pattern 1040 and the third candidate pattern 1050) may also include tolerance regions. Specifically, at each of the four corners of the first candidate pattern 1030 is a tolerance region 1031, 1032, 1033, and 1034 from which a signal is detected. Tolerance regions 1031, 1032, 1033, and 1034 are not drawn to scale because the actual tolerance reflects the uncertainty of horizontal and / or vertical periodicity and may therefore be smaller in a scaled drawing. In one embodiment, each of tolerance regions 1031, 1032, 1033, and 1034 is centered on the corresponding corner (intersection) of the first candidate pattern 1030. Tolerance regions 1031, 1032, 1033, and 1034 allow fitting a slightly rotated data matrix without having to rotate the candidate pattern 1030. This helps speed up the process of creating and selecting candidate patterns and ultimately detecting non-uniform defects.

[0110] In block S214, one of the candidate patterns is selected as a defect detection pattern, wherein the defect detection pattern is a candidate pattern of a structured pattern of a defined geometry that most closely resembles the detected local non-uniform defect. The defect detection pattern may be automatically selected by determining which candidate pattern provides the strongest signal of the total filter response from the filtered combined image. For example, selecting the defect detection pattern may include summing the filter responses at regularly spaced grid points of each of the candidate patterns. The candidate pattern with the highest summed filter response is then selected as the best candidate pattern. The grid points of each candidate pattern mark the location of a strong filter signal, thus capturing the location of the detected non-uniform defect.

[0111] The selected defect detection pattern can be fine-tuned by measuring the median signal at the horizontal and vertical (X, Y) coordinates of each row and column of pixels in the filtered combined image, for example, to provide robustness to outliers. Optionally, a rotation can be determined to provide a location of the selected defect detection pattern. For example, the rotation can be inferred by a linear fit in polar coordinates. And optionally, the lattice periodicity can be inferred by a linear fit so that rows and columns that are not visible in the filter combined image can be inferred. In addition, a scan offset can be selected in block S214, where the selected scan offset most closely corresponds to the location of the detected localized inhomogeneous defect.

[0112] Fig.12Screen shots of illustrative filtered combined images showing selected defect detection patterns according to representative embodiments are shown. Specifically, combined image 1201 shows a defect detection pattern selected for a spot non-uniformity defect, and combined image 1202 shows a defect detection pattern selected for a ring non-uniformity defect. In combined image 1201, hash marks indicate defect detection patterns selected for non-uniform spots, and circles indicate detections of spot non-uniformity defects. Of course, these detections contain a certain error range, so it is assumed that they are arranged in a regular pattern, i.e., the hash marks are actual locations. As shown in FIG. Fig.12 As shown in , the hash marks are substantially aligned with the actual spot unevenness defects. In the combined image 1202, there are four groups of hash marks, each group of hash marks indicating a defect detection pattern selected for a corresponding annular unevenness defect. Portions of each annular unevenness defect are visible in the combined image 1202 surrounded by the selected defect detection pattern.

[0113] In block S215, the detected local non-uniform defects are quantified using the (fine-tuned) defect detection pattern. Parameters such as diameter and eccentricity, as well as statistical parameters such as mean, median, maximum and minimum signal strength, as well as noise level and other parameters can be extracted from the areas marked as non-uniform defects.

[0114] In block S216, at least a portion of the defect detection pattern is displayed on a display (e.g., display 130) along with the filtered combined image. The displayed portion of the defect detection pattern shows the location of the detected local non-uniform defects in the structured pattern of the defined geometry. The detected local non-uniform defects can be used for many purposes. For example, the master panel can be accepted or rejected based on the extent to which the local non-uniform defects have been detected. Moreover, the detected local non-uniform defects can be used to identify the manufacturing equipment that caused the non-uniform defects so that the manufacturing equipment can be fine-tuned or adjusted or replaced.

[0115] According to various embodiments, non-uniform defects on the main panel and / or the flat panel display included therein can be automatically detected, classified and / or quantified. Compared with manual inspection, the automated non-uniform defect detection process according to various embodiments results in increased production volume, reduced labor costs and standardized results. The process also provides flexibility, enabling users to add target defect patterns for detection as needed. In addition, the positioning and quantification of defect patterns allow for graded scoring of panel quality, which is superior to a simple pass / fail assessment. For example, based on which panels can be accepted or rejected based in part on the presence and extent of non-uniform defects, the graded scoring of panel quality can be included in the quality assurance determination. Moreover, the detected pattern type and positioning make it possible to determine which processing equipment in the production line is responsible for causing a specific non-uniform defect. Therefore, as mentioned above, the production line can be improved by fine-tuning, adjusting or replacing the equipment that causes the non-uniform defect, thereby increasing the yield.

[0116] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed embodiments.

[0117] Other variations to the disclosed embodiments may be understood and effected by those skilled in the art in practicing the claimed invention, by studying the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude the inclusion of other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0118] Although representative embodiments are disclosed herein, those skilled in the art will appreciate that many variations are possible in light of the present teachings and still fall within the scope of the appended claims. Accordingly, the present invention is limited only by the scope of the appended claims.

Claims

1. A method of detecting non-uniform defects in a main panel (105) during manufacturing, the main panel (105) comprising a plurality of flat screen displays, the method include: preparing a combined image (1201) from the image data of the master panel (105); enhancing the quality of the combined image (1201), wherein the enhancing the quality comprises removing artifacts from the combined image (1201); filtering the quality-enhanced combined image (402) to detect localized inhomogeneity defects, the localized inhomogeneity defects comprising at least one structured pattern of a defined geometry; applying a plurality of different candidate patterns to the filtered combined image (1201); selecting one of the plurality of candidate patterns as a defect detection pattern, and selecting a defect detection pattern from the plurality of candidate patterns that is closest to the structured pattern of the defined geometric shape of the detected local non-uniform defect; as well as At least a portion of the defect detection pattern and the quality-enhanced combined image (1201) are displayed on a display (130) to show the location of the detected local non-uniform defect in the structured pattern of the defined geometry.

2. The method of claim 1, wherein the quality-enhanced combined image (402) is filtered to detect localized inhomogeneous defects. include: filtering out relevant spatial frequencies corresponding to a length scale of the detected localized inhomogeneity defect to provide a first filtered image; as well as The first filtered image is convolved using a set of templates corresponding to the defined geometric shapes respectively. The method of claim 1 , wherein the defined geometric shape comprises a spot and / or a plurality of rings.

4. A method for detecting non-uniform defects in a flat panel display before separation from a main panel (105) during a manufacturing process, the method include: Acquiring a plurality of images of the master panel (105) and storing image data corresponding to some of the plurality of images, wherein the image data comprises pixel signals received from a pixel matrix on the master panel (105); combining the stored image data of the plurality of images to provide a combined image (1201); filtering the combined image (1201) to detect inhomogeneous defects using at least one defect pattern specific template to detect a structured pattern of a defined geometry corresponding to the inhomogeneous defect; Creating candidate patterns, each of which has a two-dimensional coordinate system of different sizes, the two-dimensional coordinate system attempting to demonstrate the location of the detected non-uniform defect; selecting a defect detection pattern and a scanning offset that most closely corresponds to the location of the detected non-uniform defect from the created candidate patterns, the selected defect detection pattern providing the strongest signal from the total filter response of the filtered combined image (1201); Fine-tune the selected defect detection pattern; as well as The detected non-uniform defects are quantified using the fine-tuned defect detection pattern.

5. The method of claim 4, wherein each of the candidate patterns comprises horizontal and vertical periodicity of the detected non-uniformity defect and horizontal and vertical scanning offsets, thereby creating a rectangular pattern. The method of claim 5 , wherein at least one of the candidate patterns is rotated by introducing a rotation angle parameter.

7. The method of claim 4, wherein obtaining the multiple images of the main panel (105) includes scanning the main panel (105) using a scanning head (120), wherein electrical signals or optical signals from the scanning head (120) are respectively connected to pixels on the main panel (105) in a predetermined order to provide the pixel signals.

8. The method of claim 4, wherein the primary panel (105) comprises a glass substrate.

9. The method of claim 4, wherein the stored image data of the plurality of images is combined to provide the combined image (1201) include: downsampling the plurality of images to reduce storage requirements for storing the image data; Arranging the downsampled multiple images into a two-dimensional pattern, and combining the two-dimensional pattern into a single larger image; Correcting the contrast and background levels of the downsampled multiple images of the single larger image to provide the combined image to appear uniform (1201); as well as Marking is performed in a separate binary matrix region of the combined image (1201) consisting of actual data, rather than in an unknown region, wherein the binary matrix region is used to weight a filter response from filtering the combined image (1201).

10. The method of claim 9, further comprising cleaning the combined image (1201) before filtering the combined image (1201) to detect the non-uniform defect, cleaning the combined image (1201) include: removing measurement artifacts from the stored image data before combining the two-dimensional patterns into the single larger image; as well as Artifacts are suppressed in individual frames of this single larger image before correcting contrast and background levels.

Images