Semiconductor manufacturing equipment, inspection equipment, inspection method, and method for manufacturing semiconductor equipment

Abstract

The objective is to provide technology that can improve the sensitivity of visual inspection. [Solution] The semiconductor manufacturing apparatus comprises an imaging device, an illumination device, and a control device. The control device is configured to (a) irradiate the die with illumination light using the illumination device and capture image data using the imaging device, (b) process the acquired image data to extract candidate pixels that meet a first predetermined condition, and (c) search for candidate lines based on the extracted candidate pixels and combine the candidate pixels to detect abnormal candidates.

Description

Technical Field

[0001] The present disclosure relates to a semiconductor manufacturing apparatus, and is applicable to, for example, a die bonder that inspects for defects such as scratches on a die.

Background Art

Prior Art Documents

Patent Documents

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

Means for Solving the Problems

[0007] According to this disclosure, it is possible to improve the sensitivity of visual inspection. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 is a schematic top view showing an example of the configuration of a die bonder in an embodiment. [Figure 2] Figure 2 is a diagram illustrating the schematic configuration as seen from the direction of arrow A in Figure 1. [Figure 3] Figure 3 is a block diagram showing the schematic configuration of the die bonder control system shown in Figure 1. [Figure 4] Figure 4 is a flowchart showing a method for manufacturing a semiconductor device using the die bonder shown in Figure 1. [Figure 5] Figure 5 shows the optical system of the pickup unit in the embodiment. [Figure 6] Figure 6 is a flowchart showing a crack detection method in an embodiment. [Figure 7] Figure 7 is a simulated image of crack detection formed on the die. [Figure 8] Figure 8 is a graph showing the brightness values ​​of the crack region and the surrounding region. [Figure 9] Figure 9 illustrates the cases in which striped patterns appear on the die surface and in which they do not. [Figure 10] Figure 10 shows the flow chart for crack inspection using the Hough transform. [Figure 11] Figure 11 shows the target pixel in a two-dimensional coordinate system. [Figure 12] FIG. 12 is an image diagram showing a histogram of two-dimensional parameters by θ and R. [Figure 13] FIG. 13 is a diagram showing a plurality of strip-shaped candidate intervals. [Figure 14] FIG. 14(a) is a diagram showing a provisional straight line obtained when the search width is narrow. FIG. 14(b) is a diagram for explaining the recounting of neighboring pixels with the search width widened with respect to the provisional straight line shown in FIG. 14(a). FIG. 14(c) is a diagram showing an approximate straight line of the pixels recounted in FIG. 14(b). [Figure 15] FIG. 15 is a diagram for explaining the problems of the least squares method. [Figure 16] FIG. 16 is a diagram for explaining the case where straight line extraction is performed based on detection points. [Embodiment for Implementing the Invention]

[0009] Hereinafter, embodiments will be described with reference to the drawings, provided that in the following descriptions, the same reference numerals may be assigned to the same components and repeated descriptions may be omitted. The drawings may be schematically represented in terms of the width, thickness, shape, etc. of each part compared with the actual aspect in order to make the description clearer. Also, the dimensional relationships and ratios of each element among a plurality of drawings do not necessarily match.

[0010] [Configuration of Semiconductor Manufacturing Apparatus] As shown in FIG. 1, the die bonder 1 generally includes a wafer supply unit 10, a pickup unit 20, an intermediate stage unit 30, a bonding unit 40, a transfer unit 50, a substrate supply unit 60, a substrate carry-out unit 70, and a control unit 80. The Y2 - Y1 direction is the front - rear direction of the die bonder Ⅰ, and the X^2 - X1 direction is the left - right direction. The wafer supply unit 10 is arranged on the front side of the die bonder 1, and the bonding unit 40 is arranged on the rear side.

[0011] As shown in FIG. 1, the wafer supply unit 10 includes a wafer cassette lifter 11, a wafer holding stage 12, and a peeling unit 13.

[0012] The wafer cassette lifter 11 moves up and down a wafer cassette (not shown) in which a plurality of wafer rings WR are stored to the wafer transfer height. A wafer alignment chute (not shown) aligns the wafer ring WR supplied from the wafer cassette lifter 11. A wafer extractor (not shown) takes out the wafer ring WR from the wafer cassette and supplies it to the wafer holding stage 12, or takes it out from the wafer holding stage 12 and stores it in the wafer cassette.

[0013] A wafer W is adhered (stuck) onto a dicing tape DT, and the wafer W is divided into a plurality of dice D. The dicing tape DT is held by the wafer ring WR. The wafer W is, for example, a semiconductor wafer or a glass wafer, and the dice D are semiconductor chips, glass chips, or MEMS (Micro Electro Mechanical Systems). A film-like adhesive material DF called a die attach film (DAF) may be pasted between the wafer W and the dicing tape DT. The adhesive material DF cures by heating.

[0014] The wafer holding stage 12 moves in the X1 - X2 direction and the Y1 - Y2 direction. Thereby, the picked-up die D moves to the position of the peeling unit 13. Further, the wafer holding stage 12 rotates the wafer ring WR within the XY plane. The peeling unit 13 moves in the vertical direction. The peeling unit 13 peels the die D from the dicing tape DT.

[0015] As shown in Figures 1 and 2, the pickup unit 20 includes a pickup head 21, a pickup head table 23, and a wafer recognition camera 24. The pickup head 21 is provided with a collet 22 that adsorbs and holds the peeled die D at its tip. The Z1-Z2 direction is the vertical direction. The pickup head 21 picks up the die D from the wafer supply unit 10 and places it on the intermediate stage 31. The pickup head table 23 moves the pickup head 21 in the Z1-Z2, Y1-Y2, and X1-X2 directions. The pickup head table 23 may also rotate the pickup head 21. The wafer recognition camera 24 recognizes the pickup position of the die D to be picked up from the wafer W and performs visual inspection of the die D.

[0016] As shown in Figures 1 and 2, the intermediate stage section 30 includes an intermediate stage 31 on which the die D is placed, and a stage recognition camera 34 for recognizing the die D on the intermediate stage 31. The intermediate stage 31 is equipped with suction holes for attracting the placed die D. The placed die D is temporarily held on the intermediate stage 31. The intermediate stage 31 is both a placement stage on which the die D is placed and a pickup stage on which the die D is picked up.

[0017] As shown in Figures 1 and 2, the bonding unit 40 includes a bond head 41, a bond head table 43, a substrate recognition camera 44, and a bond stage 46. The bond head 41 is provided with a collet 42 that adsorbs and holds the die D at its tip. The bond head table 43 moves the bond head 41 in the Z1-Z2, Y1-Y2, and X1-X2 directions. The bond head table 43 may also rotate the bond head 41. The substrate recognition camera 44 images the substrate S and recognizes the bond position. Here, the substrate S is, for example, a wiring board, a lead frame, a glass substrate, etc. Multiple product areas are formed on the substrate S. The product areas ultimately become one package. The product areas are hereinafter referred to as package areas P. Also, position recognition marks (not shown) for the package areas P are formed on the substrate S. When the die D is placed on the substrate S, the bond stage 46 is raised to support the substrate S from below. The bond stage 46 has a suction port (not shown) for vacuum adsorption of the substrate S, and can fix the substrate S in place. The bond stage 46 also has a heating section (not shown) for heating the substrate S.

[0018] With this configuration, the bond head 41 corrects its pickup position and orientation based on the image data from the stage recognition camera 34 and picks up the die D from the intermediate stage 31. Then, the bond head 41 bonds the picked-up die D onto the package area P of the substrate S based on the image data from the substrate recognition camera 44, or bonds it by stacking it on top of a die D that has already been bonded to the package area P of the substrate S.

[0019] As shown in Figure 1, the transport unit 50 has transport claws 51 that grasp and transport the substrate S, and a transport lane 52 on which the substrate S moves. The substrate S moves in the X1 direction by driving, for example, a nut (not shown) of the transport claws 51 provided on the transport lane 52 with a ball screw (not shown) provided along the transport lane 52. With this configuration, the substrate S moves from the substrate supply unit 60 along the transport lane 52 to the bonding position, and after bonding, moves to the substrate discharge unit 70 and is handed over to the substrate discharge unit 70.

[0020] The substrate supply unit 60 takes out the substrates S that have been transported in a transport jig (not shown) from the transport jig and supplies them to the transport unit 50. The substrate discharge unit 70 stores the substrates S transported by the transport unit 50 into the transport jig.

[0021] As shown in Figure 3, the control system 8 comprises a control unit 80, a drive unit 86, a signal unit 87, and an optical system 88. The control unit 80 is also called a control device or controller. The control unit 80 is broadly configured as a computer comprising a control / arithmetic unit 81 mainly composed of a CPU (Central Processing Unit), a storage device 82, an input / output device 83, a bus line 84, and a power supply unit 85. The storage device 82 has a main memory 82a and an auxiliary storage device 82b. The main memory 82a is composed of RAM (Random Access Memory) that stores processing programs, etc. The auxiliary storage device 82b is composed of an HDD (Hard Disk Drive) or SSD (Solid State Drive) etc. that stores control data and image data necessary for control. In addition, it is possible to connect an external storage device to the control unit 80.

[0022] The input / output device 83 includes a monitor 83a that displays the device status and information, a pointing device such as a touch panel 83b for inputting operator instructions and a mouse 83c for operating the monitor 83a, and an image acquisition device 83d for acquiring image data from the optical system 88. The input / output device 83 further includes a motor control device 83e and an I / O signal control device 83f.

[0023] The motor control device 83e controls the drive unit 86. The drive unit 86 includes the XY table (not shown), pickup head table 23, and bond head table 43 of the wafer supply unit 10.

[0024] The I / O signal control device 83f acquires signals from various sensors in the signal unit 87 and controls switches and potentiometers that control the brightness of lighting devices and other equipment in the signal unit 87, as well as valves that control vacuum suction.

[0025] The optical system 88 includes a wafer recognition camera 24, a stage recognition camera 34, and a substrate recognition camera 44, which serve as imaging devices. The wafer recognition camera 24, stage recognition camera 34, and substrate recognition camera 44 quantify light intensity and color.

[0026] The control and calculation unit 81 receives necessary data via the bus line 84, performs calculations, and sends information to control the pickup head 21 and other components, as well as to the monitor 83a and other components.

[0027] The control unit 80 stores image data captured by the wafer recognition camera 24, stage recognition camera 34, and substrate recognition camera 44 via the image acquisition device 83d in the storage device 82. Based on the stored image data, the control / calculation device 81 uses programmed software to recognize the positions of the die D and the package area P of the substrate S, and to perform visual inspection of the die D and the substrate S. Based on the positions of the die D and the package area P of the substrate S calculated by the control / calculation device 81, the motor control device 83e moves the drive unit 86. Through this process, the position of the die D on the wafer W is recognized, the pickup head table 23 and the bond head table 43 are operated, and the die D is bonded onto the package area P of the substrate S.

[0028] The control unit 80 can be configured by installing the above-mentioned program stored in an external storage device onto a computer. The external storage device includes, for example, an HDD, USB memory, or SSD. The auxiliary storage device 82b and the external storage device are configured as computer-readable recording media. Hereinafter, these will be collectively referred to simply as recording media. In this specification, when the term recording media is used, it may include only the auxiliary storage device 82b, only the external storage device, or both. Note that the provision of programs and data to the computer and the provision of programs and data from the computer to the external device may be performed using communication means such as the internet or a dedicated line, without using external storage devices.

[0029] As shown in Figure 4, a part of the semiconductor device manufacturing process (method of manufacturing a semiconductor device) is carried out using the die bonder 1. In the following description, the operation of each part constituting the die bonder 1 is controlled by the control unit 80.

[0030] (Wafer loading process: Process S1) A wafer cassette containing wafer rings WR is loaded into the wafer cassette lifter 11. The loaded wafer rings WR are supplied (transported) to the wafer holder 12. The wafers W have been inspected beforehand by an inspection device such as a prober, and wafer map data indicating defects in electrical characteristics has been generated. This wafer map data is stored in the storage device 82 of the control unit 80. Defect marks may be added to the wafers W instead of wafer map data. The following manufacturing process is performed on good dies D or dies D without defect marks in the wafer map data.

[0031] (Substrate loading process: Process S2) A transport jig (not shown) containing the substrate S is fed into the substrate supply unit 60. In the substrate supply unit 60, the substrate S stored in the transport jig is removed from the transport jig. Then, the substrate S is supplied (transported) to the bonding unit 40 via the transport unit 50.

[0032] (Pickup process: Process S3) After process S1, the wafer holder 12 is moved so that the desired die D can be picked up from the dicing tape DT. The die D is photographed by the wafer recognition camera 24, and image data is acquired through the photography. By processing the image data, the amount of displacement (in the X, Y, and θ directions) of the die D on the wafer holder 12 from the die position reference point of the die bonder 1 is calculated. The die position reference point is a predetermined position of the wafer holder 12 that is held as the initial setting of the device. By processing the image data, a visual inspection of the die D is performed.

[0033] The die D is peeled off from the dicing tape DT by the peeling unit 13 and the pickup head 21. The die D, peeled off from the dicing tape DT, is attracted and held by a collet 22 provided on the pickup head 21, and is transported to and placed on the intermediate stage 31.

[0034] The die D on the intermediate stage 31 is photographed by the stage recognition camera 34, and image data is acquired through the photography. By processing the image data, the amount of displacement (in the X, Y, and θ directions) of the die D on the intermediate stage 31 from the die position reference point of the die bonder 1 is calculated. The die position reference point is a predetermined position on the intermediate stage 31 that is held as the initial setting of the device. Visual inspection of the die D is performed by processing the image data.

[0035] The pickup head 21, which has transported die D to the intermediate stage 31, is returned to the wafer supply unit 10. Following the procedure described above, the next die D is peeled off from the dicing tape DT, and thereafter, die D is peeled off one by one from the dicing tape DT following the same procedure.

[0036] (Bond process: Process S4) The transport unit 50 transports the substrate S to the bond stage 46. The substrate S placed on the bond stage 46 is imaged by the substrate recognition camera 44, and image data is acquired by the image capture. By processing the image data, the amount of displacement of the substrate S from the substrate position reference point of the die bonder 1 (in the X, Y, and θ directions) is calculated. The substrate position reference point is a predetermined position of the bonding unit 40, which is held as the initial setting of the device.

[0037] In step S3, the suction position of the bond head 41 is corrected based on the amount of displacement of the die D on the intermediate stage 31 calculated, and the die D is picked up by the collet 42. The die D is bonded to a predetermined location on the substrate S supported by the bond stage 46 by the bond head 41, which has picked up the die D from the intermediate stage 31. The die D bonded to the substrate S is photographed by the substrate recognition camera 44, and an inspection is performed based on the image data acquired by the photograph to determine whether the die D has been bonded to the desired position (relative position inspection of die D and substrate S), etc.

[0038] The bond head 41, having bonded die D to the substrate S, is returned to the intermediate stage 31. Following the procedure described above, the next die D is picked up from the intermediate stage 31 and bonded to the substrate S. This process is repeated until die D is bonded to all package areas P on the substrate S.

[0039] (Substrate unloading process: Process S5) The transport unit 50 transports the substrate S to which the die D has been bonded from the bonding unit 40 to the substrate unloading unit 70. In the substrate unloading unit 70, the substrate S is removed, stored in a transport jig, and unloaded. The transport jig containing the substrate S is unloaded from the die bonder 1.

[0040] As described above, die D is mounted on substrate S and discharged from die bonder 1. Subsequently, for example, the transport jig containing substrate S with die D mounted on it is transported to the wire bonding process, where the electrodes of die D are electrically connected to the electrodes of substrate S via Au wire or the like. Then, substrate S is transported to the molding process, where die D and Au wire are sealed with molding resin (not shown), thereby completing the semiconductor package.

[0041] [Inspection Department] Scratches formed on die D are inspected using a wafer recognition camera 24 provided on the pickup unit 20, for example, as shown in Figure 5.

[0042] As shown in Figure 5, a wafer recognition camera 24 with a lens 25 attached is positioned above the surface of the die D to be inspected. The field of view CV of the wafer recognition camera 24 includes the die D to be inspected and some or all of the adjacent surrounding dies Da, Db, etc. The oblique light illumination devices 26a and 26b are illumination devices such as oblique light bars, and irradiate illumination light IL near the outside of the die D to be inspected at a predetermined angle with respect to the optical axis OA. Here, the illumination light IL is irradiated toward the dies Da, Db adjacent to die D. The irradiation surfaces of the oblique light illumination devices 26a and 26b extend in the Y direction. The irradiation direction of the illumination light IL in the horizontal direction of the oblique light illumination device 26a is in the X1 direction. The irradiation direction of the illumination light IL in the horizontal direction of the oblique light illumination device 26b is in the X2 direction. The wafer recognition camera 24 and the oblique light illumination devices 26a and 26b constitute the inspection apparatus.

[0043] The oblique light illumination devices 26a and 26b realize a dark-field method that darkens the background and brightly illuminates what is to be viewed. Visual inspection (dark-field inspection) in a dark-field inspection system is performed in areas other than the specular reflection region, which is derived from the placement of the oblique light illumination devices 26a and 26b. Here, the specular reflection region is the region in which specular reflection images of illumination projected onto the surface of dies Da, Db, etc., which mainly exhibit specular reflection characteristics, can be captured. The specular reflection region is mainly formed on dies Da, Db adjacent to the die D to be inspected. However, the specular reflection region may become part of the area of ​​the die to be inspected, or it may be on the inside of the die to be inspected depending on the warping or deformation of the die. In dark-field inspection, the visualization of defects is performed by the reflection of light from the sides inside the minute defects. When defects such as cracks occur continuously in a linear manner, the sides are also continuous, and the defects are visualized depending on the direction of irradiation of the illumination light IL. For this reason, in the horizontal direction, the illumination light IL is irradiated from a direction different from the direction in which the defect extends so that light hits the sides.

[0044] Furthermore, a coaxial illumination device may be provided below the wafer recognition camera 24. The coaxial illumination device is used, for example, as illumination (bright-field illumination) when recognizing the position of die D.

[0045] Below the stage recognition camera 34 located in the intermediate stage section 30, lighting devices similar to the oblique light lighting devices 26a, 26b and coaxial lighting devices of the pickup section 20 are provided. Below the substrate recognition camera 44 in the bonding section 40, lighting devices similar to the oblique light lighting devices 26a, 26b and coaxial lighting devices of the pickup section 20 are provided.

[0046] [Method for detecting damage (inspection method)] As shown in Figure 6, the control unit 80 performs an inspection step to detect defects (e.g., cracks K) formed on the die D.

[0047] (Diatomic imaging: Step S11) Die D is imaged using dark-field illumination as shown in Figure 5. Here, die D has been deemed a good product by probe inspection, and its position has been identified. Under dark-field illumination, crack K appears brighter than the background, as shown in image 1-A in Figure 7. Image 1-A is a clear image of crack K. In image 1-A, crack K extends from the upper right (X1 direction and Y1 direction) to the lower left (X2 direction and Y2 direction) of die D on the drawing.

[0048] The image 2-A shown in Figure 7 is an image of a faint crack K. In image 2-A, crack K extends from the upper right to the lower left of the die D drawing, similar to image 1-A. However, the difference in brightness between crack K and its surroundings in image 2-A is small.

[0049] (Light area extraction: Step S12) Regions are extracted by extracting either light or dark regions. Light regions are extracted by binarization. Binarization can be performed using one of the following methods: simple brightness binarization, difference binarization which searches for points of change from a reference image, difference binarization performed with edge-enhanced images (Sobel filter, Previtt filter, Roberts filter, Laplacian filter, differential filter, etc.), connectivity and area binarization using blob detection, MIN filter, MAX filter, median filter, or simple brightness binarization after applying each edge-enhanced filter. Image 1-B shown in Figure 7 is an image in which crack K and its surroundings are separated (region separation, region extraction) from image 1-A by binarization. Images 2-B and 2-C shown in Figure 7 are images in which image 2-A has been binarized with different thresholds. Neither image 2-B nor image 2-C shows sufficient region separation. Binarization is a process that converts an image, composed of color elements and shades, into two colors, a first color and a second color, according to a first predetermined condition (for example, a predetermined threshold). In the following explanation, the bright areas extracted by binarization are shown as white as the first color, and the dark areas as black as the second color; however, the bright areas may be the second color and the dark areas as the first color.

[0050] As shown in image 1-A, cracks K that can be clearly imaged due to their large width, for example, have sufficient contrast with their surroundings. Therefore, as shown in image 1-B, cracks K in the binarized image are sufficiently accurately separated from their surroundings.

[0051] However, as in image 2-A, if the crack K is narrow or for other reasons, the captured image of crack K will be faint (low in brightness), and the crack K will be buried in the surrounding (background) noise. Therefore, even if we try to extract the region of crack K by binarization, we cannot extract a sufficient region, and the region of crack K will be fragmented, as in image 2-B. As in image 2-C, even if we adjust (lower) the binarization threshold to extract more of the crack K region, the surrounding noise region will also be detected (extracted).

[0052] [Crack fragmentation] Camera images captured using an image sensor contain various types of noise, including photon noise. Furthermore, the surface of the die D containing the crack K to be detected (extracted) is not necessarily uniform; it may exhibit periodic patterns such as memory cells, or variations in density in circuit regions. To extract a single connected crack K as a series of interconnected regions, the entire area of ​​the crack K must exceed the maximum brightness of the image background. While minute noise can sometimes be eliminated through filtering processes such as morphology, this section will explain cases where these methods are insufficient to remove the noise.

[0053] As shown in Figure 8, the brightness values ​​of the crack K region and its surrounding region are different. As indicated by arrow B in Figure 8, the brightness value (BV) of the region to be extracted (AE), such as crack K, is sufficiently different from the brightness value of the surrounding region. In this case, region extraction on the image is possible by setting a threshold between the lower limit of the brightness value of the region to be extracted (BEmin) and the upper limit of the brightness value of the surrounding region (BSmax). Here, Rth is the range in which a threshold can be set.

[0054] However, as shown by arrow C in Figure 8, if the lower limit of the brightness value of the region to be extracted is lower than the upper limit of the brightness value (BSmax) of the surrounding region, accurate region separation cannot be achieved regardless of how the threshold is adjusted. As a result, the region to be extracted may be missed, or noise may remain in the surrounding region, preventing the region from being extracted. For example, if the threshold is set slightly higher than BSmax, as shown in image 2-B, the region to be extracted will be mixed with the region that is not extracted, resulting in fragmentation of the crack region.

[0055] In addition to cases where the image appears faint (low brightness) when captured due to reasons such as narrow crack width, uneven surface brightness of die D caused by warping or bending of die D, and uneven illumination (image 2-A), the following cases may also occur.

[0056] As shown in Figure 9, the memory cells in the die D of the memory system are arranged periodically, so the surface of die D may appear as a striped pattern. In the HR image shown in Figure 9, a striped pattern is visible, consisting of rows of lines extending perpendicular to the horizontal direction of oblique illumination (directions of arrows Ax1 and Ax2) and in the direction of Y1-Y2. In this case, the crack K formed on die D is almost invisible. On the other hand, in the VR image shown in Figure 9, when the horizontal direction of oblique illumination (directions of arrows Ay1 and Ay2) is perpendicular to the illumination direction in the HR image (directions of arrows Ax1 and Ax2), the striped pattern is not visible, and the crack K becomes visible. Thus, the striped pattern is emphasized or suppressed depending on the direction of oblique illumination.

[0057] By selecting the direction of oblique illumination, the striped pattern on the surface of die D is suppressed. However, as shown in the VR image, the brightness of cracks K on the surface of die D is also affected by the periodicity of the striped pattern, and may appear fragmented.

[0058] Fragmentation of the crack region causes the following problems:

[0059] In the 2-B image shown in Figure 7, the fragmented crack K is not recognized as a single blob, but only as individual small blobs. Furthermore, in the 2-C image shown in Figure 7, the individual blobs are not significantly different in size or shape from the blobs generated by noise, making it difficult to distinguish them.

[0060] Therefore, fragmented detection regions (extracted regions) cannot be identified as a single label, and thus cannot be identified as a single crack K. Faint cracks K that fall below or above the threshold and are at the limit of the inspection sensitivity may not be detected, resulting in missed detection of cracks K. Thresholds can be used that are numerical representations of the size (area or length) of the bright region, roundness, or the state of projection. The state of projection is determined by performing projection processing on the x and y axes, and numerical representations are obtained of factors such as whether the shadow is long but the area is small, or whether the proportion occupied by the shadow is above a certain level.

[0061] (Area determination: Step S13) The area and shape (these are called regions) are determined through labeling processes, and the presence or absence of crack K is determined. Here, the area is the number of pixels. As shown in image 1-B in Figure 7, if a long bright region is extracted and the area of ​​that bright region is greater than or equal to a predetermined value (first threshold), it is determined that there is a crack. On the other hand, as shown in image 2-B in Figure 7, when the crack region is fragmented, if the area of ​​one bright region is smaller than a predetermined value (first threshold), it is determined that there is no crack.

[0062] Therefore, in this embodiment, for example, if the area of ​​one bright region is greater than or equal to the second threshold, a Hough Transform is used for determination. The second threshold is slightly smaller than the first threshold. Here, "slightly smaller" means, for example, about 5% to 20% smaller. If the area of ​​one bright region is smaller than the second threshold, it is determined that there are no cracks. In other words, if the number of target pixels within the region is extremely small (if all pixels are connected but do not reach the area threshold or length threshold) or if none of the labels meet the threshold for crack detection, the Hough Transform is not performed, and it is determined that there are no cracks. If the area is greater than or equal to the second threshold, there is a possibility that the cracked region is fragmented.

[0063] (Hough transform determination: Step S14) In this embodiment, even if the separated (extracted) crack regions are fragmented, they are considered as a single crack K if they are scattered along a predetermined linear path. Whether the scattered crack regions are along a predetermined linear path is determined using the Hough transform. In other words, if there are extracted regions that converge along a certain continuous linear path, a process is performed to determine whether that constitutes a single crack K. This determination is carried out using the Hough transform. Note that the Hough transform can also be applied even if the crack K is not fragmented.

[0064] [Hough Transform] As shown in Figure 10, the control unit 80 uses the Hough transform to search for the crack K.

[0065] (Determining the search area: Step S141) The area to search for linear bright regions (for example, white or black regions obtained through binarization) is determined.

[0066] (Setting search criteria: Step S142) The spacing of the bright areas, and the width and length of the search range for neighboring pixels are set. The spacing is the interval pixel threshold (IPT), which will be described later. The width of the search range for neighboring pixels is the search width (SW), which will be described later. The length of the search range is the join range (ML), which will be described later.

[0067] (White dot count: Step S143) When the background is set to black, pixels suspected to be cracks (K) are treated as white points (white pixels WP), and these white points are tallied. The tallied points are listed in terms of their number and the coordinates of each white point. The number of white points is also checked at this time. Note that white points may be connected or scattered. The label size may also be calculated when white points are connected. When the light region is set to black, the background is set to white, and pixels suspected to be cracks (K) are treated as black points (black pixels), and these black points are tallied. Hereafter, when the light region is set to black, white points will be read as black points, and white pixels as black pixels. White points (white pixels) and black points (black pixels) are also called candidate points (candidate pixels).

[0068] If there are too many white spots, it may be considered another abnormality and the die D may be judged as defective. Conversely, if there are sufficiently few white spots, it may be judged as a good product at that point.

[0069] (Line search: Step S144) The hypothetical line with the largest set of white dots can be found using the Hough transform. Here, the hypothetical line is a geometric shape, such as a straight line or a curve like a circle.

[0070] For example, when finding a straight line, the Hough transform uses θ rotation and the distance from a reference point such as the origin as variables to find the region with the highest density of straight lines passing through each white point. The region with high density is divided into intervals based on both θ and the distance from the origin, and a two-dimensional histogram is obtained. The interval SC with the highest count is identified as the first candidate. The straight line in the first candidate interval SC is the straight line being sought, and this line is called the provisional line TSL. The Hough transform can be interpreted as drawing a large number of provisional lines TSL, each divided into intervals based on θ and the distance from the origin, and then searching for a provisional line TSL with a larger cluster of white points.

[0071] [Searching for a hypothetical straight line] Figure 11 shows the target pixels in two-dimensional coordinates. As shown in Figure 11, an arbitrary straight line SL is drawn at an angle θ passing through the origin O. A perpendicular line PL is drawn for all white pixels WP, perpendicular to the straight line SL, passing through each white pixel WP. The straight line SL is divided into equal intervals, and the number of perpendicular lines PL passing through the white pixels WP in each divided section SC is counted. In the example shown in Figure 11, there are four perpendicular lines PL in one section SC and one perpendicular line PL in another section SC. The length of each section SC is assumed to be constant, but the length may be changed depending on the settings to correspond to the width of the cracks that may occur. Also, in order to determine the distribution more accurately, each section may overlap to a certain extent (overlap). This has the effect of preventing the number of perpendicular lines PL from being dispersed in adjacent sections due to a section boundary happening to be in the center of the distribution.

[0072] Next, the same process is performed by changing θ. Similarly, perpendicular lines PL are drawn through each white pixel WP, and areas with a high density (large number) of perpendicular lines PL are found. This process is repeated while changing θ within a range of 180 degrees (for example, from -90 to 90 degrees), and the θ and region in which the section SC with the highest density of perpendicular lines PL appears is found.

[0073] By making the interval SC finer, a more accurate angle can be obtained. However, this increases processing time. This denser set of perpendiculars PL becomes a candidate for the line to be detected.

[0074] The coordinates of the densest interval SC are given by equation (1) below, where R is the radius. x·COSθ+y·SINθ=R ···(1)

[0075] Using the left-hand side of equation (1), if we know θ and the coordinates (x,y) of the white pixel WP, we can find R, and thus easily create histograms for different values ​​of R. As shown in Figure 12, a two-dimensional parameter histogram is created for each θ and R, and the combination of R and θ that maximizes the density (number: N) of perpendicular lines PL is found.

[0076] Thus, the Hough transform for line detection is a method for finding the peak in a two-dimensional parameter histogram using θ and R. Therefore, the resulting two-dimensional parameter histogram essentially has resolution in both θ and R. Also, the line at the obtained peak is the maximum value in the interval SC, and is not an approximation formula (approximation line).

[0077] Since this is an interval histogram, identical intervals SC are considered as band-shaped regions. Therefore, even with the same combination of white pixels WP, multiple candidate intervals CSC can be generated. It is not easy to determine which is the central interval SC, i.e., the center line. Also, the maximum number of lines in the histogram is not limited to one location. The first line found among the maximum number of lines is considered the first candidate provisional line TSL.

[0078] As mentioned above, a straight line is represented by two parameters. On the other hand, a circle is represented by three parameters. A circle with center (p,q) and radius r is represented by the following equation (2). (XP) 2 +(yq) 2 =r 2 ...(2)

[0079] As mentioned above, in the case of a straight line, the search is performed by changing θ. On the other hand, in the case of a circle, the search is performed by changing p and q. If p, q and the coordinates (x,y) of the white pixel WP are known, then r can be determined. The same process as in the case of a straight line is performed to detect the circle.

[0080] (Calculation of the approximation line: Step S145) [Candidate section] As shown in Figure 13, there are multiple strip-shaped candidate segments CSC. For example, the three white pixels WP shown in Figure 13 are all included in each candidate segment. The first candidate segment CSC found is considered the first candidate. The extension direction (provisional line TSL) of the first candidate segment CSC found may deviate from the line that appears to be the center. Therefore, an approximate line ASL is drawn using the white pixels WP included in the first candidate segment CSC found, and this approximate line ASL is considered the center line.

[0081] In other words, since the provisional line TSL obtained in step S144 is not exactly an approximation line, the white pixels WP (neighboring pixels) that are judged to be close to the provisional line TSL are recounted, and the approximation line ASL of those white pixels WP is obtained using the least squares method or similar. In this case, since lines that are nearly parallel to the y-axis cannot be accurately determined by the least squares method, if the slope of the provisional line TSL is outside the range of, for example, ±45 degrees (when the slope is close to the y-axis), x and y are swapped in the calculation.

[0082] [Method for calculating the approximation line] As mentioned above, the initial hypothetical straight line TSL obtained by the Hough transform is inaccurate and prone to errors such as duplicate detection, so it is uncertain whether it is the optimal solution. Therefore, in this embodiment, the following is done.

[0083] First, as shown in Figure 14(a), the candidate lines for the provisional line TSL are narrowed down by narrowing the search width (SW), for example. A narrower search width (SW) can sometimes improve the accuracy of the desired provisional line. The five provisional lines TSL1 to TSL5 shown in Figure 14(a) were each searched using a narrow search width (SW). As shown in Figure 14(a), in each provisional line TSL1 to TSL5, the white pixels WP labeled "N" are neighbor pixels, and the pixels labeled "E" are not neighbor pixels. The candidate line with the most neighbor pixels is the provisional line TSL.

[0084] Furthermore, depending on the candidate line search algorithm, the crack shape (crack width, deviation from the straight line), and the imaging process, it may be possible to improve accuracy by widening the search width (SW) or by keeping the search width (SW) the same while determining the provisional straight line TSL.

[0085] The number of neighboring pixels for the first candidate line, provisional line TSL1, is "5". The number of neighboring pixels for the second candidate line, provisional line TSL2, is "7". The number of neighboring pixels for the third candidate line, provisional line TSL3, is "7". The number of neighboring pixels for the fourth candidate line, provisional line TSL4, is "7". The number of neighboring pixels for the fifth candidate line, provisional line TSL5, is "5". Therefore, the second candidate line, provisional line TSL2, is determined to be provisional line TSL. Provisional line TSL is determined based on whether a white dot pixel is a neighboring pixel, that is, based on the distribution state of the white dot pixels.

[0086] Next, as shown in Figure 14(b), the search width (SW) is optimized around the hypothetical line TSL to a width that takes into account the actual possible crack width and the deviation width relative to the line, and the neighboring pixels are recounted. Here, pixels that were not neighboring pixels in Figure 14(a) also become neighboring pixels, and the number of neighboring pixels is "11".

[0087] Finally, as shown in Figure 14(c), the approximate straight line ASL is obtained using the least squares method for the pixels that have been recounted and designated as neighboring pixels.

[0088] Furthermore, the approximation line obtained using the least squares method can only be found in equation (3) below. y = a·x + b ···(3)

[0089] In other words, it cannot handle equation (4) below, which covers all lines parallel to the y-axis. a·x+b·y+c=0 ···(4)

[0090] Therefore, as shown in Figure 15, although we would ideally like to find the approximate line ASL_D as the approximate line, the approximate line ASL_C is calculated instead. Thus, when the slope of the first candidate line is, for example, 45 degrees or more (θ is an orthogonal value so it must be 45 degrees or less), the x and y values ​​are swapped and the approximate line is calculated. The value of 45 degrees is just an example, and it does not necessarily have to be 45 degrees depending on the degree of variation in the straight line. It could also be divided into 30 degrees or 60 degrees. If there is a large discrepancy between the angle of the candidate provisional line TSL and the angle of the approximate line, then the x and y values ​​should be swapped.

[0091] (Calculation of combined pixels: Step S146) If the recalculated pixels meet the predetermined conditions, pixel merging is performed. A perpendicular line PR is drawn from the center of the recalculated pixels to the approximate line ASL, and the pixels to be merged are calculated from the distance between the foot F_PR of the perpendicular line PR and the adjacent foot F_PR, and the set interval (interval pixel threshold).

[0092] [Method for calculating combined pixels] The rectangles shown in Figure 16 represent adjacent pixels (white pixels WP). A square represents a single pixel, while a rectangle represents combined pixels.

[0093] Line extraction (pixel merging) is performed based on detected points (neighboring pixels) in the vicinity of the approximate line ASL, which is a pseudo-linear line.

[0094] For example, as shown in Figure 16, perpendicular lines PR are drawn from each neighboring pixel to the calculated approximate line ASL, and the intervals between the feet F_PR of these perpendicular lines PR are checked sequentially. If the interval is greater than the second predetermined condition, the interval pixel threshold (IPT), the pixels are not merged. For example, if IPT = 5, the interval IP of neighboring pixels shown in Figure 16 is greater than IPT, and the pixels are not merged. The calculation is performed regardless of whether the pixels are merged or not. This allows for a simpler calculation.

[0095] (Calculation of crack length: Step S147) The crack length is calculated from the distance between the two ends within a predetermined joining range (ML). For example, in Figure 16, the three blobs on the left are joinable. The four blobs on the right are also joinable. These joined blobs are crack candidates (abnormal candidates). The crack length KL1, which is the distance between the two ends on the left, and the crack length KL2, which is the distance between the two ends on the right, are determined.

[0096] Returning to the flow shown in Figure 6, in step S147, a substep of step S14, it is determined whether the crack length of the crack candidate obtained is greater than or equal to a third threshold length. If it is greater than or equal to the third threshold, it is determined that there is a crack (the crack candidate is a crack). If it is less than the third threshold, it is determined that there is no crack (the crack candidate is not a crack).

[0097] In the flow shown in Figure 6, step S13 is performed to determine whether or not there are cracks. Then, step S14 is performed to determine the Hough transform. Alternatively, the area determination in step S13, enclosed by the dashed rectangle in Figure 6, may be omitted, and only the Hough transform determination in step S14 may be performed.

[0098] The Hough transform described above performs a neighborhood check to determine if a point is in the vicinity of the hypothetical straight line TSL. The line used for the neighborhood check may be a curve or a branch line. In other words, a curve with a certain radius (R) may be used as a comparison reference to detect a curve, or a branch line with a certain angle may be used as a comparison reference to detect a branch line. This makes it possible to detect cracks K that tend to be non-linear.

[0099] Previously, cracks K were described as an example of damage formed on die D. The detection method of the embodiment can also be applied when the damage is a shallow groove such as a scratch, or when it is a linear pattern of dirt or foreign matter (fiber debris).

[0100] Visual inspection for detecting defects may be performed at least one of the following locations where die position recognition is performed: the wafer holder 12 of the wafer supply unit 10, the intermediate stage 31 of the intermediate stage unit 30, and the bond stage 46 of the bonding unit 40. However, it is more preferable to perform the inspection at all locations. If the inspection is performed in the wafer supply unit 10, defects can be detected earlier. If the inspection is performed in the intermediate stage unit 30, defects not detected in the wafer supply unit 10 or defects that occurred after the pickup process (defects that did not become apparent before the bonding process) can be detected before bonding. Furthermore, if the inspection is performed in the bonding unit 40, defects not detected in the wafer supply unit 10 and the intermediate stage unit 30 (defects that did not become apparent before the bonding process) or defects that occurred after the bonding process can be detected before bonding to stack the next die D, or before ejecting the substrate S. A subsequent camera (not shown) may be installed between the bonding unit and the substrate ejection unit for inspection. Furthermore, when a substrate is resupplied for a product that has already been bonded once, in order to perform lamination bonding again, a pre-inspection may be performed by a camera installed between the substrate supply unit and the bonding unit.

[0101] According to this embodiment, at least one of the following effects (a) to (e) can be obtained.

[0102] (a) The arrangement of the detected locations of the fragmented bright areas and background information allow the fragmented bright areas to be identified as a series of defects. This improves the sensitivity of defect inspection.

[0103] (b) As shown in the image 2-C of Figure 7, the impact of being buried in surrounding noise is reduced. This improves the sensitivity of the defect inspection.

[0104] (c) The effects of patterns such as memory cells in the background or uneven brightness on the die surface due to die warping or bending are reduced. This improves the sensitivity of scratch inspection.

[0105] (d) Production stability is improved because the sensitivity of defect inspection is improved.

[0106] (e) The inclusion of defective products in the laminated product into which die D is stacked is prevented. This improves the yield.

[0107] [Other forms] In the embodiment described above, in the calculation of the approximation line in step S145, a candidate line with a large number of neighboring pixels is selected from among multiple candidate lines as a provisional line TSL, the search width (SW) is changed around this provisional line TSL, and neighboring pixels are recounted. Based on the recounted neighboring pixels, the approximation line ASL is calculated. The calculation of this approximation line ASL may be omitted, and step S146 may be performed. In this case, the provisional line TSL is used instead of the approximation line ASL. That is, a perpendicular line PR is drawn from the center of the pixel to the provisional line TSL, and pixels to be considered as joined are calculated from the distance between the foot F_PR of the perpendicular line PR and its adjacent foot F_PR, and a set interval (interval pixel threshold). This allows for the detection of crack candidates.

[0108] Although the disclosures made by the Disclosers have been described in detail based on embodiments, the disclosures are not limited to the embodiments described above and are subject to various modifications.

[0109] In the embodiment, an example was described in which bright regions are used in the binarization process to separate cracks. Bright regions are areas to be separated as cracks, and specific values ​​in grayscale or color images may be used as bright regions (crack regions).

[0110] In the embodiment, an example was described in which the oblique light illumination devices 26a and 26b are arranged opposite each other. The oblique light illumination device 26b may be arranged so that its illumination direction forms an angle with that of the oblique light illumination device 26a. For example, the illumination direction of the horizontal illumination light IL of the oblique light illumination device 26a may be in the X1 direction, and the illumination direction of the horizontal illumination light IL of the oblique light illumination device 26b may be in the Y1 direction. The oblique light illumination devices are not limited to one pair; for example, in addition to the oblique light illumination devices 26a and 26b, an illumination device with the illumination direction of the horizontal illumination light IL in Y1 and an illumination device with the illumination direction of the horizontal illumination light IL in Y2 may be arranged to provide two pairs of oblique light illumination devices. Alternatively, only one of the oblique light illumination devices 26a or 26b may be provided.

[0111] In the embodiment, an example in which a film-like adhesive material DF is used was described. However, a preform section for applying a paste-like adhesive to the substrate S may be provided, and the adhesive material DF may not be used. For example, the preform section includes a preform head for applying the paste-like adhesive and a preform table for driving the preform head in the vertical and horizontal directions.

[0112] In the embodiment, an example of a semiconductor manufacturing apparatus was described in which a die D is picked up from a wafer supply unit 10 by a pickup head 21 and placed on an intermediate stage 31, and the die D placed on the intermediate stage 31 is bonded to a substrate S by a bond head 41. However, the apparatus is not limited to this, and can also be applied to a semiconductor manufacturing apparatus in which the bond head 41 picks up the die D from the wafer supply unit 10 and bonds it to the substrate S.

[0113] For example, this can also be applied to semiconductor manufacturing equipment that lacks an intermediate stage 31 and a pickup head 21, and instead uses a bond head 41 to bond the die D of the wafer supply unit 10 to the substrate S.

[0114] Furthermore, this can also be applied to a flip-chip bonder in which the die D picked up from the wafer supply unit 10 is inverted and handed over to the bond head 41, and the die D is bonded to the substrate S by the bond head 41.

[0115] The system may have multiple sets of pickup units, bonding units, and transport units (transport lanes), or it may have multiple sets of pickup units and bonding units, and only one transport unit. [Explanation of symbols]

[0116] 1. Die bonder (semiconductor manufacturing equipment) 24. Wafer recognition camera (imaging device) 26. Lighting equipment 80... Control device

Claims

1. It comprises an imaging device, an illumination device, and a control device. The control device is (a) The illumination device irradiates the die with illumination light and the imaging device photographs the die to acquire image data, (b) Image processing is performed on the acquired image data to extract candidate pixels that meet the first predetermined conditions, (c) Search for candidate lines based on the extracted candidate pixels and combine the candidate pixels to detect abnormal candidates. A semiconductor manufacturing apparatus configured to enable the following.

2. In the semiconductor manufacturing apparatus according to claim 1, The control device is The distribution state of the candidate pixels within the search width centered on the candidate line is determined, The aforementioned candidate pixels are combined to form a single abnormal candidate. Determine whether the candidate for abnormality is actually abnormal. A semiconductor manufacturing apparatus configured to enable the following.

3. In the semiconductor manufacturing apparatus according to claim 1, The control device is configured to divide θ and the distance from the origin into intervals, draw a straight line passing through each of the candidate pixels, and search for a straight line formed by the collection of the candidate pixels to find the candidate line.

4. In the semiconductor manufacturing apparatus according to claim 2, The control device is When determining the distribution state of the candidate pixels, the search width is changed to detect candidate pixels near the candidate line, and an approximation line is extracted based on the detected candidate pixels. When merging the candidate pixels, it is determined whether a second predetermined condition is met, and if the second predetermined condition is met, the candidate pixels adjacent to the approximation line are merged. A semiconductor manufacturing apparatus configured to enable the following.

5. In the semiconductor manufacturing apparatus of claim 4, The control device is After narrowing down the candidate lines, The search width is changed around the narrowed candidate line, and the candidate pixels adjacent to the candidate line within the search width are recounted. The candidate pixels are recounted and the approximation line is determined using the least squares method. A semiconductor manufacturing apparatus configured to enable the following.

6. In the semiconductor manufacturing apparatus of claim 5, The second predetermined condition is an interval pixel threshold, which is the interval between candidate pixels on the approximation line. The control device is configured to detect abnormal candidates by merging candidate pixels when the interval between each candidate pixel is less than or equal to the number of pixels of the interval pixel threshold, in a semiconductor manufacturing apparatus.

7. In the semiconductor manufacturing apparatus of claim 6, The control device is configured to determine that there is a scratch, foreign matter, or dirt if the length of the candidate pixels, which are coupled within a predetermined range along the approximation line, is greater than or equal to a predetermined length.

8. In the semiconductor manufacturing apparatus according to claim 1, The aforementioned illumination device is oblique illumination that irradiates illumination light at a predetermined angle with respect to the optical axis of the imaging device, in a semiconductor manufacturing apparatus.

9. In the semiconductor manufacturing apparatus according to claim 1, The imaging device is a camera that images the die attached to a dicing tape held in a wafer ring, in a semiconductor manufacturing apparatus.

10. It comprises an imaging device, an illumination device, and a control device. The control device is (a) The illumination device irradiates the die with illumination light and the imaging device photographs the die to acquire image data, (b) Image processing is performed on the acquired image data to extract candidate pixels that meet predetermined conditions, (c) Search for candidate lines based on the extracted candidate pixels and combine the candidate pixels to detect abnormal candidates. An inspection device configured to enable the following.

11. An inspection method in an inspection apparatus comprising an imaging device and an illumination device, (a) A step of irradiating the die with illumination light using the illumination device and taking a picture of the die with the imaging device to acquire image data, (b) A step of processing the acquired image data to extract candidate pixels that meet predetermined conditions, (c) A step of searching for candidate lines based on the extracted candidate pixels and merging the candidate pixels to detect abnormal candidates, Testing methods, including those mentioned above.

12. A method for manufacturing a semiconductor device, comprising the inspection method of claim 11.

Images