1. Introduction

Automatic inspection is becoming increasingly significant for industrial automation in manufacturing industries, with the community moving towards machine vision for precise control and maintenance of tight fabrication tolerances while allowing adaptability and flexibility [1]. Despite these advances, manual inspection by trained technicians is still commonplace even in the context of advanced, high-volume, high technology industries. This manual inspection limits throughput and quantitative consistency. To improve production yield and consistency, in-situ machine vision systems are becoming the new norm [2]. Placing these in-line vision systems early in the production workflow not only improves yield, but could provide data on tool wear, and allow for decisions about future processing steps on individual devices to be made, saving valuable resources throughout the production cycle [3,4].

Machine vision techniques can be as simple as photographing the manufactured object, detecting critical aspects of the geometry, and pattern matching for defects [5]. However, in many cases, critical defects can have dimensions that are too small to resolve from direct imaging. Some of these defects can be highly scattering features such as particles or scratches, while others can be broad but shallow surface undulations. The simplest vision systems for imaging scattering defects include dark-field microscopy and laser scatterometry [6], while pit and particle distinction can be achieved via polarisation analysis [7] or aspect ratio analysis of the defects with dark-field illumination [8]. If required, high-resolution microscopy techniques can then be implemented to determine the nature of the detected defects.

It is easy to consider the above techniques with a well-characterised base object, for example, an unprocessed, but polished silicon wafer. However, it can be beneficial to consider inspection at an even earlier stage of manufacturing, perhaps before a wafer is polished. The issue with placing such systems early in the production cycle is the device under inspection will most likely still exhibit surface features from the coarse machining. These can interfere with traditional inspection techniques when looking for defects that have a similar scale to the features. The simplest method to mitigate this issue is to avoid the background features by adaptive illumination [9]; however, this technique is only possible if there is some systematic pattern in the surface structure. The use of a single planar illumination, normal to an object with a systematic surface texture, has been shown to extract information about the surface features and defects [10]. As only one image is required for this technique it can be extremely fast, however, it can suffer from inconsistent imaging of subtle defects. Alternatively, collecting multiple images of a technical surface can be used to recover texture maps while enhancing defect signals. Reflectance Transformation Imaging is one such method where multiple images of a sample are taken and post processing allows for either selective reconstruction of the sample or relighting of the sample with adaptive illumination [11]. The extra information gained from multiple illumination angles can also be fused together to achieve a more complete representation of the defect [12].

While the surface inspection techniques mentioned above are well developed, and some deployed in advanced manufacturing facilities, these may not be suitable for all scenarios. In this investigation, we wish to detect subtle surface undulations with sub-micron depths on Hard Disk Drive (HDD) platters that are from early in the production workflow before coating and polishing. Platters at this stage of manufacture still exhibit systematic, coarse tooling marks from the lapping process which can cause preferential scattering depending on the illumination’s incident angle. Preferential scattering from such a surface can manifest as a bright strip across the platter when illuminated from a particular angle as illustrated in Fig. 1.

 

Fig. 1. (a) Schematic of the imaging system with a spherical coordinate system denoting the polar, θ, and azimuth, φ, angles of incidence. (b) Platter when illuminated with all 864 LEDs showing strong background scattering. (c), (d), And (e) show the platter when illuminated with a single LED at different azimuthal incident angles.

Download Full Size | PDF

Surface topology inspection methods can include the use of structured illumination techniques such as laser line scanning [13], phase deflectometry [14,15], and Fringe Projection Profilometry (FPP) [16,17]. These techniques typically require well defined specular or nominally diffuse surfaces to ensure predictable reflection of the illumination. The reflection properties, as well as the spatial orientation of a technical surface (such as tooling marks on a HDD platter) can cause preferential scattering, creating areas of high reflectivity which can negatively affect reconstruction efforts [18,19]. While these techniques are easily scalable, they are designed for extracting surface structures on the scale of the dimensions of the projection system, which is typically on the range of 10s of microns [20] and would be poorly suited for visualising subtle surface undulations with sub-micron depth across large a field of view.

For measuring changes in path length on the sub-micron scale, alternate approaches are required. There are many methods of recovering variations in path length on the scale of an optical wavelength by examining the optical phase. Such optical phase imaging is regularly applied within the field of biomedical microscopy, with only a few examples in industrial automation emerging such as Fourier ptychography [21,22] and Differential Phase Contrast (DPC) [23]. While DPC is traditionally used for imaging transparent samples, a technique has been reported showing how DPC can be used with diffuse, opaque, objects [24]. Such an approach may be able to reveal sub-micron surface undulations, however, implementing the approach of [24] would require construction of an illumination source which has inverse Fourier characteristics of the scatterers on the surface. This would be challenging for the structured tooling marks present on the HDD platters of interest. A technique is required that offers the simplicity of FPP, but with the sub-micron resolution of DPC.

In this paper, we present a visual inspection system that exploits systematic surface features to extract defect information from within a systematically textured surface. We show that it is possible to use an array of LEDs and by choosing which are illuminated we can achieve traditional dark-field illumination to reveal strongly scattering defects such as pits and scratches. Then using the same hardware, we show that it is possible to exploit the systematic tooling marks to achieve a form of fringe projection, but which can reveal subtle surface undulations. The utility of this approach is demonstrated by imaging different types of defects on unprocessed HDD platters, showing the potential for broader applications as a rapid, holistic, automated defect inspection system for any objects with systematically textured technical surfaces. All defects presented in this paper are from the manufacturing process and are not artificially produced by the authors.

2. Apparatus for imaging HDD platters with technical surfaces

Consider an imaging system designed to identify surface defects such as shallow scratches, sub-wavelength particles and pits on HDD platters as illustrated in Fig. 1(a). Here, the object is imaged with a bi-telecentric lens with a very narrow numerical aperture, accepting only light that enters the lens almost parallel to the optical axis. The object is illuminated at a polar angle (θ) outside the NA of the lens so that light which undergoes specular reflection from the surface should emerge at an equal and opposite angle to that of the illumination, and therefore will not be transmitted through the bi-telecentric lens. However, small defects on the platter’s surface will scatter light into a wide range of angles, with some of this scattered light entering the lens parallel to the optical axis such that it is transmitted and imaged by the camera.

In our case, the object is a 95 mm wide HDD platter. We wish to analyse this platter as quickly as possible and so choose to image the entire platter with a single field of the camera. We implemented the system illustrated in Fig. 1(a) using a bi-telecentric lens (Coolens DTCM430-190-AL, f/# = 7.5, diameter 230 mm), a 12 Megapixel monochrome camera (IDS UI-3200SE-M-GL, cropped to 2712×2712 pixels), and a circular array of blue LEDs which is positioned above the platter. Each LED acts as a distant point source and produces a broad beam of light that is sufficiently bright and uniform across the entire platter’s surface for an image of the platter to be taken with a 10 ms exposure. There are 864 individually addressable LEDs uniformly distributed about the circular array which we call a ‘halo’. The LEDs are reconfigurable at a rate of 1 kHz. The halo is positioned 200 mm above the platter resulting in a polar angle of incidence of 33° at the centre of the platter. We estimate a lateral resolution limit of approximately 35 µm which is limited by the camera’s pixel size when imaging the entire 95 mm platter. In our first measurement with the system, the platter was illuminated with all 864 LEDs simultaneously, with the recorded image presented in Fig. 1(b). The locations of the active LEDs are illustrated around the platter in blue (not to scale). When all LEDs are illuminated, the platter is not dark but is instead almost uniformly bright. It can also be seen that the platter has four arrows drawn on it, indicating the locations of known surface defects, however, there is no evidence of any significant difference in the recorded image in front of these arrows when compared to any other region on the platter.

To establish whether having individual control of the LEDs in the halo can be harnessed to aid defect detection, we imaged the platter with a single LED and recorded the image presented in Fig. 1(c). This figure shows a single bright strip across the platter. The strip extends across the platter from the location of the LED. This strip does not appear to align with the radius of the platter or the halo. We illuminated with a different LED on the other side of the platter, with the results presented in Fig. 1(d). This image shows a similar bright strip, again extending across the platter from the location of the LED. We again illuminated the platter with a third LED in between the previous two, Fig. 1(e). Here no bright strip is observed, but the top edge is slightly brighter than the rest of the platter. Close inspection of Fig. 1(e) reveals a few bright spots on the platter surface which may correspond to defects; however, these do not appear near the marked arrows.

The observation of the bright strips under certain azimuthal angles of illumination could be attributed to the surface having technical features that scatter light differently depending on the azimuthal angle of illumination. The fact that this scattered light in Figs. 1(c) and 1(d) forms an obvious bright strip spanning the platter suggests that the features on the surface are systematically structured and uniformly distributed. It is observed that no bright strips are present under some angles of illumination, indicating that the scattering structures are quite selective to the angle of incidence, Fig. 1(e). If we can predict the angles that will produce these bright strips of scattering, then it may be possible to avoid them to achieve the desired dark-field illumination. This illumination strategy may enable the detection of random defects that differ from the systematic pattern on the platter surface.

3. Traditional dark-field imaging of pits and scratches

3.1 Analysis of the systematic grind

In the previous section, we observed systematic scattering behavior when single LEDs illuminated a platter at different angles [φ in Fig. 1(a)]. We now explore the nature of the structure on the platter’s surface and how this leads to the systematic scattering seen in Fig. 1.

The high magnification, optical profilometer image of the platter presented in Fig. 2(a) shows regular scratches that are almost parallel. The scale of these systematic scratches is below the resolution limit of the imaging system, and we therefore do not expect to be able to resolve individual scratches. Illuminating perpendicular to these scratches should cause strong scattering, while illumination parallel to the scratches should not. Inspection of the grind pattern on the platter shows that these scratches are aligned tangentially to an arc with a large radius. We presume that this is perhaps the radial position of the platter relative to the lapping wheel on its final pass.

 

Fig. 2. (a) Optical profilometer image showing systematic grind and scratch defect. (b) Plot of normalized average intensity at the image sensor as a function of LED angle. (c) Geometric model used to describe the interactions between the illumination hardware and platter (Not to scale).

Download Full Size | PDF

We aim to find illumination conditions where the inherent background scattering is suppressed to conduct dark-field imaging to detect sub-pixel defects. Based on our analysis of the surface scratches, Fig. 2(a), we would anticipate that there would be a range of angles where the azimuthal angle of illumination would be aligned to the radius of some of the scratches on the surface. This condition would result in strong scattering captured by the bi-telecentric lens, while there would also be ranges of azimuth angles where the illumination was not perpendicular to any scratches on the surface. Thus, scattering would occur at an angle far from the optical axis and this would not be captured by the lens.

To test this prediction, we illuminated each of the LEDs in turn and recorded the intensity transmitted through the bi-telecentric lens to the camera. The average intensity of all the image pixels is presented in Fig. 2(b); as only a scalar value for the intensity is needed for each LED, the light in the imaging system could be tapped to a photodetector, allowing for more efficient data collection. Two ranges of azimuthal angles with significantly higher summed intensities are evident with very low intensity observed for the angles in between. This pattern can be explained as illustrated in Fig. 2(c), which shows the platter to be imaged and the halo of LEDs, which are not necessarily aligned. We assume the systematic grind pattern on the surface has the form of an arc [as indicated on Fig. 2(c) in purple, with origin (x_G, y_G)]. Consider radial lines from the grind origin that are tangential to the outer edges of the platter. These lines will intersect with the halo, defining segments of LEDs which can either illuminate the grind aligned with the radius (labelled as ‘Bright’), or regions where the LEDs produce light that are not aligned to the radii of any grind patterns on the platter (labelled ‘Dark’). As we assume a circular arc for the grind pattern, the radial lines, where aligned ‘bright’ incidence occurs, are axis-symmetric around the origin of the grind pattern which is assumed to be much larger than the radius of the halo. It can be seen from Fig. 2(c) that there will be two regions where there should be strong scattering collected by the bi-telecentric lens and that one of these regions covers a wider range of angles than the other, matching the observation in Fig. 2(b). The decrease in intensity observed in the middle of each bright range in Fig. 2(b) can be interpreted as due to the bright strip intersecting the spindle hole in the centre of the platter.

Figure 2(b) clearly shows the range of azimuthal angles over which the collected background scattering is suppressed and so it is now possible to proceed with dark-field imaging. However, we can also use simple geometry along with the information of Fig. 2(c) to determine the location of the origin of the grind pattern and the offset of the platter relative to the halo, information that may be useful in improving the manufacturing process.

3.2 Suppressing systematic scattering to achieve dark-field imaging

We now know that the unprocessed platters have a systematic grind pattern on the surface which results in predictable scattering responses. If it is possible to identify the azimuthal range of LEDs that avoid the bright response from the highly scattering surface texture, then it should be possible to illuminate the platter with this set of LEDs to achieve dark-field illumination. Under such an illumination scheme, it should be possible to identify defects in the systematic grind patterns as bright features on the otherwise dark image.

Using this methodology, and the platter from Section 3.1, we were able to identify that the grind radius was 624 mm with a platter origin of (−0.8 mm, 2.0 mm) with respect to the origin of the halo. Using the data available from Fig. 2(b), the azimuthal ranges of LEDs that result in dark-field illumination were identified (φ from 106° to 245°; and 297° to 71°), the platter was illuminated with these dark-field LEDs and the recorded image is presented in Fig. 3(a). Here, it can be seen that the platter is dimly illuminated with the four arrows drawn on the surface visible. Close inspection near the tips of these arrows [for example, the magnified image of Fig. 3(b)] reveals that there appears to be several circular defects. The optical profilometer image of shown in Fig. 2(a) includes one of these scratches. It can be seen that this scratch has a width less than 5 µm which is smaller than the pixel limited spatial resolution of 35 µm.

 

Fig. 3. (a) Results of dark-field illumination method to avoid background scattering with (b) magnified inset. (c) Binary image after Sobel edge detection process.

Download Full Size | PDF

To make it easier for machine vision to identify these defects, we employ an edge detection algorithm using the Sobel operator [25] with a global thresholding method used to binarise the output of the Sobel edge filtering step. Highlighted areas under a predetermined number of pixels were discarded to reduce the appearance of the background. The resulting processed image is presented in Fig. 3(c). Here the arrows on the platter can now be clearly seen outlined by the edge detection process, with little background noise evident. The circular defects are now clearly visible and there is a circular defect close to the tip of each arrow. These are the defects that the arrows intended to highlight, and thus we have successfully imaged these defects and could proceed to further classification steps. As a caveat, like the grind pattern, these defects will only scatter light into the imaging system when the azimuth illumination angle is perpendicular to the defect’s orientation. Hence, defects that are parallel to the grind pattern cannot be imaged using this dark-field technique, however it may be difficult to distinguish such features from the underlying grind pattern in any case.

4. Adapted fringe projection for imaging of surface undulations

4.1 Analysis of surface undulation defect

In Section 3, we have demonstrated that it is possible to analyse the systematic scattering features on an unpolished platter and use this information to quickly identify the azimuthal range of illuminations required to perform dark-field imaging. We have shown that this technique is capable of imaging particular surface defects that were not visible when using a uniform illumination method. However, we anticipate that there will be other types of defects that cannot be effectively imaged with this technique. Another platter is available with a different type of defect characterised by a surface undulation with lateral features larger than the resolution limit of the imaging system, but which are slowly varying. We would not expect this type of defect to be highly scattering, as was the case with the scratch defects imaged in Section 3; hence they may not be detected by our dark-field illumination technique.

To explore imaging this type of defect, we first configured the inspection system with all LEDs illuminated to achieve bright-field illumination as indicated in Fig. 4(a). A dark arrow indicating the nominal location of the defect is clearly evident; however, this bright-field illumination does not reveal any evidence of any defect in this location (Fig. 4(a) inset). We then followed the process outlined in Section 3 to obtain a dark-field image of this platter and produced the image presented in Fig. 4(b). There are some bright spots located near the tip of the arrow, but these are not easily distinguished from other bright spots around the platter. It would be difficult to base a robust defect classification system on this image alone.

 

Fig. 4. (a) Bright-field illumination of platter with the entire halo. (b) Dark-field method presented in Section 3. (c) And (d) show single LED bright-field images with the known defect visible. All insets are located at the expected defect location.

Download Full Size | PDF

If we consider that the expected defect is a low-frequency undulation, we might anticipate only a slight change in the angle of scattering from the systematic features on the platter. If this were the case, the defect would only be evident as a small perturbation close to the conditions where strong systematic scattering is already observed. We thus configured the illumination so that bright strips were aligned to the location in front of the arrow and look for evidence of perturbations. The results are depicted in Figs. 4(c) and 4(d) and indeed, there is evidence of a fingered or ‘wrinkle’ pattern for both illuminations (Figs. 4(c) and 4(d) insets). We can see that this pattern has complimentary responses when the bright strip is on either side of the defect, with the wrinkle appearing as a dark pattern sitting on a bright strip in Fig. 4(c) and appearing as a bright pattern on a dark background near the strip in Fig. 4(d). Evidently, when these are summed to produce the bright-field image of Fig. 4(a), the complementary images of the defect sum to be uniformly bright.

Figure 5 presents a linear contact profilometer measurement of this wrinkle defect with Fig. 5(a) showing the location of the measurement and Fig. 5(b) showing the height profile. This defect is found to have lateral periodic features on the order of 400 µm (larger than the spatial resolution of the imaging system) and a depth of approximately 400 nm. A method is required to differentiate the wrinkle defect from the bright background scattering.

 

Fig. 5. (a) Single LED bright-field image of defect, red line showing location of linear contact profilometer measurement (b) with A and B denoting direction of measurement

Download Full Size | PDF

4.2 Harnessing systematic scattering for adapted fringe projection

It is evident from Figs. 4(c) and 4(d) that surface undulations perturb the bright strips, which are a consequence of the preferential scattering, suggesting a possible means to extract these types of defects from within the grind pattern. Superficially, these bright strips look like the individual fringes that are projected onto the object of interest in traditional FPP to recover the height of diffuse, opaque surfaces. As we know from Section 3, the preferential scattering that causes the bright strips is highly predictable, and because the strips have a finite width, it should be possible to form a periodic ‘fringe’ pattern, like the projected patterns in traditional FPP, across the image of the platter, with the peak of each bright fringe corresponding to the azimuth location of an illuminating LED. If such a fringe pattern can be formed in the image, then it should be possible to use phase shifting techniques, similar to traditional FPP, to extract the surface defects perturbing this pattern.

Using the method in Section 3, we calculated the LEDs that correspond to significant scattering and discarded the angles convex to the grind, this left us with a half circle of LEDs from the halo. We then activated a sampling of LEDs so that one in every six was ‘on’ with the remainder remaining ‘off’ as illustrated in Fig. 6(a). The image captured at the camera that is presented in Fig. 6(a) appears as evenly spaced bright strips with peak intensity aligned with the illuminating LED. This image closely resembles the illumination patterns which are used in traditional FPP. As can be seen in the inset of Fig. 6(a), the wrinkle defect is faintly visible on one of the fringes. The arrow indicating the location of the defect is also clearly visible independent of whether it coincides with a bright or dark fringe.

 

Fig. 6. Adapted fringe projection achieved using multiple point source illuminations. (a) Shows the entire platter with ω = 0 illumination and (b) showing the shift in fringe pattern that generates the required illumination at the object plane for the 6-step FPP algorithm in Eq. (1).

Download Full Size | PDF

For phase recovery using phase shifting methods, multiple images are required to algebraically solve for unknown variables, with one being the phase, Ω(x, y), of the sample [26]. The general solution for the recovered phase for an N-step FPP algorithm to recover the height profile of the platter as given by:

The multiple images are taken under different illuminations where the fringes are shifted by an integer fraction, N, of the fringe period. The shift in the fringe pattern is achieved by rotating the LED illumination sequence of five off, one on, around the halo clockwise, one LED at a time. The resultant intensities for each of the six illuminations are presented in Fig. 6(b) and show the patterns closely resemble sinusoidal fringes with a consistent shift in the fringe pattern between adjacent illumination patterns.

The consistent shift in the sinusoidal pattern suggests that a phase shifting approach may be suitable to this application and may help to recover the wrinkle defect. With this illumination strategy, the six intensity images in Fig. 6(b) were used as inputs to Eq. (1) with the resultant output presented in Fig. 7(a). As is the case in traditional phase shifting approaches, the output of Eq. (1) is wrapped as the arctangent function is limited to between -π and π and a phase unwrapping algorithm [27] is required to remove the 2π discontinuities. A linear fit is applied to flatten the unwrapped phase across the platter for illustrative purposes with the resulting image, shown in Fig. 7(b), clearly showing that the wrinkle defect is visible above any background noise from the highly scattering surface texture. The marked arrow, while still visible, is faint but there is some evidence of the arrow in areas where it appears the ink may have pooled, causing a phase feature. Subtle, near horizontal lines are visible across that platter. These could be due to inconsistencies in the azimuth locations of the LEDs on the halo, resulting in relative offsets between the bright strips. The near vertical lines to the left of the platter are possibly caused by larger surface undulations than the wrinkle under inspection, but as these were not marked as defects, further investigation would be required to determine their nature. The same edge detection algorithm used in Section 3.2 (this time with a local thresholding method for binarisation) was run across the output of the phase unwrapping algorithm, non-flattened, with the binary output successfully extracting the defect in question, Fig. 7(c).

 

Fig. 7. Results for 6-step FPP algorithm. (a) Shows the wrapped phase as the output of Eq. (1), (b) is the unwrapped phase, and (c) shows the binary output from the edge detection algorithm showing the wrinkle defect has been successfully extracted.

Download Full Size | PDF

5. Conclusions

We have presented an approach for defect inspection which enables traditional dark-field imaging for visualising sub-pixel pits and scratches and an innovative adapted fringe projection imaging technique for extracting subtle surface undulations on unpolished samples. We have shown that by limiting the range of activated LEDs, the preferential scattering from the technical surface can be avoided to achieve traditional dark-field imaging, enabling the detection of sub-pixel scattering defects. Further, by using the same hardware and only reconfiguring the illumination scheme, we have shown that selective activation of LEDs which do produce strong scattering can result in periodic illumination patterns that are analogous to those used in traditional FPP, enabling extraction of subtle surface undulations that are on the order of an optical wavelength in depth. The utility of this approach is demonstrated by successful imaging of subtle surface undulation defects on unprocessed HDD platters from early in the production workflow. These techniques require less than ten images to be taken to image the entire platter in both dark-field and phase imaging modes with relatively simple post-processing showing potential for this approach in a practical manufacturing environment.