Open Access
Add to BibSonomy
Abstract
The quality-control process of polarizer production is hampered by the presence of extremely-slight transparent aesthetic defects (ESTADs). The saturated imaging method based on stripe structured backlight can effectively improve the imaging contrast of ESTADs. However, the contrast is very sensitive to the saturation degree, which requires careful manual selection. This paper presents a saturation level-guided image enhancement method that is simple to deploy in industrial settings. First, a new definition of the saturation level for structured backlit imaging with translation, scale, and rotation invariance is proposed. Then, an empirical model of contrast versus saturation level is established. Using the contrast data measured at five saturation levels, the optimal saturation level can be estimated using the parameter optimization method. The experimental results demonstrate that the method is effective, easy to use, and an improvement of imaging effects for transparent thin-film defect detection algorithms.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Polarizers, as a key component of liquid crystal displays (LCDs), are used in the screens of mobile phones, tablet computers, TVs, and other electronic products [1]. Generally, polarizers contain six layers of transparent films. Such filters have at least four common aesthetic defects: dents, foreign materials, bright spots, and scratches [2]. Transparent defects in polarizers may affect the visual perception of end users. Therefore, polarizer manufacturers and LCD panel manufacturers must detect polarizer defects before use. However, the current detection method is manual identification and marking; this method has very low detection accuracy and efficiency as well as a high cost. This approach cannot meet the demands of future industrial production. Machine vision technology can imitate the process of manual inspection and is widely used for the automatic nondestructive testing of products. For example, it can control the quality of various products [3], detect defects [4,5], and determine defect shape [6]. It is thus necessary to develop a vision-based testing technology for polarizer production.
Recently, research has been conducted on the visual detection of polarizer defects. Syu et al. developed a cost-efficient detection system for small concave and convex defects of polarizers [7]. Kim et al. designed a system for detecting multiple defects, such as bubbles, pits, threads, and alien substances [8]. Chou et al. reported an online, real-time system for detecting point and line defects in polarizers [9]. These studies have focused on standard defects that are easy to image, such as dents, bubbles, scratches, and dust. However, few reports have focused on the detection of extremely-slight transparent defects. We termed these defects extremely-slight transparent aesthetic defects (ESTADs) [10].
Detecting ESTADs poses at least two challenges. First, defect imaging is difficult. Under standard illumination, a defect has similar reflectivity and transmissivity with its surrounding background; thus, imaging contrast is very low. Second, the detection algorithm has stringent image quality requirements. Even an algorithm based on machine learning relies on a large number of defect samples, and the acquisition of effective samples is often difficult. Therefore, enhancing the imaging contrast for ESTADs is the reason for establishing an effective detector. There are two main approaches for enhancing imaging contrast. The first is to post-process the extracted images. Although the post-processing is simple to implement, it is easy to lose the original feature information of the interesting object during post-processing (such as denoising and gray scale equalization). The second is to adjust camera parameters before imaging to obtain a high-quality image. However, there is a lack of effective guidance to adjust camera parameters.
On the basis of the aforementioned considerations, in our previous study, we developed a method termed saturated imaging, which effectively improved the imaging contrast of ESTADs [11]. A method for estimating the 3D attributes of micro defects has also been proposed based on this method [12]. However, in practice, the saturation degree of stripe structured backlit images considerably affects defect imaging contrast, as do undersaturation and oversaturation. Therefore, the success of our previous work depended on the fine tuning of the exposure level, which hinders applicability in industrial automation.
In this paper, a method for improving the imaging contrast of ESTADs based on the saturation level is proposed. First, a method for evaluating the saturation level of structured backlit images is proposed. Then, an empirical model of the relationship between contrast and the saturation level is established based on a large number of sample data. After optimization is applied, the optimal saturation level of each defect sample can be obtained. Finally, the defect image with the most favorable contrast can be obtained by adjusting the camera parameters according to the optimal saturation level. The experimental results demonstrated that through this method, previous results of saturation imaging can be effectively applied to considerably improve defect imaging contrast, and the defect detection algorithm performance can be also improved.
The remainder of this paper is organized as follows. In Section 2, we explain the mechanism of saturated imaging to improve contrast and define the saturation level of stripe structured backlit images. In Section 3, the relationship between contrast and the saturation level is given, and a saturation level-guided method for identifying the optimal contrast is developed. In Section 4, experimental results are presented and discussed. Finally, conclusions are provided in Section 5.
2. Contrast enhancement using saturated imaging
In our previous research [11], it was confirmed that transparent aesthetic defects of polarizer have extremely low imaging contrast or cannot be even imaged under uniform illumination. By contrast, transparent defects can be imaged clearly in the black-and-white striped backlight [13], and may have higher contrast when the degree of imaging saturation is higher. Such imaging relies on the careful selection of the exposure level. To meet the needs of industrial automation, it is crucial to study relevant strategies for saturated imaging guidance to obtain the optimal defect contrast.
The overall configuration used for saturated imaging is the same as that in our previous study [10,13]. Our set-up is shown in Fig. 1. It is mainly composed of a phase-shiftable structured backlight source, image acquisition and control computer. The imaging acquisition system comprises a charge-coupled device (CCD) camera (Microvision MV-EM510M) with a resolution of 5.1 million pixels for image capturing. The CCD camera has a lens with a focal length of 8 mm (Computar M0824-MPW2). On the tablet computer, we generate the phase-shiftable binary stripe patterns.
Many approaches can be used to obtain saturated images, such as enhancing the brightness of the light source, increasing the width of the white stripes, extending the exposure time, expanding the aperture, and using a higher gain. Among them, adjusting the exposure time of the camera is the most straightforward method for achieving saturated images. When the exposure time is 10 ms, the image is underexposed (see for example Fig. 2(a)), and ESTADs are difficult to observe. With an increase in the exposure time, ESTADs can be imaged as a bright spot in the black stripe, and the imaging effect can be considerably enhanced. The clearest imaging effect emerges when the exposure time is approximately 100 ms, as shown in Fig. 2(d). However, when the exposure time reaches 200 ms, image noise begins to become prominent. When the exposure time is 300 ms, the brightness of the defect area reaches saturation, and the signal loss of the image is high. This is not conducive to subsequent defect detection, as shown in Fig. 2(e). Therefore, an appropriate saturation level can effectively improve ESTAD imaging contrast. This imaging theory is also applicable to dents, scratches, bubbles, and other defects. The image sequences of different dent saturation levels are shown in Figs. 2(f)-(j). In this section, we briefly discuss related mechanisms and define a mathematical indicator to describe the saturation degree for determining the maximum appropriate imaging contrast level.
Fig. 2. Sequential exposure images of two defect samples. (a∼e) represent Sample 1 with an ESTAD and (f∼j) depict Sample 2 with a dent. The exposure time, from left to right, is respectively 10, 50, 100, 200, and 300 ms. The defect sizes range from approximately 33 to 300 µm.
2.1 Mechanism
Every black-and-white striped image that we study has different values for width, orientation, and brightness. For a typical 8-bit grayscale stripe image, a pixel is completely saturated when the gray scale of the pixel reaches 255. The region with a gray scale lower than 255 is termed the unsaturated region, and whereas that with a gray scale equal to 255 is called the saturated region [14], as shown in Fig. 3. The noise in a stripe image increases with an increase in the saturation degree.
Figures 4(a)-(f) presents a sequence of stripe images with different exposure times. From left to right, the stripe images range from those that are underexposed to those that are overexposed. The grayscale histogram of the stripe pattern is displayed in Figs. 4(a*)-(f*). We can recognize the brightness saturation process of the stripe pattern from the histogram. In the entire saturated imaging process, the gray peak of the histogram gradually shifts to the right; the mean of the grayscale image gradually increases and approaches 255, and the area with a high brightness level also gradually increases. Figure 4 shows that the higher the mean value is, the higher is the image’s saturation degree. Therefore, the grayscale value of the stripe image may be a crucial indicator for measuring image saturation degree.
Fig. 4. Visualization of the image brightness saturation process. (a∼f) present the stripe patterns of different saturation levels, (a*∼f*) show the corresponding grayscale histogram, and E is the exposure time.
Chen et al. simulated the saturation effect of different fringe patterns by establishing a mathematical model and intuitively demonstrating the saturation process of sine fringe [15]. However, we use a binary black-and-white light source. Due to the polarization effect, when the binary stripe light passes through it, the image captured by the camera is no longer a binary pattern. The ideal step between black and white stripes becomes a continuous change, as shown in Fig. 5. Inspired by Chen’s work, we use a quadratic polynomial model and the periodicity principle to construct an intensity curve for the stripe pattern. The ideal 8-bit grayscale stripe pattern function in one period is formulated as follows:
Fig. 5. Simulation of black and white patterns. (a∼d) show stripe patterns with different saturation degrees; (e∼h) present the cross-sectional curves of corresponding stripes. From left to right, ${S_0}$ is taken as 1, 3, 5, 7, and 9, respectively.
The proposed mathematical model can effectively explain the saturation process of black-and-white striped images. We postulate that the enhancement of defect imaging is based on stripe saturation; that is, the approach to imaging saturation increases the diffuse light passing through ESTADs, thus enhancing defect imaging contrast. We also notice that the transparent microdefects of polarizers can be viewed as a microscale planoconvex lens model because of their physical shapes. A typical one is shown in Fig. 6.
Fig. 6. Saturated imaging of a sample defect (left) and its physical shape scanned by a laser confocal microscope (right).
To illustrate the contrast enhancement effect of defect imaging, we use the TracePro software program to configure an optical simulation system based on the lens model, including a similarly structured light source, a polarizer with point defects, and a receiving screen (for details on configuration readers, please refer to reference [11]). Image irradiance is altered by adjusting the brightness of the light source, such that the saturation process is simulated. The irradiance map results are shown in Fig. 7.
Fig. 7. Imaging results for defects at different brightness levels in the optical simulation: (a∼i) present the irradiance map and its cross-sectional curve of defect imaging. (a) shows the complete irradiance map and cross-sectional curve.
The simulation results of TracePro (Fig. 7) demonstrate that when the brightness of the light source is low, defects cannot be imaged. With an increase in the brightness of the light source, the defect begins to form a bright spot in the black stripe, and the defect and background gradually become brighter until they both reach maximum brightness. We compare the irradiance of the defect and background and studied the related contrast (Table 1).
Table 1. Defect simulation results.
The analysis and simulation results reveal that the black-and-white striped backlight can enhance the imaging contrast of ESTADs in the black stripe. However, optimal contrast cannot be achieved because of insufficient or excessive saturation. The previous evaluation of saturation degree is mainly based on human visual perception, which is subjective [11]. The exposure time or other imaging parameters are generally set according to experience. Thus, related research results cannot be applied to on-line industrial inspection, nor can they be generalized. As a result, we need to examine how to define saturation degree quantitatively and how to obtain optimal defect contrast based on the saturation degree.
2.2 Saturation level
According to our previous discussion, defect region contrast changes with the saturation degree of structured light based imaging. Optimal contrast is sensitive to the saturation degree. For color images, saturation refers to color richness. For gray images, saturation is when the brightness of the whole image approaches the highest grayscale value. Under different imaging conditions, such as image and object distance, the brightness of the light source, exposure time, and aperture, the images captured by the camera have different saturation degrees. For example, with other conditions unchanged, longer exposure time yields a higher degree of saturation. To our knowledge, the saturation degree of grayscale images has not been defined mathematically. In this subsection, we provide a definition of saturation degree for binary stripe backlit images.
Measurement of the image exposure degree has been widely used in research on multiexposure image fusion [17,18]. To our knowledge, previous studies on imaging exposure degree have all been based on Gaussian models. For example, Mertens et al. proposed a well-exposedness model to calculate natural color image exposure and used a Gaussian model to calculate the exposure degree of a pixel in each channel of the color image [17]. The purpose of these studies was to avoid underexposure or overexposure in the local area of an image. By fusing multiple images with different degrees of exposure, they could obtain images with sufficient clarity. To improve the dynamic range of camera imaging, Robertson et al. used a Gaussian model to calculate the exposure degree of an entire image, which was input as the mean grayscale value of images [18]. In addition, Que et al. used a Gaussian model to calculate the exposure degree of an entire image for exposure measurement [19].
An analysis of the aforementioned studies reveals that their calculation methods for the single-channel image exposure level can effectively measure the saturation degree of image brightness. Inspired by this, we determine that the saturation degree of binary stripe images can be defined as follows:
where M is the mean grayscale value of the normalized stripe image, and $\mu$ is the expected value set as 1. When the mean of the image is 1, the image has reached its highest saturation level. ${\sigma _g}$ represents the standard deviation. The mathematical indicator is a function of the mean grayscale value of the stripe image. Therefore the definition should be consistent with the camera response curve.For the normalized grayscale image, the saturation process is represented by a mean grayscale value increase from 0 to 1. Figure 8 shows the saturation level versus the mean grayscale value curves of different ${\sigma _g}$. In our set-up, when ${\sigma _g}$ takes the value of 0.3, the curve cloest to the camera response. This curve can reveal the saturation level of each grayscale stripe image.
In the saturated imaging experiment, the black-and-white striped backlight may have different phases, stripe width ratios, and rotations. Supose that the stripe period is $T = {W_b} + {W_w}$, where ${W_b}$ and ${W_w}$ are the width of the black and white stripes, respectively. To prevent these factors from influencing the calculation of the saturation level, we take two periods $2T$ and a fixed window length h along the direction perpendicular to the stripe and use Eq. (2) to calculate the saturation level, as shown in Fig. 9. On the basis of this operation, a translation, scale, and rotation-invariant definition of the saturation level can be obtained.
The backlight stripe configuration mainly includes the width ratio of black and white stripes $k = {{{W_b}} \mathord{\left/ {\vphantom {{{W_b}} {{W_w}}}} \right.} {{W_w}}}$, stripe period T and window length h. The value of k is empirically set to 1.3 for several reasons. First, in the case of saturated imaging, the black stripe area is gradually “eroded” as the saturation level increases. The mechanism can be found in our previous work [11]. Especially when the target is relatively small, higher saturation level is needed, then the erosion phenomenon is more serious. Our effective detection area refers to the black stripe area in the image. If k is too small, the effective detection area will become too limited to scan the field of view (FOV) efficiently. If the ratio is too large, on the contrary, the saturated imaging can not effectively improve the contrast as less luminous flux goes from white stripe to black stripe. Secondly, we use the three-step scanning method in our implementation, which is to move the stripe pattern once according to a fixed phase step (such as ${T \mathord{\left/ {\vphantom {T 3}} \right.} 3}$) after each scan. The purpose is to make the set of black stripes cover the whole FOV, otherwise missed detection probably occurs. Therefore the choice of k needs a proper compromise, and we found 1.3 is a good empirical choice. As for the stripe period T and window length h, we only consider two setting requirements. First, none of the window area goes out of image boundaries when performing arbitrary rotation. Second, they are not set too small because the saturation calculation is sensitive to noise when using small window size. In our configuration, there are nearly 14 periods of fringe in the FOV ($2560 \times 1920$), i.e., $T = 184$ pixels, ${W_b} = 104$ pixels and we set $h = 400$.
3. Optimal saturation level
Generally, with an increase in the saturation level, the defect image becomes clearer. However, when the brightness of the defect region is maximized, the higher saturation level reduces the imaging contrast, as shown in Fig. 7 and Table 1. Using the aforementioned definition of the saturation level, we attempted to determine a method for identifying the optimal saturation level to obtain the optimal imaging contrast for ESTADs. In this section, we propose an empirical relationship between the contrast and saturation level of defect imaging and then reveal the optimal saturation level through curve fitting.
3.1 Optimal saturation level
Defect imaging quality is closely related to imaging contrast. Adjusting the experimental parameters revealed that the higher the imaging contrast value of the defect, the easier it is for the defect to be detected; this improves the performance of the image processing algorithm. Contrast generally refers to the difference in brightness between regions in an image. The Weber contrast [16] approach is among the commonest methods for determining contrast. It is suitable for related calculations in a situation where both the target and image background are of uniform brightness or the brightness difference between them is obvious. In our case, when the brightness of the image reaches saturation, the ESTADs can be imaged as a bright spot on the black stripe, such that the target is obvious [11]. Therefore, we use the Weber contrast calculation method, as follows:
where ${G_d}$ is the maximum grayscale value of the defect region, and ${G_b}$ is the mean of the image background. The larger the value of C is, the better is the imaging contrast of the defect.The saturation level defined in Section 2 can be used. The relationship between imaging contrast and the saturation level of a typical ESTAD is shown in Fig. 10. With an increase in the saturation level, defect contrast gradually increases to the highest value, ${C_{\max }}$, and is then gradually reduced to zero. In other words, a saturation level exists that results in optimal defect imaging contrast. Defect contrast is not monotonous with respect to the saturation level. In Subsection 3.2, we present a mathematical model to establish the relationship between contrast and the saturation level for the identification of the optimal contrast level. Using the optimal saturation level to obtain the maximum contrast value causes the grayscale value of the defect region reach its maximum (i.e., 1).
Fig. 10. Typical imaging contrast versus saturation level curve of ESTADs. The defect width is 80 µm.
3.2 Contrast versus saturation curve
We also demonstrated that ESTADs of different sizes require different saturation levels to obtain the optimal imaging contrast. The relationship between contrast and the saturation level showed a certain degree of regularity through the experimental analysis of a large number of defect samples. Twenty ESTAD samples of different sizes are used to construct the contrast versus saturation level curve, as shown in Fig. 11. Contrast is sensitive to noise near the optimal saturation level. This is why it is difficult to manually select the optimal saturation level. In Fig. 11, the dots indicate the maximum contrast of the curve, and its radius is proportional to the width of the defect. The smaller the defect, the higher the saturation level required to optimize the contrast level.
Fig. 11. Contrast versus saturation curve of a group of real samples of different sizes. Different samples are distinguished by color. The position and radius of dots respectively indicate the maximum of the curve and the width of the defect measured using a laser confocal microscope.
We formulate the regulation in Fig. 11 using a three-parameter curve cluster as follows:
Consider a set, $S = \{{{S_1},{S_2}, \cdots ,{S_n}} \}$, which contains n saturation levels. In this set, we can test each sample to obtain its corresponding contrast, such that a contrast set is yielded as $C = \{{c_1^{}, c_2^{}, \cdots, c_n^{}} \}$. Then, according to Eq. (4), regression can be applied to obtain an optimal $\theta$ with regression loss ${L_2} - norm$. In other words, the curve parameters can be optimized using the following equation:
Using this method, we do not need to use a large number of saturation levels to scan defect samples but only a small number, say five, of scan. Our experience indicates that n is set as 5. More implementation details are discussed in Subsection 3.3.
3.3 Implementation issues
To fully implement the process of saturated imaging guided by the aforementioned saturation level indicators for polarizer defect detection, we recommend using the scheme explained as follows, and the scheme is shown in Fig. 12.
Step 1: Set the black stripe width ${W_b}$ and the width ratio k. Readers are referred to Subsection 2.2 for the setting guide.
Step 2: Adjust the exposure time, gain, and other parameters to ensure the saturation level meets n values $\{{{s_i}, i = 1, \ldots , n} \}$, respectively. Because the calculation of the saturation level is relatively straightforward, the efficiency for adjusting parameters can reach 20 Hz. Considering the coarse-to-fine strategy, usually, one round of parameter adjustments is completed within 200 ms. In our experiment, by setting n as 5 and $S = \{{0.1,0.2,0.3,0.4,0.5} \}$, favorable results can be obtained.
Step 3: According to the imaging parameter settings in Step 2, the image is acquired, and the contrast $\{{{c_i}, i = 1, \ldots, n} \}$ is calculated based on the aforementioned saturation levels.
Step 4: According to the optimization of Eq. (5), the optimal curve parameters and the optimal saturation level ${S^{opt}} = {\beta^{\ast}}$ are obtained. Because only five contrast inputs are applied in practice, we can discretize the input in advance to form a five-dimensional look-up table. Using the look-up table method can avoid the need for calculation optimization and can obtain a run with higher efficiency.
Step 5: Adjust the exposure time, gain, or other parameters to meet the optimal saturation level, ${S^{opt}}$, and capture a saturated image.
Step 6: Run defect detection, segmentation, and other post-processing algorithms to obtain the defect detection or segmentation results under this optimal contrast level. The 3D attributes of defects can also be estimated according to the 3D measurement technology proposed in reference [12].
Step 7: Move the stripe pattern of the backlight source by ${T \mathord{\left/ {\vphantom {T 3}} \right.} 3}$ along the direction perpendicular to the stripe, and loop back to Step 2 until totally three scans are completed. This is called three-step scanning [10,20] which ensures that the union of black stripes of all scans covers the whole field of view of the camera.
Step 8: Combine the defect post-processing results from the three-step scanning approach.
4. Results and discussion
The experimental hardware configuration in this study is the one mentioned in Section 2. The computing environment, MATLAB 2017b, runs on a machine with an i5-7400 processor and 8 GB of RAM.
In our experiment, the optical path is fixed, and the width ratio of black and white stripes is 1.3. By simply changing the exposure time of the camera, we obtained a series of imaging results for ESTADs at different saturation levels; eight samples of different sizes are shown in Table 2. The defect width ranges from 27 to 133 µm (measured using a laser confocal microscope), and the ESTAD regions are cropped to 50 × 50 pixels. Defects with a width of less than 33 µm are difficult to observe in images when the exposure time is less than 100 ms; such defects become clearer clear when the exposure time is 400 ms. Defects with a width of 40∼79 µm are difficult to observe in images when the exposure time is less than 50 ms. When the defect width is greater than 90 µm, only after an exposure time of over 50 ms can the defect be easily identified. Therefore, the larger the defect width is, the shorter is the exposure time required to obtain a clear defect image. By contrast, the smaller the defect is, the longer is the exposure time required.
Table 2. Saturated sequence image of ESTADs (image size: 50 × 50 pixels).
With an increase in exposure time, the noise of defect images [Table 2(a), (d) and (g)] gradually increases, affecting defect imaging quality. For this reason, we attempted to avoid overexposure.
These results demonstrated that defects of different sizes require different saturation levels to achieve optimal imaging contrast. This study revealed that for ESTADs, image uncertainty in the black stripe is mostly related to noise. Therefore, we used several blind-source image quality evaluation methods (i.e., signal-to-noise ratio (SNR), entropy, and standard deviation [21]) to evaluate the imaging quality of backlit stripes by increasing the saturation level. The image quality versus saturation level curve of a number of defect samples with different optimal saturation levels is shown in Fig. 13, where the dot represents the location of the optimal saturation level. The results showed that the image quality in the black stripe is not considerably affected without oversaturation.
Fig. 13. Curve for image quality and image saturation level of some defect samples. (a) image SNR, (b) image entropy, and (c) image standard deviation. The position of the optimal saturation level of defects is represented by colored dots, and their size is proportional to the defect width.
As for the process for determining the optimal saturation level proposed in Section 3, we conducted statistical analysis to determine the deviation between the optimized results of 52 samples and the real results (obtained by scanning in small steps on the saturation level), as shown in Fig. 14. Readers can download the sample set from https://doi.org/10.4121/ uuid:13cbec19-499d-430f-835f-c3c484bcd1e8. The optimal saturation level deviations are relatively concentrated, and the distribution is close to a Gaussian distribution. Most of the optimal saturation and contrast deviation values are within 0.03 and 0.3, respectively, which shows the effectiveness of the mathematical relationship between contrast and the saturation level. Because the optimal contrast is at the peak position, the optimal results are sensitive to noise. If sufficient computational power is available, we can scan at more different saturation levels. By contrast, in the implementation process of Step 3, the scanning position at different saturation levels runs independently, and parallel computing can also be used to improve efficiency.
We conducted confirmatory experiments for the improvement in defect detection performance using the guided method proposed in this paper. Taking the most straightforward target segmentation as an example, the segmentation method involves using the standard maximum class of variance method (Otsu algorithm) to obtain segmentation results, as shown in Fig. 15. In a certain optimal imaging saturation level range, that is, the red arrow region, the defect region can be precisely separated from the background. However, if it is not in this saturation level range, the defect cannot be detected. This demonstrates the effectiveness of our saturation level-guided method. To evaluate the detection performance, we expand the dataset with 52 negative samples (without defect). When using the pipeline of our proposed method and Otsu’s algorithm, no missed detection and false detection are found, i.e., true positive rate is 100% and false positive rate is 0%. Without using our proposed method, those defect areas cannot be successfully imaged. Otsu’s algorithm results in a recall rate of 0%.
As for the detection algorithm, numerous effective research results have been obtained for small target detection. Deng et al. proposed a method termed weighted image entropy to detect infrared micro targets [22]. Kang et al. reported a method of ship detection in remote sensing images based on a regional convolution neural network (CNN) [23]. Regardless of the type of detection algorithm used, high imaging quality is guaranteed with the successful application of automated visual inspection.
When our saturation level-guided method is applied to full field-of-view defect detection, we are unsure whether a defect exists or where it is at the beginning. Therefore, it is necessary to initially locate suspected defects. We suggest using a predetection approach that involves presetting a saturation level and using a lower segmentation threshold on the basis of the saturation imaging to obtain results with a low false-negative rate in order to locate the suspected defects. Moreover, the high false-positive detection rate caused by the low threshold is reduced after subsequent image enhancement.
If a large defect exists in the field of view, the saturation level calculation is affected. Therefore, the guided method proposed in this paper is unsuitable for contrast enhancement around large-size defects. However, this does not affect quality inspection, because the surrounding areas of large defects are cut before the polarizer film leaves the factory.
5. Conclusion
Optical simulations and experiments on saturated imaging have shown that using a stripe structured backlight is an effective method for improving the imaging contrast of ESTADs in polarizers. However, the image contrast of defect area greatly depends on the saturation level of imaging, which requires manual fine tuning. This approach is not suitable for automatic vision detection. In this paper, we propose a complete and applicable scheme. First, the saturation level of a structured backlit image is mathematically defined. The saturation level calculation is translation, rotation, and scale-invariant. Second, under the guidance of the saturation level, the optimal imaging contrast for ESTADs can be identified using a curve family model and through its parameter optimization. The experimental results showed that the proposed scheme can effectively improve the imaging without human intervention. Therefore, the scheme can be applied to automatic visual defect detection of polarizers and in the production of other thin films. Considering the difficulty in obtaining defect samples, a study is currently underway in which we use approaches such as generative adversarial networks [24] to generate a large number of effective samples.
Funding
National Natural Science Foundation of China (61571306); Science and Technology Planning Project of Shenzhen Municipality (JCYJ20180305123922293).
Acknowledgments
The authors would like to thank Shenzhen Shengbo Optoelectronics Technology Co., Ltd. who helps to collect the defect samples. The authors also acknowledge the anonymous reviewers who gave valuable suggestion.
Disclosures
The authors declare no conflicts of interest.
References
1. J. Ma, X. Ye, and B. Jin, “Structure and application of polarizer film for thin-film-transistor liquid crystal displays,” Displays 32(2), 49–57 (2011). [CrossRef]
2. C.-F. J. Kuo, C.-Y. Lai, C.-H. Kao, and C.-H. Chiu, “Integrating image processing and classification technology into automated polarizing film defect inspection,” Opt. Lasers Eng. 104, 204–219 (2018). [CrossRef]
3. S. S. Martínez, J. G. Ortega, J. G. García, A. S. García, and E. E. Estévez, “An industrial vision system for surface quality inspection of transparent parts,” Int. J. Adv. Manuf. Technol. 68(5-8), 1123–1136 (2013). [CrossRef]
4. C.-S. Cho, B.-M. Chung, and M.-J. Park, “Development of real-time vision-based fabric inspection system,” IEEE Trans. Ind. Electron. 52(4), 1073–1079 (2005). [CrossRef]
5. S.-H. Chen and D.-B. Perng, “Automatic optical inspection system for ic molding surface,” J. Intell. Manuf. 27(5), 915–926 (2016). [CrossRef]
6. W.-C. Wang, S.-L. Chen, L.-B. Chen, and W.-J. Chang, “A machine vision based automatic optical inspection system for measuring drilling quality of printed circuit boards,” IEEE Access 5, 10817–10833 (2017). [CrossRef]
7. H.-N. Yen and M.-J. Syu, “Inspection of polarizer tiny bump defects using computer vision,” in 2015 IEEE Int. Conf. Consum. Electron. (ICCE) (IEEE, 2015), pp. 525–527.
8. Y.-G. Yoon, S.-L. Lee, C.-W. Chung, and S.-H. Kim, “An effective defect inspection system for polarized film images using image segmentation and template matching techniques,” Comput. Ind. Eng. 55(3), 567–583 (2008). [CrossRef]
9. C.-F. J. Kuo, C.-H. Chiu, Y.-C. Chou, and Engineering, “Research and development of intelligent on-line real-time defect inspection system for polymer polarizer,” Polym.-Plast. Technol. Eng. 48(2), 185–192 (2009). [CrossRef]
10. Y.-l. Deng, S.-p. Xu, H.-q. Chen, Z.-h. Liang, and C.-l. Yu, “Inspection of extremely slight aesthetic defects in a polymeric polarizer using the edge of light between black and white stripes,” Polym. Test. 65, 169–175 (2018). [CrossRef]
11. C. Yu, P. Chen, X. Zhong, X. Pan, and Y. Deng, “Saturated imaging for inspecting transparent aesthetic defects in a polymeric polarizer with black and white stripes,” Materials 11(5), 736 (2018). [CrossRef]
12. Y. Deng, X. Pan, X. Wang, and X. Zhong, “Vison-based 3d shape measurement system for transparent microdefect characterization,” IEEE Access 7, 105721–105733 (2019). [CrossRef]
13. W.-w. Lai, X.-x. Zeng, J. He, and Y.-l. Deng, “Aesthetic defect characterization of a polymeric polarizer via structured light illumination,” Polym. Test. 53, 51–57 (2016). [CrossRef]
14. J. Hu, O. Gallo, K. Pulli, and X. Sun, “Hdr deghosting: How to deal with saturation?” in Proceedings of the IEEE Conf. Comput. Vis. and Pattern Recognit. (2013), pp. 1163–1170.
15. B. Chen and S. Zhang, “High-quality 3d shape measurement using saturated fringe patterns,” Opt. Lasers Eng. 87, 83–89 (2016). [CrossRef]
16. G. T. Fechner, D. H. Howes, and E. G. Boring, Elements of psychophysics (Holt, Rinehart and Winston New York, 1966).
17. T. Mertens, J. Kautz, and F. Van Reeth, “Exposure fusion: A simple and practical alternative to high dynamic range photography,” in Comput. Gr. Forum (Wiley Online Library, 2009), pp. 161–171.
18. M. A. Robertson, S. Borman, and R. L. Stevenson, “Dynamic range improvement through multiple exposures,” in Proc. 1999 Int. Conf. Image Process. (Cat. 99CH36348)(IEEE, 1999), pp. 159–163.
19. Y. Que, Y. Yang, and H. J. Lee, “Exposure measurement and fusion via adaptive multiscale edge-preserving smoothing,” IEEE Trans. on Instrum. Meas. 68(12), 4663–4674 (2019). [CrossRef]
20. Y.-l. Deng, S.-p. Xu, and W.-w. Lai, “A novel imaging-enhancement-based inspection method for transparent aesthetic defects in a polymeric polarizer,” Polym. Test. 61, 333–340 (2017). [CrossRef]
21. D.-Y. Tsai, Y. Lee, and E. Matsuyama, “Information entropy measure for evaluation of image quality,” J. Digit. Imaging 21(3), 338–347 (2008). [CrossRef]
22. H. Deng, X. Sun, M. Liu, C. Ye, and X. Zhou, “Infrared small-target detection using multiscale gray difference weighted image entropy,” IEEE Trans. Aerosp. Electron. Syst. 52(1), 60–72 (2016). [CrossRef]
23. M. Kang, K. Ji, X. Leng, and Z. Lin, “Contextual region-based convolutional neural network with multilayer fusion for sar ship detection,” Remote Sens. 9(8), 860 (2017). [CrossRef]
24. M. Frid-Adar, I. Diamant, E. Klang, M. Amitai, J. Goldberger, and H. Greenspan, “Gan-based synthetic medical image augmentation for increased cnn performance in liver lesion classification,” Neurocomputing 321, 321–331 (2018). [CrossRef]
