A novel computer vision approach to classify the mechanical stability of natural rubber latex concentrate

Abstract

Natural rubber (NR) latex is sensitive to mechanical influences which can occur at almost every stage of its manufacturing process. Moreover its mechanical stability can also change during storage or be modified with the addition of suitable soaps such as oleates, laureates, and stearates [1]. Hence the Mechanical Stability Test (MST) is of vital importance to the rubber industry as it gives an indication of the quality of the latex. Currently the assessment is performed manually by trained laboratory technicians, following the procedures as defined by the ISO 35 standard mechanical stability test. However, the test is highly dependent on the human visual capability and the experience of the laboratory technician performing the test, potentially leading to either inconsistent or inaccurate results. In this paper, we proposed a computer vision-based mechanical stability classification system to minimise the potential for biasness in the current standard test. We investigated this with a novel feature descriptor – Histogram of Size Distribution (HSD) that is based on the coagulum count and size. Experimental results demonstrated that the proposed system was able to provide essentially good classification accuracies on the data tested.

Keywords

Colloids natural rubber latex quality testing computer vision particle size distribution decision support

1. Introduction

Natural rubber (NR) latex is a complex, renewable biosynthesis polymer commonly extracted from the Hevea Brasiliensis tree. Its unique characteristics such as high strength, excellent elasticity and processability make it a useful raw material for mass production for a wide range of common products, for example surgical gloves, tyres, foley catheters, etc. [2]. Hence the quality of the raw NR latex which has a direct impact on the manufacturing processes as well as the performance of the end-product is of major industrial and economic importance [3]. Unfortunately, its quality varies with climate, temperature, rainfall, altitude, drainage of the soil, age and state of health of the tree [4]. Moreover unavoidable mechanical manipulations (shear force) introduced during either the pumping, transportation or processing of the concentrate can destabilise the NR latex significantly. Hence there is a need for a standard quality assessment in order to ensure that a particular batch of NR latex is suitable to meet the requirements for a specific application [4]. As rubber was planted on a commercial scale as early as 1877 [5], many of the processes, equipment and techniques used to manufacture rubber products have not changed much but are either still used or referred to, till this day. This is especially true for many of the testing equipment and techniques [6, 7, 8, 9]. To ensure the quality and processibility of the latex concentrates, various tests [6, 7, 8, 9, 10, 11, 12, 13] have been proposed and they essentially attempt to give a measure of the quality of the latex concentrate. These tests either exploit the rubbing or high-speed stirring mechanism as a mechanical influence to sufficiently degrade the quality of the NR latex. Apart from these techniques, a few studies have centred on the development of the test machinery [7, 9, 10, 12, 13] to measure the mechanical aspects of the NR latex. The first known piece of work on the estimation of latex stability using the aforementioned technique was with the Hamilton Beach Mixer [10]. Novotny and Jordon [6] subsequently proposed improvements to the mechanical stability test involving high speed stirring with a constant speed motor that is able to control the formation of the solids very carefully. Murphy [7] devised a test apparatus designed for a much simpler hand rubbing test under controlled conditions as a mechanical stability test for NR latex. The apparatus consisted of a moulded rubber nose which can be rotated with a sun and planet motion over a latex sample that had been spread onto a glass plate. The end point of the test was defined as the number of seconds the film of latex commerce to break up into small flocculum. The Crude Rubber Committee of the Division of Rubber Chemistry, American Chemical Society proposed a standard procedure for determining the mechanical stability of latex [11]. The effect of different variables, such as, total solid content, speed of stirring, temperature and size of the sample on mechanical stability time was studied by Dawson [8]. Maron and Ulevitch’s [9] contribution to this body of work involved a mechanical stability test that subjects the latex to a shearing force exerted by a metal disk under load, rotating in contact with a polyethylene surface. The coagulum formed in the operation of the machine for a definite time and under specified conditions was taken as a measure of the mechanical stability of the latex. Effectively, this gives a measure of the ability of the latex to withstand the colloidal destabilisation caused by the mechanical influences such as agitation. Gorton and Swinyard [13] proposed an alternative which uses a high-speed cone-plate viscometer to perform the stability test of NR latex. This can shorten the duration of the test as the device operates at a high and constant shear rate – which was achieved by rotating the cone at speeds of 1,000 rpm. The end point in this test is defined as the time for a 30% increase over minimum torque reading. This then gives a certain measure of how well the latex performs under such mechanical influences. In short, the existing approaches can be grouped into two main categories. The first category implements the hand-rubbing mechanism which has long been used by rubber technologists to obtain an approximate evaluation of latex mechanical stability. The end-point which is a direct indication of quality of NR latex is given by the time taken for the latex film to start breaking into small flocculus. The first sign of such flocculus is an indication of colloidal destabilisation. In practice, the hand rubbing test is more convenient and fairly fast. However, in order to have a more systematic and accurate mechanical stability indication, a relatively larger number of users have adopted the high speed stirring test rather than hand-rubbing test. Nevertheless, more recent investigations have focused on finding a more accurate, consistent and convenient metric for the end-point, for instance, the torque reading of the stirrer, the amount of flocculus formed for a definite duration, as well as the time required for occurrence of visible flocculation. However these may still suffer from high variability and subjectivity which will ultimately affect the test results. The predicament arises from the variations in the NR latex sample preparation, test procedure, apparatus used, test conditions, as well as different quality metrics implemented even due to the lack of consensus between the related technical personnel as well as the stakeholders. The usability and consistency of the mechanical stability test was greatly improved with the establishment of a standard test by the International Organisation for Standardisation (ISO) [14], videlicet ISO 35:2012 that aims to regularise the conditions and procedures for the test.

The standard test is widely adopted by the rubber-processing industry today. Nonetheless some challenges still remain. For example, the definition of the quality metric used in ISO 35 namely Mechanical Stability Time (MST) requires the individual analysts to spot the first occurrence of flocculation1

¹
Flocculation is a process in which particles adhere to form small clumps or masses.

and which is not easily quantifiable. The observation is made by the human naked eyes, something that is highly dependent on the operator’s visual capability as well as their experience, which can culminate in inconsistent or even inaccurate test results. Hence, we believe computer vision analysis of the latex samples can help minimize, if not eliminate many of these problems. Our main contributions in this work are twofold: a) a computer-vision based approach to provide a more robust and consistent characterisation of the mechanical stability for natural rubber latex concentrate and b) a new feature set that was used to help characterise the MST more effectively.

The rest of the paper is organised as follows. Some of the more significant prior work related to the use of computer vision are discussed in the next section. This is followed by Section 3 which describes the details of the proposed method. The next section after this includes details of the data collection as well as the experimental setup. Section 5 analyses the experimental results obtained. Finally, some conclusions are drawn in Section 6.

2. Prior work

Figure 1.

Proposed framework for MST classification.

Computer vision techniques have been successfully used in many fields [15, 16, 17, 18, 19, 20]. More specifically, there is some similar work done to investigate the use of computer vision to determine the size of objects. Kamel et al. [21] investigated the use of image processing to compute the particle size distribution of toulene droplets with microphotography under non-ideal lighting conditions. Each droplet was identified using a border following technique. Furthermore each toulene droplet was assumed to be elliptical and the areas were computed based on this assumption. The results were compared with those obtained from a commercial particle size analyzer – the OPTON TGZ-3. Particle size and the size distribution are also very important in many industrial sectors and Zhang et al. [22] were particularly interested to determine the size of coal particles. They studied 467 coal particles from four different sizes. Using a computer vision approach nine measurements, representing the width of the coal particles along various axes were computed, these were then compared with those they had obtained from physical measurements using nothing more than a pair of Vernier calipers. While they only managed to obtain 70% correlations with the physical measurements along most of the axes but they managed to achieve a maximum of 85% measurement accuracy along one of the axes. Shanti et al. [23] investigation focussed on determining the diameter of gravel which were verified against sieve analysis results. The objects were quite large and can be easily seen with the naked eye. The images were subjected to standard image processing techniques. For objects which are of much larger sizes, Nandi et al. [24] investigated the use of machine vision-based techniques for automated mango fruit sorting based on the size and maturity of the fruit. The maturity measurement was based on the RGB colour intensities using the Gaussian Mixture Model (GMM). The four different maturity categories of five different mango varieties were validated by three independent experts. These then form the ground truth. Unlike latex particles, the sizes of the mangoes are significantly larger. Other similar work on the use of computer vision to determine the size of objects involved airborne dust particles [25], grains [26, 27], wheat straw [28] and bacteria colonies [29]. Lai et al. [30] proposed hashing to encode an image into a vector of visual features followed by quantisation or projection to generate a new binary code which formed the inputs to a deep neural network to correctly identify a large set of different images. Nevertheless, their approach is significantly different for MST classification as we are looking for the first occurrence of large particles at each stage of the MST testing. While their approach is reported to be accurate, it is more sophisticated and large computing resources are required due to the more complex network architecture used to learn the input patterns. Approaches based on Deep Learning usually works well when there is huge amount of data to train the data and in the case of latex, to generate such a large amount data would not be feasible nor cost efficient.

In summary, computer vision has been used to determine particle sizes and this has been used to investigate a wide range of different objects of various shapes and sizes. Nevertheless this had been done with the sole purpose of computing the actual size of the objects. In this paper, we proposed a computer vision-based approach which can minimise much of the subjectivity and inconsistency in the quality assessment of NR rubber latex. Our earlier investigation centred on the feasibility of statistical features such as coagulum count and coagulum size distribution on the MST determination [31] for latex quality assessment with encouraging results that showed a good correlation with the manually-annotated MST. We are extending this by introducing a new feature, Histogram of Size Distribution (HSD) that is based on coagulum size distribution as part of our proposed computer vision-based approach to complement the ISO 35 standard procedure to determine the MST for NR latex more accurately and consistently. This can help eliminate human biasness in the results while providing a quantifiable measure of the quality.

3. Proposed method

This section explains the details of our proposed framework which is summarised in Fig. 1. It consists of 4 key steps with the first 3 processing the images for eventual classification.

As shown in Fig. 1, the first step is to binarise the input image by selecting an appropriate threshold to separate the object-of-interest – coagulum from the background pixels. Technically image binarisation depends purely on the histogram analysis of colour intensity and can be expressed as,

$\displaystyle B\left(x,y\right)=\left\{{\begin{array}[]{ll}1&\textit{if}\ I% \left(x,y\right)>T\\ 0&\textit{if}\ I\left(x,y\right)\leqslant T\\ \end{array}}\right.$ (1)

where $T$ is the selected binarisation threshold, with values ranging between 0 to 255. $(x,y)$ are the coordinates of image $I$ and $B$ is the segmented binary image. In order to identify and extract the various disjoint connected pixel regions, the morphological dilation operator, $S$ was utilized to process the binary image $B$ , pixel-by-pixel, i.e.

$\displaystyle w=(B\oplus S)\cap B\neq\emptyset$ (2)

where the dilation operator $S$ is a 3 $\times$ 3 kernel. The maximum intensity value among the pixels obtained by the structuring operator $S$ is also computed. All the adjacent pixels that carry the same value are then identified and marked with a similar label through the dilation process iteratively for $k=$ 1, 2, 3 $\ldots$ until all the pixels of $w$ have been labeled, i.e. when $w^{k}=w^{k-1}$ . Parameters such as the labels and the sizes of the connected regions are tabulated into a size distribution. The size distribution is then used to plot a histogram with appropriate binning. This HSD forms the feature set for the classification process. The basic reasoning behind the utilization of HSD is that MST can often be characterized rather well by the distribution of floculum size.

$P_{\textit{HSD}}=[p_{1},p_{2},\ldots,p_{n}]$ , and $Q_{\textit{HSD}}=[q_{1},q_{2},\ldots$ , $q_{m}]$ , represent the feature vectors of the reference set and test set respectively. Both data sets are fed into $k$ -Nearest Neighbour (kNN) classifier where the $k$ -features of $Q$ are identified in the reference data set, $P$ . This can be achieved by determining the $k$ -minimum statistical distances between the given feature vector, $q$ and all the other feature vectors, $p$ in the reference database i.e.

$\displaystyle D_{\textit{min}}=\mathop{\min}\limits_{n=1,2,\ldots,N}\{D(Q,P_{n% })\}$ (3)

where $N$ is the size of the reference data set. The statistical distance used is the Euclidean distance, i.e.

$\displaystyle\left\|Q-P\right\|=\sqrt{\sum\limits_{i=1}^{j}{(q_{i}-p_{i})}^{2}}$ (4)

where $j$ is the length of feature vector $p$ and $q$ . The reference feature vectors would have $k$ minimum distances for the given feature vector. Then, the most prevalent class which has the highest occurrence among the class of $k$ similar features is summarised, and the corresponding class label is treated as a match.

4. Experimental setup and data sets

4.1 Preparing NR latex for MST

The standard approach for the mechanical stability tests starts by preparing about 80 g (Fig. 2a) of the NR latex concentrate diluted with an ammonia solution to about 55% total solids (Fig. 2b). This mixture is then immersed into a water bath (Fig. 2c) to ensure that this is maintained at a constant temperature of 36–37 ${}^{\circ}$ Celsius. The diluted and warmed latex is immediately strained through a stainless sieve to remove any impurities (Fig. 2d) before being agitated with the Latex Test Machine (LTM) (Fig. 2e) which stirs the liquid at a speed of 14,000 rpm $\pm$ 200 rpm.

Figure 2.

Processes involved to conduct the MST test manually.

The stirring continues but samples of the mixture (Fig. 2f) were taken at regular intervals for analysis until the MST point is reached (Fig. 2g) [32]. Each sample was dropped into a petri-dish containing distilled water (Fig. 2f) and then placed into the imaging chamber with controlled lighting. Photographs of the latex samples in the petri-dish were then taken with a standard Digital Single-Lens Reflex (DSLR) camera fitted with a 60 mm f2.8 macro lens. This is repeated until signs of mechanical destabilisation, i.e. large number of obvious coagulum is observed. The whole process involving taking samples at regular intervals to identify the MST takes at least 20 minutes to complete. The data collected would involve three basic classes, viz., pre-MST, MST and post-MST.

4.2 Data sets

In the 50 laboratory experiments, we collected a total of 620 high-resolution images. Under the close supervision of an experienced laboratory technician, the entire data set is then carefully annotated into 3 classes: pre-MST, MST, and post-MST comprising 400, 87, and 133 images respectively. Due to the significant class imbalance in the data set which causes the learning process to be biased towards the majority classes, with the minority classes being ignored in machine learning, we down-sampled the two majority classes by randomly selecting the images to comply with the number in the minority class, resulting in a more evenly-distributed set of images for each class. The images were then cropped into a smaller size of 800 $\times$ 800 with a sliding window overlapping by 75%, resulting in a total of 7,569 images. Some examples of the images are shown in Fig. 3.

Figure 3.

Images of mechanical stability test in various classes: (a) Pre-MST, (b) MST and (c) Post-MST.

We then performed 70:30 cross-validation on these data to generate a reference set and test set respectively.

4.3 Data classification with k-Nearest Neighbour

The $k$ -Nearest Neighbour (kNN), a fundamental and simple classification method was first proposed by Fix and Hodges [33]. It is one of the oldest and simplest pattern classification algorithms, especially useful when there is little or no prior knowledge about the distribution of the data [34]. During the training phase the training vectors consisting of features in a multidimensional space and labelled with the respective class are presented and learnt by the classifier. In the classification phase, $k$ is a user-defined constant, and an unlabelled vector (a query or test point) is classified by assigning the label which is most frequent among the $k$ training samples and similar to that query point. This is based on the Euclidean distance function that measures the similarity for each pair of patterns.

5. Results and discussions

Figure 4 shows the results of using different binarisation thresholds on the MST test images and measuring the average classification accuracy. The best accuracy achieved was 91.45% for a threshold, $T=$ 190. This is the optimum point to separate the floculum from the background as this is the point where the floculum has a significantly higher pixel value than the background.

Figure 4.

Effect of binarisation threshold T on MST classification accuracy. This experiment is performed under optimal parameters: nearest neighbour, $k=$ 6, and bins of HSD, $n=$ 60.

Figure 5a shows the effect of number of bins, $n$ , on MST classification accuracy and from this figure, it may be seen that the classification accuracy of the MST data set improves with an increase in the number of bins before levelling off. By tabulating the coagulum in the range 1 to 4,000 pixels into 60 bins with a step size of 67 pixels per bin, the highest accuracy was achieved at 92.86%. Increasing the number of bins after 60 did not affect the classification accuracy, but it did increase the computational complexity. Undeniably, effective binning turns out to be essential for good performance affecting both accuracy as well as computational complexity. Hence an optimal bins of 60 was used throughout the study.

Figure 5b shows the effect of the number of clusters, $k$ on the MST classification accuracy. It can be seen that the value of $k$ has only a minimal effect on the overall classification accuracy for $k>$ 3.

Figure 5.

Effect of number of bins n (a) and number of nearest neighbour k (b) on MST classification accuracy. This experiment was performed under optimal parameters: binarisation threshold, $T=$ 190.

Figure 6.

Images of the latex concentrate for various MST classes.

Nevertheless, for each individual MST class, the accuracy increases sharply when $k$ was increased from 2 to 3 and shows only a slight increase after that. On the other hand, accuracy for pre- and post-MST only showed a slight drop for larger values of $k$ . However, the optimum point is when $k=$ 6 with the overall accuracy at its highest of 92.86%.

Figure 6 shows a sample of the images from the MST data set representing the three different classes, viz. pre-MST, MST and post-MST (Fig. 6a–c), their corresponding binary images as well as those which had been processed by thresholding. Besides the fixed-threshold binarisation with the threshold $T$ set at 190 (Fig. 6d–f), we also investigated Otsu’s adaptive thresholding method [35] which accounts for the intra- and inter-variations of an image. This adaptive thresholding is well known for its ability to adjust to a variety of global illumination levels found in different images. The images in Fig. 6g–i are the binary images from an adaptive threshold $T$ of 79, 127, 150 respectively. It is evident from Fig. 6d to i that the fixed-threshold approach gives a better binary representation than the adaptive thresholding method. Hence, we adopted the fixed-threshold binarisation for all the images.

Figure 7 shows the features based on HSD obtained from the binary images of the three different classes. It is obvious that the features are quite dissimilar for each different MST class as their characteristics are very distinctive even to the naked eyes. For some of the figures shown here, even though there may not be any particles detected in the other bins, there are always some particles in the first bin. Furthermore, when comparing the number of particles for this first bin, the lowest count is observed for the pre-MST and the highest from MST. The Post-MST features, on the other hand, have a wider distribution of particles of different sizes and not just in the first bin. Nonetheless, it can be seen that for each class, the highest number of particles are in the first bin.

Table 1 shows the confusion matrices for the MST classification accuracy obtained with the optimal parameters. For each category, classification accuracies were obtained for both fixed binarisation threshold and Otsu’s adaptive threshold. These are shown in Table 1a and b. The overall classification accuracy for fixed binarisation threshold and Otsu’s adaptive threshold were 92.86% and 85.45% respectively. Again, the results demonstrated that an effective fixed binarisation threshold performs significantly better than the adaptive threshold in the MST data set for individual class accuracies as well as overall classification accuracy. Amongst all classes, the highest accuracy achieved is 97.07% for pre-MST followed by post-MST classification at 96.03%.

Table 1

Confusion Matrix for MST Classification accuracy

(a) Fixed Threshold
	Classification accuracy
	Pre-MST (%)	MST (%)	Post-MST (%)
Pre-MST	97.09	14.15	1.46
MST	2.91	85.45	2.51
Post-MST	0.00	0.40	96.03

(b) Adaptive Threshold
	Classification accuracy
	Pre-MST (%)	MST (%)	Post-MST (%)
Pre-MST	83.73	19.18	4.63
MST	12.43	79.23	1.98
Post-MST	3.84	1.59	93.39

This experiment was performed using the optimal set of parameters: nearest neighbour, $k=$ 6, and bins of HSD, $n=$ 60. For the fixed binarisation threshold, $T=$ 190.

Figure 7.

HSD generated from different classes.

We used the principal components to further analyse the results of these Mechanical Stability tests. Principal Component Analysis (PCA) is a common technique which has been used for dimension reduction [36] and pattern analysis [37]. It is a feature extraction technique that generates new features which are linear combination of the initial features [38, 39]. PCA maps each instance of the given dataset present in a $d$ dimensional space to an $m$ dimensional subspace such that $m<d$ . The set of $m$ new dimensions generated are called the Principal Components (PC) and each principal component is directed towards maximum variance excluding the variance already accounted for in all its preceding components. Subsequently, the first component covers the maximum variance and each component that follows it covers lesser value of variance. In general, the Principal Components $C_{i}$ can be represented as,

$\displaystyle C_{i}=\sum\nolimits_{j=1}^{d}{\alpha_{j}X_{j}}$ (5)

where $C_{i}$ – Principal Component $i$ ; $X_{j}$ – original feature $j$ ; $\alpha_{j}$ – numerical coefficient for $X_{j}$ ; $1\leqslant i\leqslant m;1\leqslant j\leqslant d$ .

Since the objective is to have the greatest possible variance, this can be achieved by selecting the most appropriate value of the numerical coefficient, $\alpha$ with the constraint that the sum of the squares of elements of $\alpha$ equals to one. With the right values of $\alpha$ , the initial data values in $X$ can now be transformed to a lower dimension in $C$ . There are several well known methods to perform this transformation to compute the principal components, for example, eigenvalue decomposition, the Kosambi-Karhunen–Loève (KL) transform [40], etc. Nevertheless we adopted the more common approach to determine the principal components with singular value decomposition (SVD).

Using only the first two components generated from the PCA, the visual maps of each MST data set are shown in Fig. 8. There is a clear distinction between the individual classes which indicates the effectiveness of the HSD features selected. In the case of the pre-MST class it is nicely clustered at the bottom left corner of the map in Fig. 8a. They are not as scattered as the other two classes. This may explain why the pre-MST class always has the best classification accuracy. Moreover there is also a better segregation of the data for the other two classes – as shown in Fig. 8b and c. Nevertheless MST has the lowest classification accuracy at 85.45%.

Table 2

Comparison of classification accuracies for various features

Class	Features
	HOG (%)	SIFT (%)	Wavelet (%)	HSD (%)
Pre-MST	90.34	70.63	97.08	97.09
MST	96.56	77.51	93.25	85.45
Post-MST	73.54	75.26	88.10	96.03
Overall	86.82	74.46	92.81	92.86

Figure 8.

Visualisation of MST data set as scatter plots.

Misclassification of MST as pre-MST constitutes the largest error overall – accounting for 14.15%. The reason behind this can be explained when the entire data set is visualised as shown in Fig. 9. We can now see that there is a significant amount of overlap between the pre-MST and MST classes at the bottom left corner leading to the misclassification error of MST as pre-MST that was previously shown in Table 1.

Figure 9.

Visualisation of overall MST data set on scatter plot.

We also investigated how the proposed HSD performs against other sets of features generated by three common techniques.

5.1 Histogram of oriented gradient (HOG)

The first one that we investigated generates a robust set of features based on a locally normalised HOG which has been extensively used for object detection [41]. This set of feature is based on evaluating a well normalised local histogram of image gradient orientations in a dense grid. Essentially Dalal and Triggs showed that HOG can be a good feature set to represent an object as local object appearance and shape can be characterised by the distribution of local intensity gradients. The HOG feature set is generated by dividing the image into small regions and accumulating a local one – dimensional histogram of gradient directions of the colour intensity for each region.

5.2 Scale invariant feature transforms (SIFT)

The next set of features investigated were widely used with high object recognition accuracies. SIFT basically involves generating a set of features sampled at a large number of repeatable locations using a staged filtering approach. With these features a simple nearest neighbour classifier was able to reliably identify the correct objects [42]. SIFT features were generated for the MST data.

5.3 Wavelets

The final technique was inspired by the work of Yu and Slotine [43] who developed an algorithm for the rapid categorisation of various types of images. Their proposed algorithm has a two-layered architecture where the first layer performs a wavelet transform followed by grouplet transforms in the second layer. It was able to generate a set of features which are both scale- and translation-invariant, achieving results which exceeded those from other state-of-the art techniques in object recognition. This was also used to generate a set of features from the MST data sets.

Feature sets generated from each of these three techniques were tested with kNN and their results summarised in Table 2. Amongst the features considered in this work, the proposed feature, HSD gave the best overall performance of 92.86% amongst the four features investigated.

Moreover, the HSD performed significantly better than any of the other 3 techniques for the pre- and post-MST tests. For these two classes, the average classification accuracy for the HSD features is 96.56% (97.09% pre-, 96.03% post-) compared with 92.59% (97.08% pre-,88.10% post-), 72.95% (70.63% pre-, 75.26% post-) and 81.94% (90.34% pre-, 73.54% post-) for wavelet, SIFT and HOG respectively.

The proposed technique offers a number of significant advantages over the existing manual MST determination. Firstly, it eliminates the subjectivity and inconsistency caused by human biasness, yielding an independent and yet more robust set of test results. Besides, it can assist the laboratory technician to identify the MST end point. Additionally, the proposed technique is simple and cost-effective, with minimum additional hardware needed – basically a DLSR camera and normal computer. Most importantly, the method fully complies with the ISO 35 testing standard for determining the mechanical stability of NR latex concentrate.

6. Conclusions and future work

Even though several techniques may be used to test the mechanical stability of the latex concentrate, a commonly used test that is widely adopted by the rubber industry is the one based on ISO 35 standard. Nevertheless there is a significant amount of subjectivity in identifying the end point due to either inconsistency in interpreting the results or human errors that are related to the human visual system. In this paper we proposed a computer-vision based solution to help determine the mechanical stability for natural rubber latex concentrate.

The proposed HSD feature set provides a good balance between classification accuracy with computational efficiency by exploiting the information on the distribution of coagulum size, generated from the ISO 35 standard that is widely used to determine the quality of NR latex. Our proposed computer-vision based approach is more superior to the conventional manual approach as it assists the human operators in their tedious manual inspection task. Only non-specialised and common equipment is needed to capture the images during the MST testing. Besides, the computation needed to generate the set of HSD features is neither huge nor requires additional or specialised computing hardware. The set of HSD features is also intuitively useful for human interpretation if necessary.

The results compares quite well too with the results of three other modern methods used to generate the feature sets. Among these three other methods investigated here, Yu and Slotine’s [43] two-layered wavelet-grouplet transform for image classification is known to be able to generate a set of features which are both scale- and translation-invariant. It has been reported that their two-layered wavelet-grouplet transform performed quite well, with results that exceeded those from other state-of-the art techniques in object recognition. And yet, our experimental results also showed that the HSD feature set was capable of differentiating the various MST classes with an average overall classification accuracy of 92.86%, surpassing the two-layered wavelet-grouplet transform and other feature sets. Furthermore due to its the implementation simplicity, the proposed HSD technique can facilitate cost-effective computerised mechanical stability test of natural rubber latex concentrate.

Despite having obtained encouraging results with this new approach there are several potential areas for improvement. Some preliminary work involving multi-thresholding of the MST images with artificial bee colony (ABC) have shown promising results [44], and this is something that can be explored further. Finally, we would like to investigate this approach on other types of colloids that are commonly used to manufacture a variety of industrial products.

Footnotes

Acknowledgments

The authors would like to acknowledge the support of Ministry of Education (MOE) Fundamental Research Grant Scheme under project number (FRGS/1/2014/TK04/TARUC/02/1) for the work reported here. The authors would also like to thank Chooi Pheng Khoo and her technical team at the R&D Department of Weir Minerals (M) Sdn. Bhd. who have unselfishly spent their time and effort to help us collect the data described in this paper. They have also been very patient in assisting us to identify the three main classes of the MST tests for NR latex concentrate on the images collected.

References

Namboodiripad

Sekhar

. Improvements in or relating to the preparation of latex concentrate by centrifugation, creaming or evaporation: WIPO, WO2005023875 A1[P]. March 17 2005.

Wang

Peng

Lin

Zhang

She

. The impact of esterification on the properties of starch-natural rubber composite. Composites Science and Technology, 2009; 69(11-12): 1797-1803.

MATRADE (2016), Malaysia’s Trade Statistics. Available from: http://wwwmatrade.gov.my/en/malaysian-exporters/services-for-exporters/trade-a-market-information/trade-statistics. [Accessed: 26

{}^{\rm th}

October 2016].

Blackley

. Investigation of some latex properties. In: Polymer Latices Science and Technology, 2

{}^{\rm nd}

ed. London, UK: Chapman & Hall, 1997, pp. 426-434.

Rubber plantation in Malaysia. Available from: https://www.scribd.com/doc/44125674/Rubber-Plantation-In-Malaysia, [Accessed 8

{}^{\rm th}

March 2017].

Novotny

Jordon

. Ind. Eng. Chem., Anal. Ed. 1941; 13: 189.

Murphy

. Proc. Rubber Tech. Conf. 1938, London. 159.

Dawson

. Mechanical Stability Test for Hevea Latex. Analytical Chemistry, Sept. 1949; 21(9): 1066.

Maron

Ulevitch

. Mechanical Stability Test for Rubber Latices. Analytical Chemistry, July 1953; 25(7): 1087.

10.

Noble

. Latex in Industry. Palmerton Publishing Co., New York, 1936, 181.

11.

Crude Rubber Committee, Div. of Rubber Chemistry. Rubber Chem. Technol., 1941; 14: 290.

12.

Amran

MKAB

Mansor

MNB

Othman

. Method of quantitative measurement of mechanical stability time (MST) of latex suspensions and the apparatus for use in the method. WO Patent App., PCT/MY2011/000050[P], 22 Nov. 2012.

13.

Gorton

ADT

Swinyard

. The measurement of latex viscosity and stability using the cone and plate viscometer. NR Technology, 1979; 10(4): 73.

14.

Natural Rubber Latex Concentrate – Determination of Mechanical Stability. ISO Standards 35: 2012.

15.

Zhu

Lei

Liu

Shi

. Face Alignment Across Large Poses: A 3D Solution. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, June 2016, 146-155.

16.

Ofir

Galun

Nadler

Basri

. Fast Detection of Curved Edges at Low SNR. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, June 2016; 213-221.

17.

Zhu

Zanibbi

. A Text Detection System for Natural Scenes with Convolutional Feature Learning and Cascaded Classification. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, June 2016; 625-632.

18.

Qiao

Liu

Shen

Hengel

AVD

. Less is More: Zero-Shot Learning from Online Textual Documents with Noise Suppression. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA. June 2016; 2249-2257.

19.

Yang

Price

Cohen

Lee

Yang

. Object Contour Detection with a Fully Convolutional Encoder-Decoder Network. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA. June 2016; 193-202.

20.

Shen

Zhao

Jiang

Wang

Zhang

Bai

. Object Skeleton Extraction in Natural Images by Fusing Scale-Associated Deep Side Outputs. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, June 2016; 222-230.

21.

Kamel

Akashah

Leeri

Fahim

. Particle Size Distribution in Oil-Water Dispersions Using Image Processing. Computers and Chemical Engineering, 1987; 11(4): 435-439.

22.

Zhang

Yang

Ding

. Analysis of Large Particle Sizes Using A Machine Vision System. Physicochemical Problems of Mineral Processing, 2013; 49(2): 397-405.

23.

Shanthi

Porpatham

Pappa

. Image Analysis for Particle Size Distribution. International Journal of Engineering and Technology, 2014; 6(3).

24.

Nandi

Tudu

Koley

. Machine Vision Based Techniques for Automatic Mango Fruit Sorting and Grading Based on Maturity Level and Size. In: Sensing Technology: Current Status and Future Trends II, Smart Sensors Measurement and Instrumentation 8, A. Manson et al. (eds.), 2014, 27-45.

25.

Mazzoli

Favoni

. Particle size, size distribution and morphological evaluation of airborne dust particles of diverse woods by Scanning Electron Microscopy and Image Processing program. Journal of Powder Technology, 2012; 225: 230-244.

26.

Pan

. Cereal Grain Size Measurement based on Image Processing Technology. Proceedings of the International Conference on Intelligent Control and Information Processing, 2010; 209-212.

27.

Peregrina-Barreto

Terol-Villalobos

Rangel-Magdaleno

Herrera-Navarro

Morales-Hernández

Manríquez-Guerrero

. Automatic grain size determination in microstructures using image processing. Measurements, January 2013; 46(1): 249-258.

28.

. Particle Morphology Analysis of Biomass Material Based on Improved Image Processing Method. Journal of Analytical Chemistry. 2017; Article ID: 5840690, 9 pages.

29.

Yoon

Lawrence

Park

. Automatic Counting and Classification of Bacterial Colonies Using Hyperspectral Imaging. Food and Bioprocess Technology, October 2015; 8(10): 2047-2065.

30.

Lai

Pan

Liu

Yan

. Simultaneous Feature Learning and Hash Coding with Deep Neural Networks. in: Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition (CVPR2015), 2015; 3270-3278.

31.

Tan

Chan

Lai

Chew

Chua

. A Novel Particle Characterization Technique in Enhancing the Mechanical Stability Test for Natural Rubber Latex Concentrate. in: Proceedings of the 2nd Asia Pacific Conference on Rubber (APRC), 14

{}^{\rm th}

–16

{}^{\rm th}

, October 2015; Phuket, Thailand, 72-76.

32.

Akmal

Othman

Mansor

. Invention of RRIM MST Tester for Quantitative Method of Mechanical Stability Time (MST) of Natural Rubber Latex Concentrates. Proceedings of the Plastics and Rubber Institute Malaysia (PRIM) Annual Polymer Technology Seminar 2013, 23

{}^{\rm rd}

March 2013.

33.

Fix

Hodges

. Discriminatory analysis, nonparametric discrimination: consistency properties. Tech. Rep. 4; USAF School of Aviation Medicine, Randolph Field, Texas, 1951.

34.

Wang

Neskovic

Cooper

. Improving nearest neighbour rule with a simple adaptive distance measure. Pattern Recognition Letters 2007; 28: 207-213.

35.

Otsu

. A threshold selection method from gray-level histograms. IEEE Trans. Sys., Man., Cyber. 1979; 9(1): 62-66.

36.

Vasan

Surendiran

. Dimensionality reduction using Principal Component Analysis for network intrusion detection. Perspectives in Science, 2016; 8: 510-512.

37.

Yeung

Ruzzo

. Principal component analysis for clustering gene expression data. Bioinformatics, 2001; 17(9): 763-774.

38.

Jolliffe

. Principal Component Analysis. John Wiley & Sons, Ltd., 2014.

39.

Hotelling

. Analysis of a complex of statistical variables into principal components. Journal of Educational Psychology 1933; 24: 417-44.

40.

Brauckhoff

Salamatian

May

. Applying PCA for Traffic Anomaly Detection: Problems and Solutions. in: Proceedings of IEEE INFOCOM, 15–25 April 2009; Rio de Janeiro, Brazil. 2866-2870.

41.

Dalal

Triggs

. Histograms of oriented gradients for human detection. IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), San Diego, CA, USA, 2005; 1: 886-893.

42.

Lowe

. Object recognition from local scale-invariant features. Proceedings of the Seventh IEEE International Conference on Computer Vision, Kerkyra, 1999; 1150-1157.

43.

Slotine

. Fast Wavelet-based Visual Classification. in: Proc. IEEE International Conference on Pattern Recognition (ICPR), Tampa, 2008.

44.

Lai

Phoon

Chew

Goh

Chua

. Image Multi-Thresholding to Characterise the Mechanical Stability of Rubber Latex Concentrate. Proceedings of the International Conference on Digital Image Computing: Techniques and Applications (DICTA 2016); Nov/Dec 2016; Gold Coast, Australia: 372-379.

A novel computer vision approach to classify the mechanical stability of natural rubber latex concentrate

Abstract

Keywords

1. Introduction

1 Flocculation is a process in which particles adhere to form small clumps or masses.

4.1 Preparing NR latex for MST

5. Results and discussions

5.2 Scale invariant feature transforms (SIFT)

5.3 Wavelets

6. Conclusions and future work

Footnotes

Acknowledgments

References

¹
Flocculation is a process in which particles adhere to form small clumps or masses.