Nondestructive Detection of Stem Content in Tobacco Strips Using X-Ray Imaging Analysis

Tobacco leaves (Figure 1a) were harvested and cured in 2013 in Yunnan Province, China. These leaves were threshed with a tobacco threshing machine. The threshed materials were composed of leafy strips, fully stripped stems, and stems with leaf attached (Figure 1b–d). Four types of stems referred to as thick stem (diameter > 2.38 mm), slender stem (diameter < 2.38 mm), long stem (length > 20 mm), and short stem (length < 20 mm) were also prepared as testing materials in the experiment. All tobacco samples contained leafy strips and the four types of stems were adjusted to the moisture content of 17–18% (moisture) before testing. Table 1 shows the apparent density and area density of leaves, thick stems, and slender stems. The area density is defined as the sample mass per unit area, which is equal to the apparent density multiplied by thickness. The density difference, especially the area density difference between leaves and stems, is the main reason that tobacco stems can be identified with X-ray.

Figure 1

An experimental platform was designed on the basis of X-ray detection and analysis method. As demonstrated in Figure 2, this platform includes X-ray generator, imaging detector, feeding hopper, sample conveyor, high-performance computer, discharge hopper, and weight sensor. During testing, the samples containing leaf strips and stems are first fed into the feeding hopper, and then spread on the conveyor in a single layer without overlapping. When samples are transported through the X-ray inspection area, the X-ray with its specific intensity penetrates the samples and the belt, and the samples are then continuously detected by the X-ray imaging detector. The speed of the conveyor belt affects both the dispersion of samples and image quality. When the speed of the conveyor belt is not higher than 0.6 m/s, no obvious influence on the accuracy of the test can be observed due to the maintenance of image quality. At the same time, the 0.6 m/s speed could also lead to the samples being spread in a single layer on the conveyor belt. Considering the above, the speed of the conveyor belt was set to 0.6 m/s in the present system. The gray image was analyzed by high-performance computer to recognize the stems and calculate their mass according to the stem recognition and quantitative algorithm. After that, the samples were conveyed to the discharge hopper. Then, weight sensor at the bottom of discharge hopper could obtain the mass information of samples.

Figure 2

In view of the low density and thickness of plant leaves, the experimental platform used a low energy X-ray generator with a radiation voltage of 30–120 KeV (XRP-75/1000; COMET X-ray Equipment Trading Ltd. Co., Flamatt, Switzerland). Two types of imaging detectors were investigated to assess their applicability for quantitative detection of stems, including a line-scan image sensor (X-DCU; Dandong Along Radiative Instrument Ltd., Co., China) and a TDI (Time Delay Integration) image sensor (C12200; Hamamatsu Photonics Ltd., Co., Japan). The two detectors have the same detecting resolution of 0.4 mm × 0.4 mm per pixel.

The detected gray image of a sample containing leaves and stems is shown in Figure 3. The image processing flowchart of stem recognition and quantization algorithm is given in Figure 4. When leaves and stems are penetrated by X-rays, the gray images obtained by detector show different gray levels in the leaf and stem zones due to differences in density and thickness of the two kinds of tobacco materials. It allows the recognition of stems in the gray image by gray-level threshold and size determination of target area. In Figure 3, the stems showed gray levels < 37,000, whereas the gray level of a tobacco leaf was about 43,000. The gray level of the background was generally > 47,000. According to the differences in gray distribution of leaf and stems (11), the gray threshold used for the stem identification algorithm was selected as 40,000.

Figure 3

Figure 4

In our previous work (5), the details of image preprocessing, segmentation, and stem identification have been given, as seen from step 1 to step 3 in Figure 4. The present work developed the quantitative algorith of stem mass further.

In this quantitative algorithm, the identified stem regions are traversed to calculate the cumulative value of logarithmic function of gray ratio. The coefficient B can be determined by the calibration test of free stem sample in advance. Then the stem mass in tobacco leaves sample can be obtained. The calculation method is given in section 2.4. In addition, the thick and long stems can also be quantified by measuring the diameter and length of stem region, as seen in step 5 of the image processing flowchart.

Low energy X-ray imaging has been widely used for automatic inspection and quality evaluation of agriculture products due to its penetration capabilities (12–13). X-ray has different attenuation after penetrating leaves and stems in terms of different porosity and thickness. The captured X-ray image can show different gray levels in stem and leaf regions. Our previous work has provided the details of X-ray image processing for identifying stems from the mixture of leaves and stems. The following algorithm was used to extract the mass information of identified stems in X-ray image.

The intensity I of monochromatic X-ray at a depth of t in a medium after transmitting is represented as the following (12): [1] I=I0e−μt I = {I_0}{e^{ - \mu t}} where I₀ and μ are the intensity at the top surface of the medium and the attenuation coefficient of the medium, respectively.

The attenuation coefficient of the medium, μ, can be described as follows: [2] μ=σ(Z,E)n \mu = \sigma \left( {Z,E} \right)n where σ and n are the probability of interaction between atoms and photons and the number of atoms per unit volume, respectively. In particular, σ is a function of atomic number, Z, and photon energy, E. [3] n=NρA n = {{N\rho } \over A} where N, ρ, and A are the Avogadro constant, the mass density of medium, and the molar mass of atom, respectively.

According to Equation (2) and Equation (3), Equation (1) can be transformed into: [4] I=I0e−σ(Z,E)NAρt I = {I_0}{e^{ - {{\sigma \left( {Z,E} \right)N} \over A}\rho t}}

Equation (4) can be simplified to the following Equation (5), in which σ, N, and A are assumed to be constant based on the assumption that the substances are consistent in the tobacco stems. [5] I=I0e−Ba I = {I_0}{e^{ - Ba}} where B = σ(Z,E)N/A is a constant; a = ρt is the area density of the medium. Equation (5) can be consequently transformed into: [6] Ba=lnI0I Ba = \ln {{{I_0}} \over I}

Notably, it can be easily found that the degree of X-ray attenuation is directly related to the area density of the tobacco stems according to Equation (6). The stem image is actually composed of n pixels. The area density of the medium corresponded with each pixel is a₁, a₂, a₃, …, a_n, respectively. The data of each pixel are substituted into Equation (6), and the following can be obtained: [7] B∑i=1nai=∑i=1nlnI0Ii B\sum\nolimits_{i = 1}^n {{a_i} = } \sum\nolimits_{i = 1}^n {\ln {{{I_0}} \over {{I_i}}}}

The area of single pixel, S, is a constant for a certain imaging system. It is added to both sides of Equation (7), which can be transformed into the following: [8] ∑i=1nSai=SB∑i=1nlnI0Ii \sum\nolimits_{i = 1}^n {S{a_i} = } {S \over B}\sum\nolimits_{i = 1}^n {\ln {{{I_0}} \over {{I_i}}}}

Because ∑i=1nSai \sum\nolimits_{i = 1}^n {S{a_i}} is equal to the mass of each stem, m, Equation (8) can be converted to the form: [9] m=SB∑i=1nlnI0Ii m = {S \over B}\sum\nolimits_{i = 1}^n {\ln {{{I_0}} \over {{I_i}}}}

In the image of tobacco stems detected by X-ray detector, the various gray values in different regions represent different degrees of attenuation. G₀ is the gray value of the image background. Furthermore, the gray value of each pixel is G₁, G₂, G₃, …, G_n, respectively. Equation (10) can depict the relationship between attenuation and gray value. [10] lnI0Ii=lnG0Gi \ln {{{I_0}} \over {{I_i}}} = \ln {{{G_0}} \over {{G_i}}}

Finally, according to the above series of derivation, the following equation can be obtained: [11] m=SB∑i=1nlnG0Gi m = {S \over B}\sum\nolimits_{i = 1}^n {\ln {{{G_0}} \over {{G_i}}}}

During the real application, the ratio S/B needs to be calibrated in advance. Namely the samples of stems of a certain quality that were fully stripped from tobacco leaves were detected to get their ∑i=1nln G0/Gi \sum\nolimits_{i = 1}^n {\ln \,{G_0}/{G_i}} value in X-ray image, then the S/B was obtained by the ratio of sample mass to ∑i=1nln G0/Gi \sum\nolimits_{i = 1}^n {\ln \,{G_0}/{G_i}} . After that, the S/B could be used for subsequent testing.

Due to the fact that mass information is associated with the thickness and mass density of samples, the thick or long stem content can be derived from their gray level distribution. The gray level distribution is used to estimate the mass information of tobacco stems in X-ray image after calibration. Furthermore, the content of thick (or long) stems in the mixture of tobacco materials or pure tobacco stems can be calculated based on the mass algorithm above.

The thick stem ratio is defined as the content of tobacco stems with a diameter > 2.38 mm in the tobacco materials according to the Chinese Tobacco Industry Standard (4). The tobacco stem with a diameter < 2.38 mm is called a slender stem and is defined by the slender stem ratio. In this work, the thick stem rate was calculated by weighing the thick stems post offline screening, which is inefficient and high-cost processing. Therefore, we established an algorithm to efficiently determine the content of thick or slender stems in tobacco materials based on the X-ray image identification system. The algorithm was set up based on the diameter determination, shape determination and mass determination of tobacco stems.

a)

Diameter determination:

As shown in Figure 5, the stem image was segmented by the circumscribed rectangles, which can obtain the diameter of each stem fragment, d₁, d₂, d₃, …, d_n. The average diameter of the target stem, D, can be estimated according to the below equation: [12] D=(∑i=1ndi)n D = {{\left( {\sum\nolimits_{i = 1}^n {{d_i}} } \right)} \over n}

Figure 5

The performance the of X-ray detector is the key factor influencing the image quality of tobacco stem. In order to improve the accuracy and stability of tobacco stem image, the traditional line-scan detector and the TDI image detector were compared by testing the same stem samples. Figure 6 shows the obtained images by two detectors. The image from TDI detector presented the gray function value ∑i=1nln G0/Gi \sum\nolimits_{i = 1}^n {\ln \,{G_0}/{G_i}} in stem regions with the range of 14,200–14,500 when 5 repeated tests were carried out. In comparison, the gray function value in stem regions was only 6400–6850 by applying the traditional line-scan detector. It is clear that the TDI detector has a higher sensitivity of the quantitative analysis for the same stems. In comparison to the traditional line-scan detector, the TDI detector could synchronize the motion of the object by controlling a group of linear array combinations. In this way, as the image moves from one line to another, the accumulated charge also moves, providing a higher resolution at a weaker X-ray intensity.

Figure 6

The X-ray intensity was further optimized for the detection of tobacco stems in the present experimental system. The contrast of stems and leaves in X-ray images was investigated when X-ray intensity varied from 25 KeV to 100 KeV, as shown in Figure 7. It can be seen that 35 KeV intensity of X-ray resulted in the highest contrast. Considering that a high contrast helps to improve the recognition rate of stems on images, the optimal working condition of the X-ray generator was set as 35 KeV.

Figure 7

In order to evaluate the repeatability of the quantitative detection system, about 5 g fully stripped stems were randomly scattered on the conveyor belt for several times. As shown in Figure 8, the image of stems was obtained by the X-ray imaging unit. The value of ∑i=1nln G0/Gi \sum\nolimits_{i = 1}^n {\ln \,{G_0}/{G_i}} for each image in Figure 5 is shown in Table 2. After 6 times of experiments, the average values of gray function were at 16,616. The relative deviation of each image ranged from −0.17% to 0.16%. Due to the linear dependence of stem mass on the gray function value, the results shown in Table 1 indicated that the method and the detection system are reliable in the sense of repeatability for quantitative detection of the sample stems.

By blending different amounts of free stems into isolated leaves, the samples with different levels of stem contents (i.e., from 1.0% to 6.0%) were prepared to evaluate the accuracy of quantitative detection method (Table 3), according to its actual range during tobacco primary processing. At each level, 500–1000 g of the blended samples were tested to obtain the experimental value of stem contents. The experimental results by the machine vision system showed high consistency with the actual contents of tobacco stems. The relative errors ranged only from −3.64% to 2.76%.

In the tobacco primary processing, a traditional ISO method (3) is applied to determine the stem content in leaves. This detection method mainly includes the sample breakage, air classification, and stem weighting. The aim of sample breakage in the ISO method is to free all stems that are still part of leaves. Then the stems are separated from the samples by air classification and weighed to obtain the stem content of the sample.

The quantitative results of stem content obtained from the traditional ISO method and the present X-ray image analysis method were compared with the same test material. According to the sample amount required by the ISO method, 3 kg stripped leaves were blended with 30 to 120 g of stems to prepare 4 samples with different stem contents. Each sample was firstly analyzed by the presented method, then recollected and tested by the ISO method. Table 4 shows the quantitative results of the two methods. The relative error of ISO method is in a range of 3.32–9.79% with an average relative error for the 4 levels 7.08%.

However, the relative error of the presented method is in a range of 0.95–2.07% with an average relative error of only 1.59%.

For some leafy agriculture products, more attention needs to be paid to the content of larger stems in leaves due to their more significant negative effects on the quality of products. For example, the detection of thick stems (diameter > 2.38 mm) is necessary for the quality evaluation of tobacco leaves.

Currently, the above detection methods mainly depend on manual picking. It means the tester has to manually select the larger stems out from certain amounts of leaf product via a vernier caliper, and then determine the content of the thick stems as well as long stems. In the presented work, the detection algorithm of thick stems was established by X-ray imaging. For example, 100 free stems containing 50 thick stems were used as a test sample to evaluate the detection performance for the identification system. The results are summarized in Table 5. The number of mis identifications for target thick stem was no more than 4. The identification accuracy was 94.67%. These results demonstrate that the developed system is valid when identifying thick stems.

Similarly, a test sample containing 100 stems containing 50 long stems was also used to evaluate the experimental system. It was found that at least 49 of total 50 long stems were identified, as shown in Table 6. The identification accuracy was higher than 98.00%. The results indicate that the experimental system and method can significantly improve the detection accuracy and efficiency for long stems.