Study on the Selection of Processing Process and Parameters of Platycodon grandiflorum Seeds Assisted by Machine Vision Technology

Abstract

Seed processing is an important means of improving seed quality. However, the traditional seed processing process and parameter adjustment are highly empirically dependent. In this study, machine vision technology was used to develop a seed processing method based on the rapid extraction of seeds’ material characteristics. Combined with the results of clarity analysis and the single seed germination test, the seed processing process and parameters were determined through data analysis. The results showed that several phenotypic features were significantly or highly significantly correlated with clarity, but fewer phenotypic features were correlated with viability. According to the probability density distribution of pure seeds and impurities in the features that were significantly correlated with seed clarity, the sorting parameters of length, width, R, G, and B were determined. When the combination of width (≥0.8 mm) + G (<75) was used for sorting, the recall of pure seeds was higher than 91%, and the precision was increased to 98.6%. Combined with the specific production reality, the preliminary determination of the Platycodon grandiflorum seed processing process was air separation—screen (round hole sieve)—color sorting. Then, four commercialized Platycodon grandiflorum seed lots were sorted by this process using corresponding parameters in the actual processing equipment. Subsequently, the seed clarity and germination percentage were significantly improved, and the seed quality qualification rate was increased from 25% to 75%. In summary, by using machine vision technology to quickly extract the material characteristics of the seeds, combined with correlation analysis, probability density distribution plots, single feature selection, and combination sorting comparisons, the appropriate processing process and corresponding sorting parameters for a specific seed lot can be determined, thus maximizing the seed quality.

1. Introduction

Platycodon grandiflorum (Jacq.) A. DC. is a perennial herb of Campanulaceae, and the root is used as a medicine to promote lung expectoration, drain pus and reduce swelling [1,2]. To meet the market demand, China has started the large-scale standardized artificial cultivation of Platycodon grandiflorum. However, most of the Platycodon grandiflorum seeds sold in the market are directly harvested from the field, and the quality of seed lots varies [3]. Seed clarity, seedling emergence rate, and seedling neatness of some seed lots are low. Platycodon grandiflorum seed quality has two existing local standards. In the local standards of the Inner Mongolia Autonomous Region (DB/T 15.1297-2017) [4], according to clarity, germination percentage, seed utilization value (clarity × germination percentage × 100), and the number of other plant seeds, Platycodon grandiflorum seeds are divided into three levels, including primary seed germination percentage ≥95%, secondary ≥85%, tertiary ≥75%, primary seed clarity ≥98%, secondary ≥96%, and tertiary ≥94%. The local standard of Hebei Province (DB/T 13.1320.3-2010) [5] stipulates that the germination percentage of Platycodon grandiflorum seeds is ≥75% and the clarity is ≥95%.

Seed processing is an important means to improve the clarity, purity, and vitality of seeds after harvesting, and the process is mainly based on the physical characteristics of seeds to eliminate impurities, heterogeneous seeds, and unsaturated, insect-infested, or deteriorated seeds [6]. Compared with crop seeds, the production and processing of Chinese herbal medicine seeds are still at a relatively crude stage [7]. Seed processing procedures generally include air separation, air screen, pocket cylinder, specific gravity separation, and color sorting. Seeds of different crops, different varieties of the same crop, different origins of the same variety, and different years of production have different characteristics and contain different impurities, so there are differences in the appropriate processing techniques and parameters. It is clearly stated in textbooks that the selection of seed processing process should take into account the type of seed, the nature and type of impurities, and the final required seed quality indicators [8]. However, there has been a lack of relevant operating procedures on how to operate. The production relies on the experience of the operator or the pre-treatment of the small electric sieve equipment several times to determine the processing process, and the small electric sieve is limited to pre-sorting based on width and thickness. Due to the lack of effective data support, the operator requirements are high. The reason for this dilemma is that the acquiring of material characteristics is cumbersome. Bao et al. (2022) and Darfour et al. (2022) still used vernier calipers to measure the length, width, and thickness of seeds when measuring the material characteristics of Isatis indigotica seeds and maize grain [9,10].

Machine vision technology, which has developed rapidly in recent years [11,12], combines computer technology and image processing technology, with the characteristics of high computational power, low price, non-destructive, and high efficiency. There are several seed testers in the market, and they are mainly used for the rapid determination of population morphological features of breeding materials in the field. The Seed Science and Technology Research Center of China Agricultural University has launched Seed Identification, PhenoSeed, AIseed, and other testing systems for mass and rapid extraction of the phenotypic features of a single seed since 2012 [13,14,15,16]. In recent years, many scholars have used machine vision technology combined with different algorithms to detect seed clarity, authenticity, and vigor [14,15,16,17]. The color sorter is the most successful case of machine vision technology for seed sorting and has been increasingly used in the last decade for seed selection [18,19,20]. The common algorithms used in color sorters consist of two types of algorithms: grayscale and color difference. The grayscale algorithm differentiates based on R, G, and B. The color difference algorithm compares the size of R/(R-G), R/(R-B), or G/(G-B), which is equivalent to comparing the size of R/G, R/B, or G/B. Existing color sorters are equipped with intelligent modeling functions that can precisely adjust the relevant parameters to improve the color sorting results.

Because of the low quality of Platycodon grandiflorum seeds on the market and the incomplete processing process, this study aimed to explore the appropriate processing process and parameters assisted by machine vision technology. Specifically, PhenoSeed software was used to rapidly extract the phenotypic features (material characteristics) of the seeds to be sorted. Combined with the results of clarity analysis and single seed germination test, the appropriate processing process and related parameters of Platycodon grandiflorum seeds were determined through correlation analysis, single feature selection, and combination sorting comparisons. In addition, another three Platycodon grandiflorum seed lots were adopted to verify the sorting effect of our method.

2. Materials and Methods

2.1. Seed Materials and Equipment

Four commercialized Platycodon grandiflorum (Jacq.) A. DC. seed lots were purchased from the market (

Table 1. Seed list.

Name	Place of Origin	Year of Production	Clarity/%	Germination Percentage/%
Lot 1	Hebei Anguo	August 2020	96.5	85.7
Lot 2	Zhangjiakou	August 2021	95.2	74.0
Lot 3	Shandong	November 2020	96.2	44.3
Lot 4	Anhui	October 2020	95.1	65.7

and Table S1), each 1 kg. A CCD scanner (Microtek MiCardWizard, Shanghai Zhongjing Technology Co., Ltd., Shanghai, China) was used to obtain images of seeds and impurities. An automated seed analysis system (PhenoSeed, developed by Seed Science and Technology Research Center of China Agricultural University and Nanjing Zhinong Yunxin Big Data Technology Co., Ltd., Nanjing, China) segmented the seeds and impurities from the background and obtained the phenotypic features (e.g., shape, texture, and color) of the samples. The air separator (Beijing Xida Agricultural Engineering Science and Technology Development Center, Beijing, China) separated seeds from light impurities and dust by adjusting the wind speed. The screen machine (Huachang Grain and Oil Machinery Co., Ltd., Taizhou, Zhejiang) separated seeds from impurities of different sizes by selecting different pore sizes of the sieve plate. A 5FXS-2 air screen machine (Beijing Xida Agricultural Engineering Science and Technology Development Center) is equipment that combines air separation and a screen. A SATAKE FMS-2000 color sorter (Satake manufacturing (Suzhou) Co., Ltd., Suzhou, China) separated the heterochromatic seeds detected by the optoelectronic system through a jet.

2.2. Method

2.2.1. Image Scanning and Extraction of Physical Features of Seeds and Impurities

A random selection of 300 Lot 1 pure seeds and 300 impurities, each with a certain gap between the seeds and impurities, were scanned using a scanner. The images were saved in TIF lossless format with a resolution of 300 dpi. The software can extract color features including R (Red), G (Green), B (Blue), H (Hue), S (Saturation), V (Value), L (Luminosity), a (range from magenta to green), b (range from yellow to blue), Gray, R/G, R/B, and G/B and size features, including length (mm), width (mm), L/W (length-to-width ratio), area (projected area, mm²), perimeter (mm), and roundness.

2.2.2. Single Seed Germination Test

About 900 pure seeds of Platycodon grandiflorum (Lot 1) were randomly selected and scanned in the same way as in Section 2.2.1. Afterward, the seeds were placed in Petri dishes with two layers of germination paper according to serial numbers for the germination test, which was performed according to GB/T 2930.4 [21]. The germination result of each seed was counted on the 14th day, using 1/2 of the length of the radicle extending from the seed as the germination standard. To ensure a balanced number of live and dead seeds samples, all 193 dead seeds and 207 randomly selected live seeds, totaling 400 seeds, were used for the study.

2.2.3. Material Characteristics and Correlation Analysis

IBM SPSS Statistics 21.0 software was used to analyze the correlation between seeds’ material characteristics, clarity and viability, to select physical features that were significantly correlated and had high correlation coefficients. Probability density distribution plots were drawn to select different parameters. The sorting effects of different parameters were compared by precision (P), recall (R), and F1. In the analysis of clarity using machine vision data, the ratio of the number of seeds to the total number of seeds and impurities was used as an indicator for easy calculation and comparison, called quantity clarity. When analyzing clarity, the precision was the quantity clarity of the selected seeds (number of seeds in the post-selection seed lot divided by total number of seeds and impurities in the post-selection seed lot), and the recall was the seed recall rate (number of seeds in the post-selection seed lot divided by number of seeds in the pre-selection seed lot). When analyzing viability, the precision was the germination percentage of the selected seeds (number of live seeds in the post-selection seed lot divided by number of seeds in the post-selection seed lot), and the recall was the live seed recall rate (number of live seeds in the post-selection seed lot divided by number of live seeds in the pre-selection seed lot). F1 is the harmonic average of precision and recall, which is a comprehensive evaluation indicator:

F1 = P × R × 2/(P + R) × 100%

(1)

2.2.4. Real Equipment Sorting Verification

The original Lot 1 was processed using the experimentally determined processing process and parameters. The clarity (weight clarity), thousand seed weight, and germination percentage of the seeds before and after sorting were measured.

2.2.5. Data Analysis

The data were summarized by Microsoft Excel 2016, followed by the analysis of variance (ANOVA) using IBM SPSS Statistics 21.0 for clarity, thousand seed weight, and germination percentage. The specific experimental procedure is shown in Figure 1.

3. Results

3.1. Correlation Analysis of Seeds’ Material Characteristics with Clarity and Viability

As shown in

Table 2. Correlation analysis of phenotypic features with clarity and viability of Lot 1 Platycodon grandiflorum seeds.

Phenotypic Features	Quantity Clarity		Viability
	Correlation Coefficient	Coefficient of Variation	Correlation Coefficient	Coefficient of Variation
Length/mm	0.636 **	0.30	/	/
Width/mm	0.236 **	0.19	0.094 **	0.13
L/W	0.323 **	0.37	−0.075 *	0.16
Area/mm2	0.762 **	0.39	0.084 *	0.20
Perimeter/mm	0.515 **	0.23	/	/
Roundness	0.201 **	0.19	0.104 **	0.12
R	−0.782 **	0.29	/	/
G	−0.851 **	0.33	/	/
B	−0.792 **	0.31	/	/
R/G	0.594 **	0.09	/	/
R/B	0.196 **	0.12	/	/
G/B	−0.377 **	0.09	/	/
L	−0.846 **	0.31	/	/
a	0.439 **	0.52	/	/
b	−0.645 **	0.35	/	/
H	−0.285 **	0.22	/	/
S	0.360 **	0.23	/	/
V	−0.811 **	0.29	/	/
Gray	−0.836 **	0.31	/	/

Note: * and ** represent significant (p < 0.05) or highly significant (p < 0.01) correlations, respectively.

, 19 phenotypic features were significantly or highly significantly correlated with quantity clarity, and 7 features (Area, R, G, B, L, V, and Gray) had correlation coefficients above 0.7. Only four phenotypic features (width, L/W, area, and roundness) were significantly or highly significantly correlated with viability, but the correlation coefficients were low, only at 0.1 or so.

3.2. Single Feature Sorting Effect

The sorting process was simulated based on several features of length (pocket cylinder), width (round hole sieve), R (color sorter), G (color sorter), B (color sorter), R/G, R/B, and G/B, taking into account the currently available seed sorting equipment and its sorting principle. According to the probability density distribution plots in Figure 2, it can be seen that there was a clear distinguishable threshold between pure seeds and impurities. In contrast, the probability density distribution plots of live and dead seeds overlapped extensively, presumably with poor differentiation. Therefore, the processing objective of this Platycodon grandiflorum seed lot was mainly to improve the clarity.

Table 3. Sorting effects of different phenotypic features.

Sorting Feature	Parameters	Quantity Clarity			Viability
		Precision/%	Recall/%	F1	Precision/%	Recall/%	F1
Length/mm	<1.4	0.0	0.0	NA	100.0	0.5	0.010
	1.4–1.8	48.6	17.0	0.252	50.0	3.4	0.063
	≥1.8	83.6	83.0	0.833	51.7	96.1	0.672
Width/mm	<0.6	0.0	0.0	NA	100.0	0.5	0.010
	0.6–0.8	17.9	5.0	0.078	25.0	1.4	0.027
	≥0.8	59.0	95.0	0.728	52.5	98.1	0.684
R	<85	90.6	92.7	0.916	51.5	93.7	0.664
	85–100	25.3	7.0	0.110	58.8	4.8	0.089
	≥100	0.5	0.3	0.004	50.0	1.4	0.028
G	<75	95.0	95.3	0.952	46.2	84.5	0.597
	75–90	28.0	4.7	0.080	62.5	4.8	0.090
	≥90	0.0	0.0	NA	40.0	1.0	0.019
B	<60	92.2	91.0	0.916	52.1	79.7	0.630
	60–75	26.0	9.0	0.134	50.7	17.9	0.264
	≥75	0.0	0.0	NA	50.0	2.4	0.046

also verifies this inference. Although the viability precision (germination percentage) reached 100% for length < 1.4 mm or width < 0.6 mm, the recall was only 0.5%, which was essentially negligible. The distribution of impurities and pure seeds on these features possessed obvious distinction, and the distinction on R, G, and B was better than that on R/G, R/B, and G/B, so the grayscale algorithm should be selected for color sorting. The distribution plots of length, width, R, G, and B of impurities and pure seeds were used to initially determine the sorting parameters to be used for further comparison of each feature, namely length: 1.4 mm, 1.8 mm, width: 0.6 mm, 0.8 mm, R: 85, 100, G: 75, 90, and B: 60, 75. As can be seen from Table 3, when length ≥ 1.8 mm, width ≥ 0.8 mm, R < 85, G < 75, or B < 60, the quantity clarity recall was above 80.0%, and the quantity clarity F1 was higher than 0.7; among the three features of R, G, and B, the quantity clarity F1 was highest when G < 75.

3.3. Combination Sorting Effect Comparison and Preliminary Determination of the Processing Process

As can be seen from

Table 4. Sorting effects of different combinations.

Sorting Combinations and Parameters	Quantity Clarity
	Precision/%	Recall/%	F1
Width (≥0.8 mm) + G (<75)	98.6	91.0	0.946
Length (≥1.8 mm) + G (<75)	98.3	79.0	0.876
Width (≥0.8 mm) + length (≥1.8 mm)	90.2	80.0	0.848
Width (≥0.8 mm) + length (≥1.8 mm) + G (<75)	99.1	76.3	0.863

, the combination of width (≥0.8 mm) + G (<75) has the best sorting effect, with a precision of 98.6%, recall of 91.0%, and F1 of 0.946. Machine vision technology cannot detect the floating speed and seed-specific gravity, but air separation and specific gravity separation are two important parts of the seed processing process, and air separation has to some extent the function of specific gravity separation. Therefore, the preliminary determinations of Lot 1 Platycodon grandiflorum seed processing process were (1) air separation, to remove dust and light impurities in the seeds; (2) screen, using a round hole sieve (sieve diameter of 0.8 mm) to remove small impurities and small seeds (air screen machine can merge step (1) and step (2)); (3) color sorting, with the seeds with G < 75 as the conformance product, and impurities or seeds with G >75 as the nonconformance product for intelligent modeling and sorting.

3.4. Real Equipment Sorting Verification

For Lot 1 Platycodon grandiflorum seeds, air separation—screen (round hole sieve ≥ 0.8 mm)—color sorting (G < 75) was used for sorting. Color sorting was added to the secondary sorting, i.e., the unselected material after the primary color sorting was subjected to another color sorting.

According to Platycodon grandiflorum seed-quality grading (DB/T 15.1297-2017) [4], the seeds were divided into three grades based on the germination percentage, clarity, and the number of grains of other plant seeds. As can be seen from

Table 5. Comparison of seed quality of Lot 1 Platycodon grandiflorum seeds after air separation, round hole sieve, and color sorting.

Processing	Clarity/%	Thousand Seed Weight/g	Germination Percentage/%	Recall (Weight)/%
Original seed	96.5 ± 0.4 d	1.13 ± 0.06 a	85.7 ± 2.1 cd	/
Air separation	97.6 ± 0.5 c	1.13 ± 0.03 a	88.0 ± 2.6 bcd	98.2
Air separation—round hole sieve	98.1 ± 0.1 bc	1.07 ± 0.10 ab	91.7 ± 0.6 ab	97.2
Air separation—round hole sieve—primary color sorting	99.1 ± 0.1 a	1.11 ± 0.04 a	94.3 ± 2.1 a	40.2
Air separation—round hole sieve—secondary color sorting	98.6 ± 0.5 ab	1.06 ± 0.03 ab	89.3 ± 3.5 abc	26.4
Air separation—round hole sieve—color sorting nonconformance product	98.1 ± 0.2 bc	0.98 ± 0.03 b	83.0 ± 4.4 d	/

Note: Different lowercase letters in the same column indicate the presence of significant differences between treatments (p < 0.05).

, Lot 1 seeds had reached 98% of clarity (weight clarity) after air separation—screen (round hole sieve ≥ 0.8 mm), which had met the first grade of clarity in the local standard of Inner Mongolia. The germination percentage had also been significantly improved, reaching 91.7%. If there was a higher demand for clarity and germination percentage, color sorting could be carried out. The clarity of primary color sorting (G < 75) reached 99.1%, and the germination percentage reached 94.3%; the clarity of secondary color sorting was 98.6%, and the germination percentage was 89.3%.

Material characteristics of the other three Platycodon grandiflorum seed lots revealed that the process was the same, but the round hole sieve apertures needed to be adjusted. According to their corresponding probability density distribution plots, 1.0 mm and 1.5 mm round hole sieves were selected, and seeds between 1.0 mm and 1.5 mm were taken for color sorting again (G < 75).

The germination percentage of all three seeds was also improved, with an average increase of 7.2%. However, due to the low original germination percentage of the seeds themselves, the germination percentage of only two seed lots increased to more than 75%, meeting the requirements of the local standard of Hebei Province for the germination percentage and the local standard of Inner Mongolia for the germination percentage of third-grade seeds.

4. Discussion

The four seed lots used in this study were all commercialized seeds purchased from the market. Although the clarity met the minimum local standard requirement of 95%, only one seed lot had a germination percentage that met the requirement. The quality pass rate of these four seed lots was only 25%, indicating that the quality of Platycodon grandiflorum seeds in the market is uneven and needs urgent improvement. Seed processing is an important link to enhance the quality of seeds. Seed companies are generally built with complete sets of seed processing lines, and the conventional configuration includes a pre-cleaning machine, air screen machine, pocket cylinder, and specific gravity separator. In recent years, the new line has generally increased the color sorter. The seed processing process is based on the characteristics of the seeds and their impurities. It requires the use of the smallest amount of machinery and equipment and the simplest process to achieve the most desirable processing results [8]. The traditional use of experience or a small electric sieve to determine the processing process (including parameters) obviously cannot meet this requirement.

In this study, machine vision technology was applied to the rapid analysis of material characteristics to determine the processing process of Platycodon grandiflorum seeds, which was the same for different seed lots, but the specific parameters should be adjusted according to the quantitative analysis of seeds’ material characteristics.

The processing process identified in this study can significantly improve seed clarity and achieve 98% of the standard requirement for first-grade seeds, but the improvement of seed germination percentage is greatly limited by the original germination percentage of the seeds. The germination percentage of the seeds used in this study ranged from 53.7% to 85.7%, and only one seed lot was qualified. After sorting using the processing process determined in this study, Lot 1 seeds were upgraded from second-grade seeds to first-grade seeds, and two of the other three unqualified seeds were upgraded to qualified commercial seeds.

Specific gravity separation is an important part of seed vigor enhancement. Rabindra found that compared with air separation, specific gravity separation better improved the germination percentage and vigor of wheat seeds [22]. In our previous study, Platycodon grandiflorum seeds were divided into different groups according to the specific gravity by liquid sorting, and the germination percentage of seeds with specific gravity ≥ 0.95 g/mL was significantly higher than that of seeds with specific gravity < 0.95 g/mL [23]. However, since machine vision technology cannot detect the specific gravity of seeds, this study only addresses the processing aspects that can be improved. In addition, the specific gravity of the selected Platycodon grandiflorum seeds after combination sorting was ≥0.95 g/mL. This shows that for the seeds with low specific gravity (such as medicinal seeds and vegetable seeds), wind separation plays a role as specific gravity separation while removing impurities. Ji obtained the same result in an experiment on pepper seed air separation [24].

When PhenoSeed extracted the phenotypic features based on the scanned seeds images, it was for the two-dimensional flat images and could not detect the thickness of the seeds, which is a technical point that needs to be improved or solved in this kind of research.

5. Conclusions

Machine vision technology can be used to assist in the selection of seed processing process and parameters by quickly extracting seeds’ material characteristics. This study first used machine vision technology to quickly extract the machine vision features of pure seeds/impurities and live/dead seeds in a seed lot, then selected significantly correlated features by correlation analysis, determined the sorting parameters of each feature by probability density distribution plots, compared the sorting effect of a single feature, and determined the most suitable process for different seed lots by multiple combinations of sorting. The laboratory has almost completed the development of AIseed Simulation software (to be released) to facilitate users to simulate sorting on personal computers to determine the appropriate seed processing process and sorting parameters.

Four Platycodon grandiflorum seed lots were sorted using the precisely selected processing process and parameters. This study significantly improved the clarity and germination percentage and increased the seeds qualification rate from 25% to 75%.