Characterization and Detection Classification of Moldy Corn Kernels Based on X-CT and Deep Learning

Abstract

Moldy corn produces aflatoxin and gibberellin, which can have adverse effects on human health if consumed. Mold is a significant factor that affects the safe storage of corn. If not detected and controlled in a timely manner, it will result in substantial food losses. Understanding the infection patterns of mold on corn kernels and the changing characteristics of the internal structure of corn kernels after infection is crucial for guiding innovation and optimizing detection methods for moldy corn. This knowledge also helps maintain corn storage and ensure food safety. This study was based on X-ray tomography technology to non-destructively detect changes in the structural characteristics of moldy corn kernels. It used image processing technology and model reconstruction algorithms to obtain the 3D model of the embryo, pores and cracks, endosperm and seed coat, and kernels of moldy corn kernels; qualitative analysis of the characteristic changes of two-dimensional slice grayscale images and 3D models of moldy corn kernels; and quantitative analysis of changes in the volume parameters of corn kernels, embryos, endosperm, and seed coats as a whole. It explored the detection method of moldy corn kernels based on a combination of X-ray tomography technology and deep learning algorithms. The analysis concluded that mold infection in maize begins in the embryo and gradually spreads and that mold damage to the tissue structure of maize kernels is irregular in nature. The overall volume parameter changes of corn kernels, embryos, endosperm, and seed coats in the four stages of 0 d, 5 d, 10 d, and 15 d showed a trend of first increasing and then decreasing. The ResNet50 model was enhanced for detecting mold on maize kernels, achieving an accuracy of over 93% in identifying mold features in sliced images of maize kernels. This advancement enabled the non-destructive detection and classification of the degree of mold in maize kernel samples. This article studies the characterization of the characteristic changes of moldy corn kernels and the detection of mildew, which will provide certain help for optimizing the monitoring of corn kernel mildew and the development of rapid detection equipment.

1. Introduction

Corn is one of the primary grain crops in China. The large embryonic part, rich in nutrients, with a high bacterial load and strong respiration, makes it highly susceptible to mold infection. Mold infections significantly contribute to the deterioration of grain quality in both farmers’ storage facilities and state grain storage warehouses. The consumption of moldy maize feed by livestock can lead to mold infections in animals, with aflatoxin B1 being a significant factor in causing liver cancer in humans. Every year, China experiences a grain loss of about 70 billion catties, with farmers’ grain storage accounting for 8% of this loss. Mildew is identified as the primary cause of this loss. Therefore, it is crucial to understand the process of maize mold infection, investigate the changes in the internal characteristics of maize kernels after molding, and accurately detect maize molds to ensure the safety of grain storage and food.

The non-destructive detection of changes in the internal characteristics of moldy maize kernels is challenging with conventional mold detection techniques. The detection of moldy maize can be achieved through sensory inspection, but this approach is ineffective and can only detect visible mold spots on the surface. Colony counting, enzyme-linked immunoassay [1], and thin-layer chromatography [2] involve laborious and expensive detection methods. Near-infrared spectroscopy [3] and terahertz [4] are non-destructive testing methods but have limited detection sensitivity, complex modeling challenges, and high technical requirements for test preparation. Electronic temperature measurement [5] and CO₂ detection [6,7] show hysteresis in mold detection. Currently, conventional image processing methods are utilized to detect the internal morphological characteristics of maize kernel samples. However, this approach requires the chemical slicing of maize, which is inefficient and environmentally polluting. Furthermore, it cannot provide isotropic information about the three-dimensional structure of a whole maize kernel [8]. In order to intuitively monitor the changes in the morphology of the embryo, endosperm, and grain at each stage of corn kernel development and quickly detect moldy corn kernels in the early stages, a non-destructive testing method that preserves the morphology of corn kernels is needed.

X-ray tomography is a new imaging technique in which X-rays are used to scan a sample in a multi-directional spatial stereoscopic manner, and the image data obtained from the scan are reconstructed into a three-dimensional model. This technique is characterized by non-contact, non-destructive scanning and high resolution, and is widely used for the non-destructive testing of materials and the structural investigation of microscopic objects. Ying Zhou et al. [9] used this technique to reconstruct the three-dimensional characteristics of wheat grain after germination in three dimensions. The researchers calibrated the volume parameters of the embryo and endosperm of wheat grain, examined the volume changes of the embryo and endosperm before and after germination, and identified that the endosperm gradually decreases while the embryo grows after wheat germination. Kang Zhu et al. [10] scanned wheat grains and used this technique to reconstruct a three-dimensional model and export the point cloud data. They integrated it with Geomagic Design for post-processing point cloud data and created a discrete element model, which was subsequently validated through experimental verification. This technology can be used for scanning wheat grain images and in processing methods for discrete element simulation. Qingping Wang et al. [11] used X-ray tomography technology to evaluate the effects of moisture–heat coupling aging on recombinant bamboo. The study led to changes in the CT value of the engineered bamboo, demonstrating the effective structural strength characteristics of the engineered bamboo and the loss of moisture–heat coupling aging. Koksel et al. [12] used synchrotron-source X-rays to analyze bubbles in non-yeast dough. Neethirajan et al. [13] used X-ray tomography CT in both horizontal and vertical directions to study wheat, peas, and other grains and found differences in the size and number of air channels between the two directions. Although traditional image processing methods and model reconstruction can be used to observe changes in the structural characteristics of wheat grains, the processing process is cumbersome, and the defects of wheat cannot be quickly identified after CT scanning.

Deep learning techniques are widely used for detecting pests and diseases in agriculture and fisheries. Jian Wang et al. [14] used the improved ResNet34 model to identify and classify images of rice pests. Fang Zhang et al. [15] used the ResNet50 model to identify sick redfin oriental triggerfish through a transfer learning approach, effectively addressing the challenges of limited samples and low detection accuracy. Baoxia Sun et al. [16] used deep learning visual recognition technology to identify ripe citrus for yield estimation and long-term defect detection. Weina Wang et al. [17] employed deep learning to develop a Hypernet PRMF model for detecting moldy peanuts, achieving a detection accuracy of 90.35%. Barboza et al. [18] improved the detection of maize weevils by utilizing X-ray imaging technology in conjunction with deep learning algorithms. Mallick et al. [19] proposed a deep learning-based method for the rapid detection and identification of pests in mung beans. Choi et al. [20] introduced a method for constructing a deep learning partitioned image dataset for pest detection in strawberries, resulting in an improved accuracy of 91.93% in detecting pests in strawberries.

This paper presents a study on the internal structural changes of moldy maize kernels and methods for detecting mold using X-ray tomography. Two-dimensional cross-sectional images of kernels were obtained non-invasively. Then, 3D models of the corn kernel, embryo, endosperm, seed coat, pores, and cracks were reconstructed. Furthermore, a ResNet50 model was developed to detect corn mold by integrating it with deep learning algorithms. Qualitative and quantitative analyses were conducted to examine the structural change characteristics of moldy maize kernels. This study used the ResNet50 model to detect maize mold and classify the severity of the mold.

2. Materials and Methods

2.1. Preparation of Materials

The experiment simulated the natural heating and hygroscopic mildew of corn in the granary, and the moldy corn grain samples were prepared using the corn’s own fungi. When a corn kernel reaches a temperature of 25 °C or higher, corn with a water content exceeding the safe moisture level of 13% becomes highly susceptible to mildew. Consequently, the corn kernels were screened and purified. The initial moisture content was reduced to 12% using a drying box. Sterile water was dispensed into a test tube using a pipette to adjust the moisture level of each corn grain to 20%. Subsequently, the corn was placed in a room with a constant temperature of 25 °C after standing for 48 h.

2.2. Experimental Methodology

Corn kernel samples were scanned using CT for each sample at incubation times of 0 d, 5 d, 10 d, and 15 d. The sliced images were processed and a 3D model was reconstructed to develop a corn kernel mold detection model based on a deep learning algorithm, which also classified the degree of corn mold.

2.2.1. Image Data Acquisition

The acquisition device was a SKYSCAN1275 X-ray tomography scanner, which scanned the object at every angle of rotation and automatically collected a set of data by rotating 360° for each scanning cycle. The process for scanning the corn sample was as follows: the sample was secured in place on the rotating column using white foam and tape and then placed into the CT scanning chamber on the rotating disc. The hatch was closed and the scanning parameters were set according to the specifications outlined in

Table 1. CT scan parameters.

CT (SKYSCAN1275)	Voltage	Amps	Resolution (of a Photo)
Parametric	50 kV	65 µA	11 µm

.

2.2.2. Image Processing and Model Reconstruction

The maize kernel image data obtained through scanning needed to undergo the sequential processes of image data preprocessing, image segmentation, and model reconstruction to obtain 2D grayscale slice images for each maize tissue region, as well as 3D models.

Image data preprocessing was used to convert the obtained CT scan data into 2D slice images. To clearly display the sliced image of moldy maize, the smoothing function was disabled, and the ring artifact value was set to 70%. As shown in Figure 1a–c.

Image segmentation was a crucial step in the 3D visualization of the maize kernels as it involved separating each tissue of the kernel from the background region. Due to the indistinct boundary contours of each tissue in the corn kernel slices, we employed an interactive threshold segmentation algorithm [20] to segment the kernel image layer by layer. In this study, the corn kernel, embryo, endosperm, and seed coat were segmented using a thresholding technique. As the amount of mold on the corn seed coat and endosperm increased, segmentation became more challenging, and only the seed coat and endosperm could be segmented together as a whole. The segmentation included pores and cracks in these four areas. As shown in Figure 1d,e.

Model reconstruction was accomplished by assigning distinct colors to each tissue of maize, which was segmented through interactive thresholding and rendering [21]. This process separated the kernel, endosperm and seed coat, embryo, pores, and cracks from the kernel one by one. As shown in Figure 1f,g.

2.3. Methods for Analyzing Structural Changes in Moldy Maize Kernels

In this study, we analyzed the changes in grayscale values and contour features of 2D greyscale images of maize kernels from 0 d to 15 d by qualitative analysis methods. Additionally, we analyzed the changes in geometrical and morphological features of maize kernels by analyzing 3D models of maize kernels, embryos, endosperm, seed coats, pores, and cracks from 0 d to 15 d.

A quantitative method was used to analyze the changing pattern of the volumetric parameters of the maize kernels, embryos, endosperm, and testa with the deepening of mycorrhizal infestation from 0 d to 15 d. The volumetric parameters were derived from the cumulative number of voxels occupied by the reconstructed model. It is known that the volume of each tissue structure of the maize kernel can be calculated by Equation (1).

$$\mathrm{V}=\mathrm{t}\mathrm{n}$$

(1)

where t is the voxel volume and n is the number of voxels.

2.4. Classification Methods for Detecting Moldy Maize Kernels

2.4.1. Dataset Creation

In this study, the ResNet50 model was fine-tuned for detecting corn mildew, and 2191 grey-scale images were divided into two classes. Images showing mold spores, low gray values of the embryo, or defects were categorized as images with mildew characteristics. Conversely, images without mildew characteristics were also identified. Among these, 1084 images exhibited no mildew characteristics, while 1107 images showed mildew characteristics, as illustrated in Figure 2. The dataset was subdivided into a training set (80% of the images) and a validation set (20% of the images); the training set consisted of 867 grayscale images without mildew features and 886 images with mold features and the validation set consisted of 217 grayscale images without mildew features and 221 images with mildew features.

2.4.2. Image Preprocessing

In order to mitigate the order of magnitude difference between the dimensional data and prevent significant errors in the prediction model resulting from the substantial difference in the order of magnitude between the input and output data, grayscale images were normalized before training and the dimensions were adjusted to 224 × 224 pixels. In order to prevent overfitting of the deep learning model caused by the limited dataset size, the existing corn kernel dataset was augmented to enhance training diversity. This was achieved through data augmentation techniques such as image rotation, flipping, and translation. The preprocessing flowchart is illustrated in Figure 3. Image rotation: the image was randomly rotated between 0 and 360 degrees. Image flip: the image was randomly flipped around the X-axis and the Y-axis. Image panning: the image was panned along the X-axis and the Y-axis, with a pixel interval of the panning interval (−20–20). Figure 3 shows the progression of an original corn kernel slice image through normalization, rotation, flipping, and panning.

2.4.3. ResNet50

In this study, a ResNet50 model was developed for the rapid detection of maize mildew, and Figure 4 illustrates the fine-tuning process of the model. Optimization was performed using an optimizer with stochastic gradient descent with momentum (SGDM) with an initial learning rate of 0.01. The learning rate was reduced by a factor of 0.2 every five cycles, with a minimum batch size of 128. A maximum of 30 theses were trained and validated using 1 iteration per 50 each. The original ResNet50 model was trained on the ImageNet dataset to obtain model weights, which were then loaded into the final fully connected layer. The ResNet50 model for recognizing and classifying corn kernel mildew was fine-tuned using an expanded image dataset augmented with additional data. In this process, 80% of the images were used for training and 20% for validation in the training set, which was constructed based on the classification of corn mildew detection. The 2D grayscale image slices of each corn kernel were utilized as a test set to evaluate the ResNet50 model for detecting corn mildew. The model output a value of “1” for a mildewed image and “0” for a mildew-free image and then calculated the proportion of images with mildew features in the results. By calculating the proportion of images with mildew features in the results, the degree of mildew in maize kernels could be classified as light, medium, or heavy.

3. Results

3.1. Characterization of Moldy Maize Kernels

3.1.1. Greyscale Slice Image Analysis

The imaging principle of X-ray tomography technology is based on the fact that the density and structure of each tissue in the scanned object differ, resulting in varying X-ray absorption capabilities. This led to the formation of grayscale images with different shades, which could be observed and analyzed to draw conclusions about the structure of each seed grain. Taking maize sample 8 as an example, Figure 5 below shows 2D sliced grayscale images of maize sample 8 in the xy and yz planes at four time points from 0 d to 15 d. When observing the xy-plane slice images, it is evident that the brightness of the embryo part gradually decreased and the grey value diminished as the molding time progressed. This decrease was notably lower than the grey value of the endosperm. Figure 5 shows sectional images of maize kernels in the yz plane, revealing a significant presence of mold spores at the embryo and endosperm pores of the kernel at 10 d, as well as mold originating from the tip of the maize kernel.

3.1.2. 3D Model Image Analysis

3D models of embryos, maize kernels, endosperm, seed coats, pores, and cracks were analyzed for each sample from 0 to 15 d after pollination. Taking sample 8 kernel as an example, the reconstructed model of the maize kernel embryo is shown in Figure 6a. The model’s embryonic face sheet was smooth at 0 d, with folds and bulges appearing at 5 d. The model became covered with dense spots and erosion pits at 10 d and 15 d. The reconstructed model of the pores and cracks of the entire maize kernel is depicted in Figure 6b, illustrating the continuously evolving morphology of the kernel’s pores and cracks from 0 to 15 d. The reconstructed model of the maize kernel endosperm and testa is shown in Figure 6c, revealing the gradual emergence of spherical structures in the testa from 0 to 15 d, which were identified as mold spores. The reconstructed model of the entire maize kernel is shown in Figure 6d. According to Figure 4, it is evident that the geometric model of the embryo was significantly damaged during the maize infestation by mold. The internal structure of the kernel, infested by molds, underwent dynamic and irregular changes, leading to uncertain morphological alterations in the internal structure of the maize kernel.

3.2. Quantitative Analysis of Moldy Maize Kernels

The volumetric parameters of maize kernel volume, embryo volume, endosperm, and seed coat were extracted from maize samples collected from 0 to 15 d after pollination. These three sets of volumetric parameters were quantitatively analyzed and plotted, as shown in Figure 7, Figure 8 and Figure 9, with trend and standard deviation plots of the volumetric parameters. The analysis revealed that the three groups of volumetric parameters for each sample increased at 5 d compared to the initial kernel. The volumetric parameters for each maize sample exhibited a consistent pattern of initial increase followed by a decrease during the incubation stage. It can be seen from Figure 7, Figure 8 and Figure 9 that with the passage of time in the mold culture, the parameters of grain volume, embryo volume, and endosperm volume of eight groups of corn grain samples increased initially and then decreased. This was because corn mold requires corn grains to absorb external water, leading to a gradual increase in the volume of corn grains, embryo, and endosperm during the moisture absorption stage of corn grains from 0 days to 5 days in this experiment. After the fifth day of culture, the metabolic activity of mold in maize grains increased dramatically. This led to significant consumption of water and biological media in the maize grains, resulting in a decrease in the volume parameters of the maize grains, embryos, and endosperm.

3.3. Maize Kernel Mold Identification and Degree Classification

3.3.1. ResNet50 Improved Model Training Results

Based on the two-dimensional sliced grayscale image of a corn kernel obtained by X-ray tomography, mold and mold-free feature images were extracted. The ResNet50 model was enhanced for the rapid detection of corn mold. The training accuracy and validation accuracy of the ResNet50 model varied during training, as depicted in Figure 10a, and the classification loss also varied during training, as shown in Figure 10b. With iterative adjustments, its training accuracy exceeded 99%, and the classification loss was less than 0.1.

3.3.2. ResNet50 Improved Model Training Results

The enhanced ResNet50 model was utilized to detect and identify 1000 images without mildew features and 1000 images with mildew features selected from each set of maize kernel slice images from 0 to 15 d, with 93% accuracy on the test set. The ratio of images with mildew features to the entire set of sliced images of maize kernels was utilized as a parameter to distinguish the degree of mildew. Mild mildew was designated as 0–30%, moderate mildew as 30–50%, and severe mildew as greater than 50%. Five groups of corn kernels were sampled from the 5-d, 10-d, and 15-d sections of the image set. A total of 100 to 1300 images were taken for mildew detection, and the results are shown in

Table 2. ResNet50 improved model test mildew feature classification table.

Image Quality	Incubation Phase	Group 1	Group 2	Group 3	Group 4	Group 5
Non-moldy	5 d	614	461	42	1052	503
	10 d	436	442	33	843	423
	15 d	356	345	23	681	353
Moldy	5 d	587	740	1159	149	698
	10 d	765	759	1168	358	778
	15 d	845	856	1178	520	848

below. The percentage of mold images detected in each group was calculated and plotted in Figure 11. The comparison of group 3 with the other groups in Table 2 indicates that the degree of mold in different maize kernels was not dependent on the incubation time but was closely associated with the amount of bacteria present on the maize seeds themselves. Figure 11 shows that 4–5 d and 4–10 d were mild mold, 4–15 d and 1–5 d were moderate mold, and the remaining groups were heavy mold.

3.4. Summary of This Chapter

In this study, CT scanning technology was used to obtain slice images of moldy corn grains. Structural models of corn grains at each stage were reconstructed to visually analyze the characteristic changes caused by mold infection. Additionally, a deep learning network model was integrated to facilitate the rapid detection of corn grain mildew. Compared with traditional methods such as dry slice culture, sensory detection, and microbial activity determination, the rapid detection method for identifying corn grain mildew defects, which integrated CT scanning technology and deep learning, could swiftly and accurately detect corn grain mildew in the early stages. Furthermore, it allowed for the non-destructive observation of internal structural changes in corn kernels following mold infection.

4. Discussion

In this experiment, eight groups of corn samples were scanned and analyzed using X-ray tomography (CT). The samples were examined at incubation times of 0, 5, 10, and 15 days (0 d, 5 d, 10 d, 15 d). The geometrical models of the embryo, endosperm, and seed grain of the eight corn samples were reconstructed at each of the four incubation stages. The reconstructed models and two-dimensional grayscale images of the embryo, endosperm, husk, and seed kernel of corn sample 8 were analyzed using qualitative analysis. The models of the embryo at incubation times from 0 to 15 days revealed a gradual increase in damage to the corn embryo during the pre-molding stage. The internal infection of the corn kernel originated from the embryo and spread gradually to the endosperm. This study also examined the geometric changes in the embryo, endosperm, seed coat, internal pores, and cracks in the corn; however, the findings were inconclusive. This finding aligns with a previous study indicating that maize embryos contain a significant amount of mold and are rich in nutrients. Additionally, it suggests that maize mold originates from the embryo covering. Utilizing quantitative analysis methods to investigate mold infestation in corn following alterations in its corn kernel embryo, endosperm, and kernel volume parameters, it was determined that, over time, eight groups of corn kernel samples exhibited growth followed by a decrease in the parameters of the embryo, endosperm, seed coat, and kernel volume. The observed non-linear pattern of change was attributed to the corn mold’s hygroscopic nature during the process and the metabolic activity of mold. Mold metabolism in maize led to the decomposition of the corn’s chemical composition, consumption of free oxygen atoms inside the corn kernel, decomposition of water and CO₂, expansion of the corn’s biological macromolecules, and reproduction of spores. Utilizing CT scanning to non-destructively acquire sliced images of corn kernels, in conjunction with the deep learning ResNet50 model for identifying and detecting moldy corn, resulted in an identification accuracy exceeding 93%. This approach enabled the use of X-ray tomography (CT) to rapidly detect corn mold by analyzing a set of sliced kernel images and classifying the degree of mold contamination in the corn. Consequently, it allowed for the accurate prediction of whether the sampled corn was moldy or not. The study of corn kernels using CT scanning combined with the ResNet50 deep learning model has the following expectations:

(1) Research on characterizing and rapidly detecting other defects of corn kernels, such as insect damage and germination, needs to be expanded in the next step.

(2) In this study, the accuracy of mold detection reached 93%. How we can enhance the algorithm model to further improve the accuracy of mold detection in corn kernels now needs to be determined.

(3) The feasibility of utilizing this method to detect defects in wheat, soybeans, broad beans, and other crops needs to be explored and analyzed.

(4) The process of mold infestation of corn kernels and different types of mold damage to corn characteristics still need to be further explored.

5. Conclusions

This study demonstrates that X-ray scanning technology is an effective means for the real-time monitoring of internal structural changes in moldy maize kernels. The scanned data were used to reconstruct two-dimensional grayscale slices of maize kernels. Image processing and model reconstruction algorithms were then used to achieve the three-dimensional visualization of maize kernels. This allowed for the non-destructive monitoring of changes in the internal structure of maize kernels with varying levels of mold, as well as the calculation of parameters such as kernel size, embryo size, pore volume, and more. Through qualitative analysis, it was concluded that corn mold originates from the embryo and gradually spreads out. The quantitative analysis indicated that the volumetric parameters of the corn kernel, embryo, endosperm, and seed coat exhibited a pattern of increasing and then decreasing during the mold cultivation process. Additionally, by integrating deep learning with the reconstructed two-dimensional slices of corn kernels, the rapid and non-destructive detection of mold and the classification of the degree of mold in corn were achieved. This combined technology not only offers a viable strategy for studying the structural characteristics of moldy maize grains but also enables quick prediction of regional mold situations during grain storage, thereby reducing unnecessary economic losses. This study offers assistance in the development of equipment for monitoring and rapidly detecting mold in maize kernels.