Abstract
Due to the tastiness of mushroom, this edible fungus often appears in people’s daily meals. Nevertheless, there are still various mushroom species that have not been identified. Thus, the automatic identification of mushroom toxicity is of great value. A number of methods are commonly employed to recognize mushroom toxicity, such as folk experience, chemical testing, animal experiments, and fungal classification, all of which cannot produce quick, accurate results and have a complicated cycle. To solve these problems, in this paper, we proposed an automatic toxicity identification method based on visual features. The proposed method regards toxicity identification as a binary classification problem. First, intuitive and easily accessible appearance data, such as the cap shape and color of mushrooms, were taken as features. Second, the missing data in any of the features were handled in two ways. Finally, three pattern-recognition methods, including logistic regression, support vector machine, and multigrained cascade forest, were used to construct 3 different toxicity classifiers for mushrooms. Compared with the logistic regression and support vector machine classifiers, the multigrained cascade forest classifier had better performance with an accuracy of approximately 98%, enhancing the possibility of preventing food poisoning. These classifiers can recognize the toxicity of mushrooms—even that of some unknown species—according to their appearance features and important social and application value.
1. Introduction
Mushrooms are the fleshy fruiting bodies of certain fungus, some of which are edible, but a minority of them are toxic [1]. Every year, a large number of people die [2, 3] from eating poisonous mushrooms. It is useful to identify whether a mushroom is poisonous according to the appearance features of the mushroom. The automatic recognition of mushroom toxicity has important social and application value in effectively preventing food poisoning [4].
Current methods of recognizing poisonous mushrooms can be roughly divided into four categories: chemical determination, animal experimentation [5], fungal classification, and folk experience [6]. At present, the research of poisonous mushrooms based on these methods not only has been imperfect but also has left much to be desired [7].
The classification of poisonous mushrooms has evolved from the biological level to the molecular level [2]. Therefore, the application of chemical determination methods to detect poisonous mushrooms is becoming increasingly popular [8]. However, there are strict requirements for the experimental conditions, which are often limited to the laboratory. Due to cumbersome handling and the great number of unstable toxins, the method of toxic chemical detection cannot be used to distinguish edible mushrooms from poisonous ones [9]. This approach requires professional knowledge and is, therefore, not suitable for the average person.
Generally, mushrooms with intact cells, bright colors, and the lack of birds and insects interacting with them are likely to be poisonous, particularly if they are found in places where animals are foraging. To investigate the above situation empirically, the animal acute toxicity test is commonly used to classify poisonous mushrooms [10]. Although the methods involved are simple, they carry some limitations, such as low efficiency, material and dosage concerns, and the varying sensitivities of different animals. Therefore, special institutions or facilities are needed to facilitate the application of these methods.
Fungal recognition includes three aspects: identification, classification, and phylogeny [11]. The development of fungal taxonomy has gone through two stages: traditional taxonomy and molecular biology. These methods have mainly been used to identify the mushrooms’ species. The aim of these methods is subjective, however, because fungi contain many species and complex morphological features. These methods are limited to applications involving the artificial cultivation of fungi and are only suitable for professionals. Therefore, the identification of poisonous mushrooms is not straightforward.
For a long time, humans have recognized poisonous mushrooms by observing the shape, color, odor, and secretion features empirically [12]. This method is more intuitive, but is of low accuracy proven by the annual poisoning events. Thus, it is not a reliable method for identifying whether mushrooms are poisonous. However, this method relies on background knowledge acquired by humans. People get a lot of background knowledge and experiences so that the recognition accuracy rate is high. Otherwise, the accuracy rate is low. In this paper, automatic identification can break through the limitation to determine whether it is toxic. The machine learning methods not only do not require background knowledge but also can identify unknown species.
These mushroom toxicity recognition methods have some limitations, such as low accuracy, unqualified detection of unknown toxins, strict requirements for the experimental environments, sufficient professional knowledge, and complex experimental cycles. To solve these problems, an automatic model for mushroom recognition based on appearance features is constructed in this paper. According to the observed mushroom appearance data, a poisonous mushroom can be automatically and accurately identified by the proposed model.
With the advent of the data age, machine learning and deep learning have become the core of artificial intelligence [13]. In recent years, machine learning techniques have been used to identify the toxicity of mushrooms. Chaoqun [14] used machine learning models to identify poisonous mushrooms in an application. The android-based toadstool identification system can effectively classify toadstool in real time [15]. Zhifeng [16] proposed decision fusion based on the stacking algorithm to improve the accuracy of classification methods. The image database of mushroom, obtained from the Internet by Python Crawler, was constructed by Shuaichang et al. [17]. The model-based transfer learning and the Adam algorithm as the model optimization method were applied to construct the model structure of mushrooms’ image recognition.
Deep neural networks require large-scale data volumes, making already complex models even more complicated. Machine learning has unique advantages for small-sample problems. For the identification of poisonous mushrooms, three different pattern recognition methods are discussed in this paper: logistic regression, support vector machine (SVM) [18], and multigrained cascade forest (gcForest) [19]. Mushroom toxicity recognition is regarded as a problem of binary classification. By observing the appearance features of mushrooms, these machine learning methods are used to determine whether a mushroom is toxic. gcForest has the following advantages: (1) feature-based learning and iterative classification through gcForest, which has the best performance; (2) no need for professional knowledge to use the system; (3) if the unknown mushroom varieties are toxic, they can be identified quickly; and (4) independence from the effects of the natural environment unlike other methods, expanding the scope of its use.
2. Methods
Driven by big data, deep neural networks (DNNs) show great potential [20]. DNNs have achieved remarkable success in various applications. However, deep neural networks have too many hyperparameters, and their learning performance critically depends on their careful tuning. At the same time, it has an impact on DNNs that is difficult to rein in [21].
In response to the above difficulties, Zhou and Feng proposed a multigrained cascade forest framework [19]. gcForest is a decision tree ensemble method. The gcForest method, which consists of multigrained scanning and a cascade forest, is explained in detail below.
As illustrated in Figure 1, after we had obtained the dataset, we checked the integrity of dataset. Firstly, we solved the problem of missing data by Process A and Process B. Secondly, the complete data was divided into test dataset and training dataset. Then, the training dataset was input to the constructed classifier model. After the model judgment, the result of classification depended on the category with high probability. Therefore, the experimental process is shown in Figure 1 in this paper.
In this paper, we use the cascade forest to discriminate the toxicity of mushrooms, which is regarded as a binary classification problem. The 22-dimensional features are easily obtained and used as data input. After running the gcForest model, the toxicity identification results of the corresponding samples can be obtained, as shown in Figure 1.
The gcForest classifier model can accurately judge whether a mushroom is poisonous in a timely manner; thus, it has strong practicality. The classifier also has the following features: (1) ease of trainability; (2) compatibility with various datasets; and (3) an adaptively adjustable hierarchy for the cascade structure depending on the desired the complexity of the model; such a small sample dataset can achieve good recognition performance.
2.1. Cascade Forest Structure
Feature learning by deep neural networks (DNNs) depends on the layer-by-layer processing of the original features [22]. Similar to DNNs, gcForest uses a cascade forest structure where the information of each layer is processed in an upper layer and the result is delivered to the next layer.
As illustrated in Figure 2, suppose that the mushrooms can be divided into two categories: toxic and nontoxic. The leaf node of each forest will output a two-dimensional class as a vector, which is concatenated for the re-representation of the original input. Therefore, the next level of the cascade will receive augmented features, and the vector dimension of the input feature will be . Namely, the feature dimension is equal to the number of enhanced features + the number of original (or transformed) features. Each forest will output two-dimensional class vectors, which are connected to the input features to produce the next original input. Additionally, each level contains several classifiers capable of ensemble learning [23].
For simplicity, we suppose that each layer of the cascade forest structure consists of two random forests and two completely random forests (CForests) [24]. Each forest is an aggregation of decision trees [25].
A completely random forest contains a number of completely random trees, generated by randomly selecting a feature for splitting at each node of the tree, and the tree is grown until each leaf node contains only the same class in Figure 3. Similarly, each random forest contains a number of trees, by randomly selecting number of features as candidates ( is the number of input features) and choosing the one with the best Gini index (which refers to the index of optimal features when CART is used for classification problems) for splitting. The tree is grown tree until each leaf node contains only the same class of instances [26].
As illustrated in Figure 3, suppose that there are two classes, each forest will generate a two-dimensional class vector. Different symbols in the leaf nodes imply different classes. First, the red color highlights the paths along which the instance traverses to the leaf nodes, and each forest will generate the class distribution by counting the percentage of different classes of training examples at the leaf node where the concerned instance falls [19]. Then, the estimated class distribution forms a class vector. This vector is concatenated with the original feature vector as the input to the next level of the cascade. Finally, the cascade result vectors are averaged to two-dimensional vectors. The class of the maximum value is used to determine whether the mushroom is poisonous.
To reduce the occurrence of overfitting produced by each forest, k-fold cross-validation is used in this algorithm [27]. Each instance will be used as training data for times to generate vectors, which are finally averaged to produce the final class vector that represents the augmented features for the next level of the cascade. In other words, these augmented features are partially input in the new cascade layer, and the performance of the entire cascade on the validation set is evaluated. And if there is no significant performance improvement, the training process will automatically terminate. For that reason, the number of cascade layers can be adjusted automatically. Contrary to complex and fixed depth neural networks, gcForest can adaptively stop training to determine the number of cascade layers required.
2.2. Multigrained Scanning
The deep neural network has a powerful advantage in dealing with spatiotemporal and sequence features. Similarly, gcForest takes into account multigrained scanning prior to the cascade structure to enhance the feature learning ability of the overall structure. As shown in Figure 4, gcForest scans the original features using a multigrained sliding window. There are 22 raw features for the sample mushroom, and a window size of 8 features is used; this results in the production of 15 feature vectors. Assuming that there are two classes of mushroom toxicity, the original 22-dimensional feature vector corresponds to 60 transformed dimensions.
Deep neural networks are effective in handling feature relationships. Inspired by this recognition, we enhance cascade forest with a procedure of multigrained scanning.
As illustrated in Figure 4, firstly, the complete appearance feature in the dataset is dimensions. A sliding sampling window with length is used to obtain subsample vectors with the size of v-dimensional feature. The process is similar to the sliding convolutional core , with the stride is 1 and the padding is 0. Secondly, each of the instances extracted is used to train a completely random tree forest and a random forest . Thirdly, each forest generates a length probability vector, where is the number of categories, here equal to two (corresponding to whether the mushroom is poisonous). Finally, the results of forests at each level are joined together to generate the output samples. A representation vector is generated in each forest, and these vectors can be concatenated together to obtain the final sample output in Figure 4. Therefore, there is output feature -dimensions [28].
Suppose the sliding windows with sizes of 8 and 12 features will generate 60-dimensional and 44-dimensional feature vector for each original training example, respectively. The transformed feature vectors, which contain augmented features by the previous grade, can be used to train the 2nd grade and 3rd grade of cascade forests, respectively. The procedure, in every three levels, will iterate until the boost of accuracy rate is less than the threshold, as illustrated in Figure 5. The repetition process of the training is completed.
Feature vectors of dataset will enter into the cascade forest structure in batches and connect with the upper output data to increase the disturbance of mushroom samples. The input feature vector of the cascaded forest structure will be connected with the output data of the first layer to form the input data of the second layer. Therefore, we have obtained the transforming features from the process. The final model is actually a cascade of cascade forests [19]. Each level in the cascade consists of multiple grades (cascade forests), and each corresponds to a grain of scanning, as shown in Figure 5.
The last layer classifies the upper-layer input data. Counting the percentage of different classes of training examples at the leaf node where the concerned instance falls, we then compute the average value in all forests to generate the maximum value of the class distribution to determine whether the mushroom is poisonous in Figure 5.
3. Results
3.1. Mushroom Dataset
The mushroom dataset provided by the University of California, Irvine was used to classify the toxicity of poisonous mushrooms [29]. The input features of the mushroom include class, cap shape, cap surface, cap color, bruises, odor, gill appendages, gill spacing, gill size, gill color, stalk shape, stalk root, stem-surface-above-ring, stem-surface-below-ring, stem-color-above-ring, stem-color-below-ring, veil type, veil color, ring number, ring type, spore print color, population, and habitat, for a total of 23 features (see Table 1). These features, which can be observed directly, are classified with the feature calculation. There are 8,124 recordings of mushroom data, which can be divided into two nearly balanced classes: poisonous (48%) and nonpoisonous (52%). We have created a new table (see Table 1) based on data attributes.
Each recording of mushroom has 22 features and one class label; however, part of the recording in the mushrooms’ dataset is missed in the stalk-root feature. There are two ways to solve the missing data problem in Table 2:(1)Using a KNN to complete the missing data, which is called Process A in this paper. The value of parameter K is set to 12.(2)Treating missing attribute values as special values [30]. A special value (“m”) is used to fill in the missing values of the stalk-root, which is called Process B.
The first row shows an example of complete data. The second row shows that the original data of the stalk-root is missing. In Process A, the KNN algorithm predicts the value of stalk-root (“c”). In Process B, a special value (“m”) is used as the new label for all of the missing data. And the range of the stalk-root is changed to .
The two processes were specifically verified and compared in the experiment to determine the feature’s effect in identifying mushroom toxicity.
The missing values of the original data are processed to obtain a complete list of character data. In the experiment, the error in the numerical data is small relative to the result, and the data can be converted to the target type data by LabelEncoder [31].
When the support vector machine classifier is used, a toxicity category value of or is generally assigned. Therefore, the range of the feature data toxicity is changed to , which is advantageous for obtaining the hyperplane of toxicity.
Three models were run in this paper: logistic regression, SVM, and gcForest.
3.2. Evaluation Standard
To evaluate the classification models, we adopt two measures.
The index for evaluating the performance of the classifier is the accuracy of classification generally. For the test dataset, it is the ratio of the correctly labeled samples to the whole pool of samples in (1).
Assuming that the model of classification is , given bywhere is the test sample size and is the indicator function.
The second is a standard evaluation performance of the binary classification used, namely, recall and precision [32]. In addition, there are F1-score and ROC curve.
The mushroom class has two possible predicted classes: edible (e) and poisonous (p). And there are four situations in confusion matrix (see Table 3): true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).
The precision and recall are given by the following relation:
Precision is defined as the number of true positives TP over the number of true positives plus the number of false positives FP [32].
Recall is defined as the number of true positives TP over the number of true positives plus the number of false negatives FN [32].
These quantities are also related to the F1-score, which is defined as the harmonic mean of precision and recall [32].
The ROC curve describes the change process of classifier performance changing with the change of thresholds settings, where the x-axis represents false positive rate and the y-axis represents true positive rate [33].
Ideally, we want the fraction of correct positive class predictions to be 1 (top of the plot) and the score of incorrect negative class prediction to be 0 (left of the plot) [33].
3.3. Experimental Analysis Using Logistic Regression
Assume that is the feature set for mushrooms containing features such as cap shape, cap surface, cap color, swelling, and odor. According to the features, we calculate the probability of this mushroom as toxic, which is called a score. This score is used as the input of the sigmoid function:
At the same time, a sigmoid function is used to convert the scores into values within the interval of [0,1].
In the logistic regression function, the maximum likelihood function is applied to obtain the parameters of the model. Then, a logistic regression model is constructed, which is turned into an optimization problem using the log-likelihood function [34]. During the period of learning the logistic regression model, the gradient descent method or other improved scores is usually used [35].
Analysis of the cleaned features may result in excessively long training time or memory overflow, due to an excessively large feature matrix. Consequently, rescreening is the most direct and effective method to screen effective features with a random logistic regression model [36]. There are differences between the data obtained from Processes A and B, and so different effective features will be obtained.
As shown in Table 4, the effective feature stalk-color-below-ring is different across the different processes. The only variable is stalk-surface-above-ring. There may be a correlation functional relationship between stalk-surface-above-ring and stalk-color-below-ring. Although the values of stalk-surface-above-ring are different processes, this feature is one of the reasons that affect accuracy.
First, the dataset is divided into 4 even parts; 3 of the parts are selected as the training set, while the remaining 3 are used as the testing set. Second, we establish a model for discriminant poisonous mushrooms on the basis of the training dataset and then judge whether the testing data indicate poisonousness. Then, statistics on the accuracy of prediction are calculated. Finally, the statistics for the effective feature analysis are compared with the statistics for the analysis using all the input features, and the results are described in Table 5.
It is easy to observe that, following Process A, the accuracy obtained with effective features and with all features differs by approximately 0.01% with maximum and minimum accuracy values of 0.955 and 0.940, respectively. Following Process B, the effects of the effective feature set and of the overall feature set on the accuracy vary greatly, with a minimum accuracy of 0.940 and a maximum accuracy of 0.955 (see Table 5). The effective feature set is an important factor in judging accuracy, but an incorrectly judged feature could reduce the accuracy of estimating mushroom toxicity.
3.4. Experimental Analysis Using SVM
The labels for the toxicity category are set to indicating whether a mushroom is toxic. The other features are converted to using LabelEncoder to represent the samples. The dataset is divided into a training set with 75% of the data and a testing set with the remaining 25% by random selection.
Support vector machine (SVM) is essentially a binary classification model [37]. The basic idea is to find the optimal classification line (surface) from the feature space. The optimal demarcation line maximizes the distance in the binary classification of data.
The minimum accuracy under Process A is approximately 0.02% higher than that under Process B (see Table 6) on the test dataset. Further experiments with KNN constraints on the original data will produce partially correct data. Therefore, increasing the correct proportion of mushroom features can improve the accuracy of mushroom toxicity classification.
3.5. Experimental Analysis on gcForest
To reduce the contrast error between the three experiments, 25% of the dataset is used as the testing data. In this paper, the gcForest model requires two stages: multigrained scanning and the cascade forest [19]. Multigrained scanning generates the features, and the cascade forests use multiple forest cascades to derive prediction probability results [38]. The criterion for selecting each parameter is that the accuracy should be the lowest; this process is then iterated. Due to the inconsistent magnitude of the parameters, the weight assignment will cause analysis errors. The influencing factors of the parameters on the experiment will not be commented upon in detail (e.g., the number of trees in a random forest, the size of the sliding window, the sliding step, the number of cascading random forest, and the number of trees in a single cascade random forest). The maximum fluctuation in the experimental results is less than 8%.
The maximum average denotes the average of the sum of the maximum values of each parameter in the interval range. In the experimental results, when the original data are further tested under the KNN constraint, the processing will produce erroneous data with a higher error than the new category (Process B). Incorrectly judged features will reduce the accuracy of discriminating mushroom toxicity. Thus, the effect of predicting incorrect data in the experiment is greater than that from filling in special values. Table 7 shows that the average accuracy of the multigrained cascade forest classifier fluctuates between a maximum of 0.9835 and a minimum of 0.9260 on the test dataset.
3.6. Analysis of the Results
Three classifiers were built on a mushroom dataset to determine whether the mushrooms are poisonous according to their features.
In the logistic regression experiment, the accuracy of each feature in determining mushroom toxicity is calculated following steps such as dimension reduction and computing accuracy. At the same time, the results from the SVM and gcForest models are analyzed separately.
Based on the mushroom dataset, SVM has a slightly higher effect with Process A than that with Process B. Predicting erroneous data in the logistic regression and gcForest models will produce more errors than will adding a new class (Process B) from a certain feature; the data requirements are stricter, and incorrectly judged features will reduce the accuracy of determining mushroom toxicity. Table 8 shows that, among the three classifiers, the average accuracy of gcForest is 0.9835 and fluctuates less than 8%. The implementation of gcForest obviously improves the accuracy of the classification, but this improvement is not stable. Thus, it is necessary to improve the experiment to further improve the effect of the gcForest classification.
In addition, we used four indexes (precision, recall, F1-score, and ROC curve) to compare the performance of the three classification algorithms in this paper (see Table 9 and Figure 6). AUC stands for the area under the ROC curve.
From Table 9 and Figure 6, it is verified that Process B is better than Process A, and more features are beneficial to classification. In other words, the effective feature set is an important factor in judging accuracy, but an incorrectly judged feature could reduce the accuracy of estimating mushroom toxicity.
According to Table 10 and Figures 7 and 8 can be described as follows. It is verified that the result of partially correct data by KNN constraint is better than that by Process B.
(a)
(b)
The gcForest classifier obtains the higher precision rate and the lower recall rate than other classifiers on the balanced dataset. In addition, we intuitively get the information that ACU and ROC have reached the highest value on the gcForest classifier from Figures 7 and 8 and Table 10. The results further prove the outstanding classifier in the three classifiers and the applicability of gcForest classifier on this dataset.
4. Discussion
In this paper, we proposed an automatic mushroom toxicity identification method. The gcForest method proposed by Zhou has a recognition accuracy of more than 98% [19]. Based on the high accuracy requirements, we analyzed three pattern classification models. Logistic regression yielded classification results by analyzing the effective features necessary to identify toxicity. The accuracy of the SVM method is better than that of logistic regression. Compared to SVM and logistic regression, gcForest achieved better results in terms of identification accuracy. Therefore, gcForest is a good method for automatically identifying whether a mushroom is poisonous.
A number of common mushroom toxicity recognition methods are currently in use. These methods use different contributions for determining toxicity, but they have a number of limitations, such as low accuracy, unsatisfactory detection of unknown toxins, the need for a strict experimental environment, and sufficient professional knowledge and complex experimental testing techniques. To circumvent the limitations of these methods and apply them to small-sample data analysis, we used machine learning. In contrast to deep neural networks, which require great effort in hyperparameter tuning, gcForest is much easier to train and can be applied to different kinds of data in different domains. The gcForest algorithm has the following advantages: (1) it has a simple structure; (2) it can be applied to datasets of different sizes; (3) the testing techniques and handling are simple; and (4) for our experiments, mushroom toxicity is recognized quickly. Feature-based learning and iterative classifiers in the gcForest method have the best performance among the three methods of machine learning used proposed in this paper. This automatic identification method is suitable for nonprofessional identification and for unknown mushroom varieties.
Among the three machine learning methods, gcForest yielded the best accuracy. However, the stability of the classifier, as shown in Table 8, needs to be improved. A reason for this error may come from the absence of features, the feature labeled stem-surface-above-ring, or it may come from the algorithm itself. Therefore, improving stability is a top priority when trying to improve the accuracy of the classifier. Since gcForest can be used for different types of datasets and recognition based on image features is more convenient than with the other classifiers [39], this method of identifying whether a mushroom is toxic can be extended to image recognition. Nonetheless, there is currently no dataset of mushroom images.
In this paper, we studied whether mushrooms were toxic, by comparing, analyzing, and summarizing four classic and traditional identification methods. According to their shortcomings, we adopted automatic identification methods to conduct the analysis based on machine learning. Based on the mushroom dataset, three pattern recognition analyses were performed. In contrast, gcForest has higher accuracy, but its stability needs to be improved. The used method identifies whether the unspecified mushroom species is toxic in a timely manner. Because this automated identification method is not affected by the natural environment, it has important social and application value in effectively preventing food poisoning. Meanwhile, people also need to improve their safety awareness of mushrooms.
5. Conclusions
In this paper, multigrained cascade forest was used to determine whether a mushroom was poisonous based on its appearance features. LabelEncoder was used to encode the processed data to form numerical data for the mushroom dataset. According to the analysis of the dataset features, the accuracy of gcForest in data classification was approximately 98%. The maximum fluctuation of its accuracy was less than 8%, however, so the stability of the classifier needs to be improved. The gcForest structure can be used not only for large data but also for small-sized samples, and the adaptive selection of cascade layers can achieve the same accuracy as fixed patterns of deep neural networks in other datasets.
At present, research on toxins in poisonous mushrooms is still underway [40], and cases of mushroom poisoning still occur. Therefore, it is necessary to establish an automatic model for the appearance feature recognition of mushrooms’ toxicity. Compared with other mushroom identification methods, the method proposed in this paper uses short cycles, is highly efficient, has low requirements in terms of the natural environment, and results in the timely identification of the toxicity of unknown species. Consequently, this method has important social and application value.
Data Availability
The mushroom data used to support the findings of this study have been deposited in the UCI (https://archive.ics.uci.edu/ml/datasets/mushroom) repository.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Authors’ Contributions
All authors have made significant contributions to this research. JXD conceived the research ideas. YYW designed the experiments, conducted the data analysis, and provided the writing of this paper. YYW performed the majority of the data processing, and the University of California Irvine [29] provided rapeseed mushrooms data. HBZ provided important insights and suggestions into this research from the perspective of a computer scientist. All authors read and approved the final manuscript.
Acknowledgments
This work was supported by the National Key Research and Development Program of China (No. 2019YFC1604700), the National Natural Science Foundation of China (Nos. 61673186 and 61871196), the Natural Science Foundation of Fujian Province of China (No. 2019J01082), the Promotion Program for Young and Middle-Aged Teacher in Science and Technology Research of Huaqiao University (Nos. ZQN-YX601 and 18014083020), and the Subsidized Project for Postgraduates’ Innovative Fund in Scientific Research of Huaqiao University.