Efficiency in Orchid Species Classification: A Transfer Learning-Based Approach

1. Introduction

Orchid is a terrestrial herb and also a flowering plant. It has the largest families among the botanical plant. It is reported that orchid has almost 800 genera and 22,500 species.^1,2 Orchid has elegant flower posture, quiet flower fragrance, rich colors and noble moral, therefore, it has high ornamental value and is deeply loved by people. Orchid is also one of the commercial crops in the floriculture sector with horticultural and medicinal value. It has received comprehensive attention in biology, evolutionary chemistry, taxonomy, cytology, chemistry, hybrid breeding and so on, because of its ornamental value and high economic value in market.

With the improvement of people’s living standards, people’s demand for orchid planting and appreciation has gradually increased. As orchid plants have many similarities in flower petals, texture and color,³ it is a hugely challenging task for people to correctly and quickly distinguish them, especially for those people who do not know the characteristics of orchid species. Plant experts may sometimes misclassify them due to their similar visual characteristics. It is also a very time-consuming work to correctly classify the orchids because it needs to extract many similar features of orchid. Besides, when planting orchids, it is necessary to classify orchid species, as they need different growing conditions. Therefore it is necessary to develop an automatic classification algorithm that is expected to make it easier to classify orchid species.

Since the current state-of-the-art general-purpose orchid classification algorithms based on different kinds of technologies, such as orchid classification algorithm based on artificial feature extraction, orchid classification algorithm based on machine learning, orchid classification algorithm based on deep learning methods and so on, we have low classification accuracy due to lack of enough training datasets,^4,5 and there are few studies specifically for the classification of orchid species, that make them unsuitable for orchid species classification.

In order to solve the problem above, an efficient classification algorithm of orchid species based on transfer learning is proposed in this paper. The achievement of the paper lies in the fact that we successfully use the transfer learning technology to realize the efficient classification for 12 orchid species. Specifically, in our scheme, first we obtained 12 types image sets of orchid with number of 12,227 images by network and field photography. Second, we analyzed and studied the characteristics existed in the acquired orchid images set. Third, we found some processing relationships between acquired orchid datasets and other datasets. At last, we used transfer learning technology to realize the efficient classification function for orchid species based on processing relationships. As a result, the problem above of low classification accuracy rate can be effectively improved by our proposed classification algorithm in this paper. The experimental results show that our proposed classification algorithm in this paper realized 96.16% classification accuracy rate compared with not using transfer learning method, which can improve the whole classification efficiency of orchid species.

The rest of this paper is organized as follows. In Sec. 2, the related works of orchid species classification are introduced. Our proposed classification algorithm is presented in Sec. 3. In Sec. 4, we show our experimental results and analyses. Finally, Sec. 5 concludes.

2. Related Works

In recent years, many research works have been taken for the classification, recognition or detection for orchid or other related plant species.

A supervised learning-based orchid types classification method based on feature and color extraction was proposed by Andono et al.⁶ for orchid plants. In this paper, they mainly used support vector machine, Naïve Bayes and k-nearest neighbor algorithm to classify 15 orchid plants. An identification method of orchids variety based on Fourier Transform Infrared Spectroscopy and Stacked Sparse Auto-Encoder is developed by Chen et al.¹⁴ for orchid genotypes. In their method, they mainly used Fourier Transform Infrared Spectroscopy and three models, namely SSAE, SVM, and KNN, to classify 13 orchid genotypes. A visual recognition system based on a deep learning approach was presented by Arwatchananukul et al.¹⁵ to identify the Paphiopedilum orchid, where they identified the orchid species and diseases by extracting the features of orchid plant leaf by image processing technology and predicted the healthiness of the orchid plant data mining technology. A classification method of orchid species based on neural network was proposed by Sani et al.⁷ for orchid species. In their method, they used the combination of convolutional neural network and Inception-v3 feature extractor of TensorFlow platform to identify 15 orchid species, and achieved about 85.7% classification rate. An orchid classification method based on the combination of color, shape and texture features was developed by Sabri et al.⁸ for orchid species, where they used support vector machine to classify three orchid species, Dendrobium, Phalaenopsis and Vanda, and achieved 82.22% classification accuracy rate. An orchid classification method based on homogeneous ensemble of small deep convolutional neural network was proposed by Sarachai et al.⁴ for orchid species. In this paper, they mainly used global prediction network, local prediction network and ensemble neural network to realize the classification function for orchid species. An orchid leaf disease detection method based on border segmentation techniques is suggested by Fadzil et al.¹⁹ for orchid disease. In their method, they mainly used filtering technology and morphological processing technology to realize the identification for orchid leaf disease. A multilabel classification method based on deep learning was proposed by Post ⁵ for orchid features, where they designed a single output multilabel classifier with transfer learning to classify six different orchid features. A visual detection and species classification method is proposed by Puttemans and Goedemé⁹ for orchid flowers, where they used object categorization and object classification techniques in an industrial context with a very limited set of training data to realize visual detection and species classification for orchid flowers. A shallot quality classification and size identification method based on HSV color models and Naive Bayes classifier was suggested by Susanto et al.¹⁰ for Shallot quality classification, where they applied Naïve Bayes and hue saturation value color model probabilities to realize 91.67% classification accuracy for Shallot quality classification. An approach based on Bayesian network was presented by Jayech and Mahjoub¹¹ to improve the content of image classification systems, where they used three Naïve Bayes: tree augmented Naïve Bayes, forest augmented Naïve Bayes and regular Naïve Bayes to realize the image classification systems, and found that regular Naïve Bayes achieved the highest classification quality in them. Chandel et al.²⁰ presented a Paphiopedilum recognition system based on Convolutional Neural Network (CNN) combined with Inception-v3 feature extractor of TensorFlow platform, and in their method, they can achieve 98.6% recognition accuracy rate for Paphiopedilum. Zawbaa et al.¹² presented an automatic flower classification approach-based machine learning, where they used support vector machine and random forest algorithms to classify eight kinds of flowers, and found support vector machine-based algorithm can provide better classification accuracy in two algorithms above under the same testing conditions. Mohamed et al.¹⁶ suggested a plant species recognition scheme based on Bag-of-Word with SVM classifier. In their method, they mainly applied support vector machine to train different strategies according to the organs and species of plants on the basis of studying SIFT and Opponent Color SIFT. Fu et al.²¹ presented a fast and accurate detection method based on improved YOLOv3-tiny model for kiwifruit in orchard, where they proved their method has small and efficient real-time kiwifruit detection compared with Faster R-CNN with ZFNet, Faster R-CNN with VGG16, YOLOv2 and YOLOv3-tiny methods. Koirala et al.²² developed a real-time fruit detection method based on R-CNN for mango fruit, where they achieved a F1-score of 0.89 on a day-time mango image dataset and orchard fruit load estimates of between 4.6% and 15.2% of packhouse fruit counts. Wang et al.²³ proposed a segmentation method based on MR-CNN for waxberry image under orchard environment, where they realized average 97% detection accuracy and 91% recall rate, respectively, compared with K-Means method under the same datasets. Nilsback and Zisserman¹³ proposed an automated flower classification based on support vector machine for a large number of classes. In their method, they combined SIFT features and hog features to classify flower, and achieved 72.80% classification accuracy. Liu et al.¹⁷ developed an image-based large-flowered chrysanthemum cultivar recognition method based on deep learning, where they mainly used vgg16 and resnet50 to recognize chrysanthemum species, evaluated its recognition performance with different networks and analyzed how to extract features by convolution neural network with visualization and feature clustering technologies.

Although the methods above are well designed and used different kinds of technologies to realize orchid classification algorithms, such as orchid classification algorithm based on artificial feature extraction, orchid classification algorithm based on machine learning, orchid classification algorithm based on deep learning and so on. However, those methods do not completely explore the internal relationship in orchid image sets and the existing ImageNet datasets. Therefore, there is still a need to further develop a more efficient orchid classification method to improve its whole classification accuracy. Different from these classification methods above, this study proposes an efficient method for classifying orchid species using transfer learning for analyzing and studying the relationship between obtained orchid datasets and ImageNet datasets. In our approach, we primarily employed transfer learning technology to enhance the classification accuracy of orchid species, thereby significantly boosting the efficiency of orchid classification.

3. Proposed Scheme

To address the issue of low accuracy in existing orchid species classification algorithms, caused by the insufficient availability of training datasets, this paper introduces a novel classification algorithm for orchid species. It is founded on the analysis and examination of the correlation between the orchid datasets we acquired and the ImageNet datasets. This innovative approach leverages transfer learning to enhance classification accuracy.

3.1. Acquisition of orchid image sets

In this paper, we mainly obtained 12 types of image sets of orchid by network and field photography with a number 12,227, where the size, shape, light intensity, shooting angle and shooting background of captured images are different from each other in our field photography. The images of 12 orchid species are shown in Fig. 1.

In our scheme, in order to process the original orchid datasets above, we divided them into training sets and testing sets according to the division ratio of 8: 2, where the number of training sets is 9777 and the number of testing sets is 2450. We normalized the length and width of the training sets to $256\times 256$ to ensure its uniformity because of their inconsistent size of orchid image. The number distribution of orchid image sets is shown in Table 1.

3.2. Motivation for transfer learning

As deep learning based on convolutional neural network needs a large number of dataset to support its training and has a problem of over fitting in its training process,²⁵ we usually need to collect more datasets to solve the problem. However, data collection is a very time-consuming and labor-intensive work, and in some cases, it is very difficult or expensive to obtain enough training datasets. Since there are some relevance between one task and another task under some circumstances, and the large training datasets can be available in the task, while enough datasets cannot be available in another task,²⁶ under such a circumstance, we need to use the existing knowledge in one task to solve the problem in another related task.

Built on fundamental principles, we have introduced the application of transfer learning technology to address the aforementioned issue in this study. Where transfer learning mainly uses the existing knowledge to transfer to related fields, that is the trained model of one task is applied to another task, so as to help the training of new tasks.²⁷ Transfer learning can reduce the scale of required dataset and the impact of over fitting, reduce the cost and time of network training, improve the accuracy of orchid species classification.^28,29

3.3. Transfer learning-based ResNet34

In this paper, we selected ResNet34 as the deep convolution neural model and used transfer learning technology to train the ResNet34 model. We select Imagenet dataset as the original dataset of model and transfer the weight of the model to our classification method of orchid species to improve its classification accuracy and reduce its training time.

In our scheme, we import the model weight of ResNet34 and adopt the fine tuning method to train the classification of orchid species, which freezes part of the convolution layer at its micro level of the model or does not freeze them at all. Where fine tuning method refers to initialize our own network with trained parameters obtained from the trained model, and then continue training with our own data. If it freezes, it usually freezes the convolution layer in the lower layer of model, and trains the remaining convolution layer that is not frozen and full connection layer. Since the frozen layer does not participate in back propagation, it can save a lot of computing resources and time. In our scheme, we divided the ResNet34 network into five convolution blocks, and frozen from the lower to higher convolution layers. The specific dividing and freezing scheme of convolution layers are shown in Fig. 2.

4. Experimental Results and Analysis

In this section, we designed a series of simulation experiments to verify the effectiveness of our presented scheme. Our designed experiment includes the following two contents: establish experimental environment and test experimental result.

4.1. Establish experimental environment

We performed our simulation experiments on Ubuntu16.04, AMDRyzen4800H, 16G memory and 6GB NVIDIAGTX2060. The specific experimental environment in our experiment is shown in Table 2.

In order to better compare the improvement influence of the deep convolution neural network on our experiment, we unify the training super parameters and set their specific parameter as given in Table 3, where the batch size is set to 32 and the epoch is set to 100. When the training of each time is completed, the testing set is used to evaluate the classification performance of network model, and the network model with the best classification result during the training processing is selected. The optimizer uses the random gradient descent algorithm, whose momentum and weight attenuation are set to 0.9 and 0.0001, respectively. The learning rate is dynamically adjusted with the network training and the attenuation strategy of learning rate is set in fixed step attenuation way in our experiment that is the learning rate is reduced every certain epoch. In our experiment, the initial learning rate is set to 0.02, and the learning rate is adjusted to 1/2 of the current learning rate every 20 epochs, so as to avoid the network convergence caused by too large learning rate. Equation (1) is the definition of current learning rate.

In our experimental setup, we opted to enhance the analysis of classification outcomes by utilizing classification accuracy and the average training time of an epoch as the key evaluation metrics for the training model’s performance on the testing dataset. A higher classification accuracy reflects a superior model classification outcome, while a lower average training time for a single epoch indicates a more efficient training model under equivalent classification performance conditions. The definitions for accuracy and average training time of an epoch are expressed as follows.

4.2. Test experimental result

In this section, we test the effects of freezing scheme of convolution layers that is the classification accuracy and average training time in unfreezing convolution layer, freezing part of convolution layer to freezing all convolution layers on are tested with our proposed method, respectively. In our experiment, we also used K-fold method to further validate our work. In our K-fold method, we first randomly divided our orchid datasets above into mutually exclusive 10 subsets, and took the average of five random partitions to ensure its randomness. We then randomly divided 10 subsets into nine groups and the remaining one into another group, which has 10 ways to divide them. We finally treated the nine subsets of the group as the training set, and the other one as the testing set in each grouping result, which can result in 10 predictions and average them. Table 4 presents the experimental results.

Table 4 shows that the classification accuracy rate of schemes B, C, D, E and F is increased by about 10.41%, 12.65%, 12.73%, 12.69% and 12.94%, respectively, compared with that without using transfer learning. It is clearly shown that transfer learning can improve the classification accuracy rate and reduce the risk of over fitting of model by freezing some convolution layers with a fine-tuning way. InceptionV3 method and InceptionV3-STN method can achieve 91.47% and 94.13% classification accuracy rate for orchid species. At the same time, from Table 4, we also can clearly see that freezing all convolution layers will lead to a decrease in classification accuracy, which is 12.94% compared with not using transfer learning method. The main reason for it is that the model cannot converge to a certain extent. Figure 3 presents the relationship between scheme classification accuracy rate and Epoch with different schemes.

Figure 3 shows that the accuracy rate with using transfer learning method begins to converge when the epoch is about 20, which is faster than that without using transfer learning, and the over fitting problem can be alleviated to a certain extent. When all convolution layers are frozen, the accuracy of training set and testing set shows some fluctuations.

The experimental results above also show that the less the convolution layer is frozen, the better the classification effect of orchid species is. The classification accuracy of the first to three layers frozen is the same as that of the first frozen layer, but the average time difference of training an epoch is very significant. Therefore, when training time is limited and the best classification effect is not pursued, scheme C can be used to achieve the best balance between classification effect and training time.

5. Conclusion

This paper introduces an enhanced approach for efficiently classifying orchid species, leveraging a transfer learning-based methodology to elevate the precision of classification outcomes. Our framework encompasses several crucial steps: Initially, we acquired a collection of 12 distinct orchid image sets, totaling 12,227 images, through a combination of network sources and field photography. Subsequently, we meticulously analyzed and explored the distinctive traits inherent within the gathered image sets. Furthermore, we identified pertinent processing relationships between the obtained orchid datasets and other available datasets. Ultimately, we harnessed transfer learning paradigm to achieve a highly efficient classification capability for orchid species. Consequently, our proposed classification algorithm effectively addresses the aforementioned issue. Empirical results underscore the potency of our proposed classification algorithm, showcasing an impressive 96.16% classification accuracy rate compared to conventional methodologies. However, certain limitations warrant acknowledgment. First, this paper’s orchid image classification encompasses a subset of only 12 common orchid species. Given the diverse array of orchid varieties in practical scenarios, future classification efforts should endeavor to encompass a broader spectrum of species. Second, while our proposed approach yields substantial improvements, the accuracy of orchid species classification could benefit from even more advanced classification techniques in the future.

Acknowledgment

The work was supported by the Guangdong Basic and Applied Basic Research Foundation (No. 2021A1515011514 and No. 2023A1515012194); the Laboratory of Lingnan Modern Agriculture Project (NT2021009); the National Natural Science Foundation of China (No. 61602187); the Sino-Singapore International Joint Research Institute Project (No. 206-A021006); The Guangdong Basic and Applied Basic Research Foundation (No. 2021A1515010923); The 111 Project (D18019) and the Key Area Research and Development Program of Guangdong Province (2019B020214003).

Conflict of Interest

The authors declare that they have no conflict of interest.

Data Availability

Data sharing is not applicable to this paper as no datasets were generated or analyzed during this study.

ORCID

Jianhua Wang  https://orcid.org/0000-0001-9587-2845

Haozhan Wang  https://orcid.org/0000-0001-9193-1158