Artificial neural networks

The artificial neural network is a mathematical model of problem-solving inspired by the functioning of human nerves (Bishop, 1995;Haykin, 2010). The simplest model of an artificial neural network is a single-layer perceptron, which is diagrammatically presented inFig. 2.

Mathematically, the single-layer perceptron is expressed as follows:(1)${\displaystyle \sum _{x=1}^{x=n}{w}_i{\mathbf{x}}_i+b}$

wherew_i,x_i andb denote weight −i, feature −i and bias consecutively. The single perceptron receives information on attribute/feature values and assigns a weight to each feature. The final result is determined by passing the sum of the values to the activation function. Several commonly used activation functions are sigmoid, hyperbolic tangent, and rectified Linear Unit (ReLU).

Deep neural networks

The artificial neural network continues to develop. Researchers proceed to model it as a multilayer perceptron (Goodfellow, Bengio & Courville, 2016). Similar to the single-layer perceptron, the multilayer perceptron receives input in the form of a set of feature values to then be weighted and the final result is entered into the activation function. The difference between single-layer perceptron and multilayer perceptron is that there is a hidden layer that automatically adds the number of nodes and creates more complex connections. Hence, multilayer perceptron can solve any problems that cannot be solved by using single-layer perceptron.

The use of multilayer perceptron continues to grow, especially in the field of computer vision. Hardware support enables more complex image operations, such as convolution procedures. With the use of artificial neural networks that have many hidden layers and complex image operations, the era of deep neural networks that is more popularly known as deep learning begins.

Related works on deep learning

The use of deep learning in classifying a collection of images becomes more popular due to the explosion of data and the availability of the high-end computational platforms. The classification of larvae images takes the advantage of the deep learning platform. To the best of our knowledge, however, only a few studies have been proposed in this field.

The first related work is proposed inFuad et al. (2018) that utilized transfer learning technique,i.e., Google inception model (Szegedy et al., 2017), to identify the larvae ofAedes Aegypti which are commonly found inside the clean liquid place. These larvae are important to classify because it helps reduce the spread of dengue fever, especially in tropical countries. There were enough pictures,i.e., more than five hundred, used in this paper due to the assorted canvas and lightning within the liquid place. The paper mentioned that the dataset was produced by laboratory work, however, it remains unclear why the authors chose to have no data pre-processing required by the study. It is common to have data pre-processing as an essential step in deep learning especially dealing with real-world pictures, although most researchers skip the explanation in their reports. The authors claimed that the less the learning rate is the better the classification accuracy. It can be understood because there are only two classes are produced,i.e., larva or not larva. In detail, the accuracy in the paper achieved 99.98% and 99.90% for learning rate 0.1, 99.91% and 99.77% for learning rate 0.01; and 99.10% and 99.93% for learning rate 0.001. The cross-entropy error in the paper achieved 0.0021 and 0.0184 for learning rate 0.1, 0.0091 and 0.0121 for learning rate 0.01; and 0.0513 and 0.0330 for learning rate 0.001. Although the metrics of accuracy and cross-entropy errors are presented, the significance of the experiments could not be justified due to the goal of the paper is to provide informative accuracies and learning errors with three learning rates. It would be interesting to see the baseline of the accuracy performed by this paper with different kinds of the larva.

Aedes Aegypti is the most popular larva to be deeply classified from its images. Another work on this larva is proposed inAzman & Sarlan (2020) that examined whether particular water storage is the suitable place for the mosquitoes to lay their larva or not. Although it is not explicitly mentioned, it can be assumed that the paper provides two different types of larva from the mosquitoes that spread the lethal epidemic of dengue fever and three other types of larvae from different mosquitoes. These five types of the larva are sorted based on the prediction accuracy by utilizing a general convolution neural network algorithm, although there is no further explanation in detail on how to implement this algorithm with specific techniques. The result shows the unbalanced accuracies for each type of larvae that varies from 0.7% to 73%. It remains unclear why the gap between the two types of dengue fever mosquito larva and three other types is huge.

The authors inAsmai et al. (2019) also analyzedAedes Aegypti larvae. The testing was conducted from ten images ofAedes Aegypti larvae and ten images of nonAedes Aegypti larvae. Various methods of convolution neural networks were utilized, such as VGG16 (Zhang et al., 2015), VGG-19 (Zhang et al., 2015), ResNet-50 (Koonce, 2021), and InceptionV3 (Szegedy et al., 2017). The results showed that Resnet-50 outperforms other methods in terms of the implementation on mobile devices, although its performance in terms of accuracy and loss was lower than VGG-19 (Zhang et al., 2015). However, combining the mobile performance from two regular metrics, such as accuracy and loss, with other provided metrics, such as file size and training time, should be further investigated because it is not commonly used in deep learning.

A different kind of larva is analyzed inShang, Lin & Cong (2020) that addressed the issue of a low number of pictures with high-quality tags. Another issue tackled by this paper was that the picture ofZebrafish larva usually contains the fuzzy classification for ten types of the larva, such as deceased, regular, and short bottom. Hence, the deep learning is non-trivial in classifying this larva. The result showed that there was an improvement of classification performance between the proposed two-layered classification technique and other deep learning methods, such as GoogLeNet (Szegedy et al., 2015), VGG-19 (Zhang et al., 2015) and AlexNet (Krizhevsky, Sutskever & Hinton, 2012), by reproducing the previous researches with the same dataset. The improvement compared to the baseline reached 22%. The obtained accuracy mean was 91% for overall types and the obtained maximum accuracy was 100% for deceased and skin types. However, it remains unclear how to obtain these numbers since there is no exact numbers provided by the paper. Moreover, it should have been more interesting to see how the technique operates, since running two layers of classification in a single workflow is obviously better than running one layer, regardless of the image characteristics.

The latest work from aquaculture researchers inKakehi et al. (2021) presented the use of deep learning in identifying oyster larvae. The work was important for the oyster farmers to speed up the process of larva identification during the oyster growing time. The result was claimed to have high performance with more than 80% for precision, around 90% or recall, and therefore around 86% for F-score. However, the overall process was not completely automatic, since human intervention was required during the process of identification. It is interesting to note that the difficulty was due to the characteristics of oyster larva that occasionally mixed with other objects that produce several layers on the images. This paper also provided three-dimension graphs of shell height, long side, and short side to help understand the dataset. These three measures were estimated by using the coordinate system of PyTorch (Ketkar, 2017b).

Another latest work was proposed inOng, Ahmad & Majid (2021) that provided a deep learning model to predict the teemingness of the larvae of house flies. The experiments were conducted with different angles of picture shooting. The variation also includes various colors and textures. The final result of the proposed method performance was precision between 88.44% and 92.95%, recall between 88.23% and 94.10%, accuracy between 87.56% and 92.89%, and F-score between 88.08% and 93.02%. Another interesting result from this paper is that the image with green and white lighting performed the best while the images with red lighting performed the worst. However, it should be further researched whether the result from this larva remains the same as the results from other kinds of larvae. Another critical thinking on this paper is that the benefits of the classification result on house fly larva to the real world remain questionable, except for the sake of experimental laboratory. In addition, it would be more interesting to see the specific technique of deep learning used in the paper instead of explaining convolution neural networks in general.

Overall, the presented literature review shows that the use of deep learning in classifying larva images is eminent. However, none of the papers, to the best of our knowledge, handled the images ofZophobas Morio andTenebrio Molitor larvae. These larvae are important for feeding broiler chicken in emerging countries; hence, the classification of the larvae can be an additional feature for other presented larva automatic detection systems. This is also due to the fact that the detection for two different larvae by naked eyes has lower accuracy as been identified by biologists (Benzertiha et al., 2019).

Transfer learning

Several transfer learning algorithms have been proposed, including VGG-19, Resnet-50, and InceptionResnetV2 (Goodfellow, Bengio & Courville, 2016;Simonyan & Zisserman, 2014;He et al., 2016), to enhance deep learning with less amount but significant enough dataset. Instead of using deep learning from scratch for the two specific larvae, transfer learning is more suitable because the previous works have been proposed for various larvae.

VGG-19 is a more advanced development of VGG-16 (Zhang et al., 2015). It has 19 layers, which is quite a contrast compared to Resnet-50 that has 50 layers, and InceptionResnetV2 which that 164 layers. The architecture for VGG-19 can be seen inFig. 3.

For practical purposes, two pre-trained models are selected in the experiment, namely, VGG-19 and Inception v3. The first model has lower layers than the second one has.

Since there is no previous research related to the classification ofZophobas Morio andTenebrio Molitor larvae classification, two traditional classification algorithms, specifically,k-NN and SVM, are employed (Géron, 2019). The results from these two algorithms are used as a reference for the baseline. In other words, the result of the pre-trained models is compared to the traditional classifiers.

Datasets

The datasets are obtained using a standard mobile phone camera commonly available on the market. In this experiment, we use a mobile phone with specifications: Memory 6 GB, Processor Exynos 9611 Octa-core, Camera: Quad camera 48 MP+12 MP ultrawide+5 MP macro+5 MP depth. The images were taken with the use of flash. These specifications are important for reproducing the research in the future.

Larva image samples were obtained from the animal market. The captured larvae are expected to be alive, although several larvae are generally mixed with skins and dead larvae. The total number of samples was 640 equally distributed for eachZophobas Morio andTenebrio Molitor. Each sample was placed on a white paper sheet and subsequently captured with a distance of approximately 10 centimeters from the larvae as previously displayed inFig. 1. Since the images are designed with homogeneous background,i.e., white paper, and the size of the larva is various, it is not necessary to perform an image augmentation on the dataset.

The dataset is divided into 10 folds during the experiment. Each fold is chosen as a testing set while the remaining folds are utilized as a training set. Moreover, the distribution of each class in the fold is considered identical for all folds. This split is known as stratified cross-validation (CV) (Breiman et al., 2017). We summarize the distribution of the dataset inTable 2. Each class consists of 320 images which is divided into 10 folds. One fold containing 32 images per class is assigned as data testing while the remaining ones are as data training. It is also important to note that each fold is assigned as testing data only once. After 10 folds are iteratively evaluated, the testing set will completely contain 320 images.

Model implementation

We carried out experiments for the above datasets using deep learning. The model used is taken from another model that has been tested and has the same objective characteristics, namely image classification. As stated in section 2, two transfer learning methods are employed in building the classification model, namely VGG-19 (Zhang et al., 2015) and Inception v3 (Szegedy et al., 2016).

The transfer learning using VGG-19 is implemented in Python programming language and run on Google Colab (Paper, 2021). The library for implementing the transfer learning is Keras (Gulli & Pal, 2017) which is imported from Tensorflow (Abadi et al., 2016). In VGG-19, the number of layers has been defined as 19. The number of layers is fixed for this model. However, there are still several parameters that we set based on the number of datasets used. First, the trainable parameter is set to false because the network will not be trained again. Then, the last classification layer is set to 2 because in this experiment there are only two labels, namelyZophobas Morio andTenebrio Molitor. On the initial layer, the image is adjusted to the size of (224 × 224 × 3). The batch size is set to 32 while the epoch is 25.

Similarly, Inception v3 is implemented in Python and executed on Google Colab. Inception v3 has 48 layers and the trainable parameter is set to false. In order to provide a fair comparison, the batch and the epoch parameters are set to the same setting as VGG-19.

Experimental results

The testing data are utilized in the final evaluation to determine the performance of the built model. As previously mentioned, two traditional classification algorithms, namelyk-NN and SVM, are selected. The results serve as a baseline because no previous study exists on the classification of the two larvae to the best of our knowledge.

The detailed results fork-NN with variousk values are listed inTable 3. The image size is scaled down to 224 × 224. The value ofk is set at an odd number since the classification is binary. The results ofk-NN and SVM are used as the baseline. The best results on a particulark ink-NN are chosen and subsequently considered a reliable representation ofk-NN in general.

Similar tok-NN, SVM is performed on resized images,i.e., 224 ×224. The selected kernel is linear due to its simplicity. The results of SVM, as well ask-NN, are subsequently compared to the two transfer learning algorithms used in this research,i.e., VGG-19 and Inception v3. The comparison of accuracy per fold betweenk-NN, SVM, VGG-19, and Inception v3 is presented inTable 4. Please note that the assignment ofk = 1 for thek-NN is obtained from the previous experimental result as listed inTable 3