Journal of information and
communication convergence engineering
(J. Inf. Commun. Converg. Eng.)

Humans can recognize new objects using minimal samples. For example, children can recognize simple objects, such as apples and strawberries, with only individual pictures in the book. Cognitive ability refers to the perception of objects. Nerve cells play an essential role in cognition. Researchers hope that a machine learning model can quickly achieve few-shot learning [12] by training a large amount of data in a particular category and then training only a small number of samples for a new category arising from a downstream task. Traditional few-shot learning considers that both the training data and test data come from the same domain. If the unknown domain is included in the downstream task, the traditional few-shot learning method is unsatisfactory.

Recently, along with the rapid development of machine learning, the development of few-shot learning in the image processing field has surpassed that in the natural language processing field, and excellent few-shot learning has been focused on [13, 14]. Few-shot learning is an application of meta-learning in the field of supervised learning. Meta-learning, also known as learning-of-learning, decomposes a dataset into multiple meta-tasks in the meta-learning phase to learn the model generalization ability when categories change. There is no need to face brand new categories in the metatesting stage. The classification can be completed by changing the existing model. The definition of small-sample learning is as follows: a few-shot learning set contains multiple categories, each with multiple samples. C categories are randomly chosen from the training set of the training phase, along with K samples for each category (total CK data), and the meta-task is composed of inputs to the model support set. Extract the batches of samples from the remaining data as model prediction objects (batch sets). The model must learn to distinguish these C categories from C*K data, and these tasks are called C-way K-shot problems.

In a learning process with only a few-shot learning session, different meta-tasks are sampled for each training (episode); therefore, the training as a whole involves different combinations of categories. This mechanism allows the model to learn the common parts of various meta-tasks, such as extracting important features and comparing sample similarities, but removes the relational task-specific parts from the meta-tasks. Models trained using this learning mechanism can classify well even when faced with new and unseen meta-tasks.

Few-shot learning models can be broadly divided into model-based, metric-based, and optimization-based models. The mathematical definition of the model-based model is shown in (1).

(1)$$P_{\theta }\left(y|x,S\right)=f_{\theta }\left(x,S\right),$$

where P represents probability, S represents the sample, x denotes the input sample, y denotes the output sample, and θ denotes the weight. f represents the entropy function, also known as full conditional embedding, and can be understood as a model predicted by the model. The mathematical definition of a metric-based model is shown in (2).

(2)$$P_{\text{θ}}\left(y|x,S\right)={\displaystyle \sum _{\left(x_i,y_i\right)\in S}{k}_{\text{θ}}\left(x,x_i,S\right)y_i}.$$

where k represents the number of samples. The mathematical definition of optimization-based model is shown in (3).

(3)$$P_{\theta }\left(y|x,S\right)=f_{\theta }\left(S\right)\left(x\right).$$

The study of monkey species is conducive to researchers on the habits, population classification, and genetic characteristics of monkeys and is of great significance for studying human evolutionary history. This study established a small deep learning model to classify approximately 1,400 monkey group images of 10 species. This dataset was taken from Wikipedia’s monkey cladogram [15], named 10 monkey species. The training dataset had 1098 images, and the test dataset had 272 images. The number of images in each category was not uniform. Compared to the Dogs vs. Cats dataset [16] (approximately 25,000 images), the number of samples in this dataset was minimal. Therefore, it was challenging to establish a deep model to achieve monkey species classification and improve the classification accuracy.

When there are enough data samples, the CNN model can be competent for most image recognition and classification tasks owing to its simple structure, small number of parameters, fast data feature extraction, and high prediction accuracy. However, when dealing with image classification tasks of small datasets, although pre-training models (such as VGG and ResNet) can extract more data features, the result of the pre-trained model will increase the number of model parameters, resulting in a deeper model hierarchy. The model training time is extremely long. Therefore, we reformed and optimized the CNN model structure without significantly increasing the model parameters, mainly including the following aspects.

First, we extracted more data features by adding a small number of convolutional and pooling layers. Here, we used a 3 × 3 convolution kernel, a convolution stride size of 1 × 1, and set the convolution layer activation function to ReLU. We set the maximum pooling size to 2 × 2 and the pooling stride size to 2 × 2. Moreover, we set up a combination module of two layers of convolution and a maximum pooling layer, and added two such combination modules before the fully connected layer.

Second, in the fully connected layer, we used regularization techniques such as dropout to reduce unnecessary neurons, provide essential data features for the classifier, and improve classification accuracy. In the fully connected layer, we set the two layers dropout values to 0.5 and 0.25.

Third, we set softmax as a classification function to realize the classification output of input image feature neurons and output the classification results of 10 categories.

Finally, in the model training process, we used the categorical cross-entropy verification function as the loss function, and we used Adam [17] as the optimizer of stochastic gradient descent to optimize the training process of the model and accelerate the model convergence. The structure of the proposed model is illustrated in Fig. 1.

Figure 1. Proposed model structure.

Data enhancement, also known as data expansion, refers to the value of limited data corresponding to more data without significantly increasing the data. Data augmentation is very effective for small sample datasets. Data enhancement can be divided into two methods: supervised and unsupervised enhancement. Supervised data enhancement can be divided into single-sample data enhancement and multi-sample data enhancement methods. Unsupervised data enhancement can be divided into two directions: generating new data and learning enhancement strategies. In this study, we mainly adopted the data enhancement method of geometric transformation, as presented in Table 1. Data enhancement by such geometric transformation can increase the number of training samples to improve the model generalization ability, effectively suppress model overfitting, and improve the accuracy of model classification.

Table 1 .

Parameters	Values
Rescale	1.0/255
Rotation (º)	40
Width shift	0.2
Height shift	0.2
Zoom range	0.2
Shear range	0.2
Horizontal flip	TRUE
Fill mode	NEAREST