Agronomy
2022, 12, 2096
2 of 18
extraction module for pest images. That is, the whole image is represented by handcrafted
features, including Gabor [
3
], HOG [
4
], GIST [
5
], SIFT [
6
], and SURF [
7
]; (2) Machine
learning classifiers. This includes techniques such as support vector machines [
8
–
10
], Naive
Bayes [
11
] and k-nearest neighbor (KNN) [
12
]. Such methods rely on the accurate extraction
of feature parameters, and once any wrong features are extracted, it is difficult for machine
learning classifiers to accurately identify pests with similar features.
Recently, deep learning techniques have attracted much attention from rice pest
researchers [
13
–
15
]. Liu et al. [
16
] classified rice pests by training a deep CNN with their
dataset covering 12 species and including about 5000 training samples. Alfarisy et al. [
17
]
collected 4511 training samples using a search engine and used CaffeNet [
18
] for rice
pest identification. Burhan et al. [
19
] comparatively studied the performances of five
deep learning models (Vgg16, Vgg19, ResNet50, ResNet50V2 and ResNet101V2). Overall,
these depth-feature-based efforts lack sufficient samples to optimize the large number of
hyperparameters of CNNs. To address the problem of limited pest species and samples,
Xiaoping Wu et al. [
20
] collected a large-scale IP102 dataset of eight pest-infested crops
covering 102 species including a total of 75,222 samples. They evaluated the performance of
state-of-the-art deep convolutional neural networks including AlexNet, GoogleNet, VGG16,
and ResNet50 on the IP102 dataset, where ResNet achieved the best results in all indicators.
However, the highest classification accuracy of only 49.4% indicated the challenging nature
of the IP102 dataset.
Deep learning [
21
] is a method based on big data [
22
], and we certainly hope that the
larger and higher quality the data are, the better the model generalization ability. At the
same time, we hope the data can cover various scenarios. However, it is often difficult
to cover all the scenarios when data are collected. This is where data augmentation [
23
]
presents an effective way to expand the data sample size. However, the relatively small
diversity and variability of the generated images using classical data augmentation has
facilitated research on GAN data augmentation [
24
], and the samples generated by GAN
introduce more variability and can further enrich the dataset to improve the accuracy of the
training process [
25
]. Nazki et al. [
26
] proposed a framework for the data augmentation of
plant disease datasets using GAN networks to address the problems of imbalance and lack
of samples in the dataset, and demonstrated the effectiveness of synthetic data augmenta-
tion of GAN over classical data augmentation for image classification and detection tasks.
Ding B et al. [
27
] proposed a focused attentive recurrent generative adversarial network
(ARGAN), and experimental results proved that the method outperformed state-of-the-
art methods. Haseeb Nazki et al. [
28
] introduced ARGAN, which differs from previous
approaches, by optimizing Activation Reconstruction loss (ARL) and adversarial loss to
render more visually compelling synthetic images of the dataset.
Considering the highly unbalanced nature of the IP102 [
20
] dataset and the challenge
presented by the highest accuracy rate of only 49.4%, we reorganized a dataset for rice
pest identification based on the IP102 dataset by Web crawler technique and manual
screening, named IP_RicePests. Specifically, the dataset includes over 8248 images in
14 categories with a natural long-tailed distribution of data. In addition, we expand the
IP_RicePests dataset to 14,000 images by ARGAN data augmentation technique and use
the parameters trained on ResNet [
29
], VGGNet [
30
] and MobileNet [
31
] networks based
on the public image dataset, ImageNet, as the initial values for network training to achieve
image classification in the field of rice pests.
