Lightweight Convolutional Neural Networks for Pepper Diseases Detection and Classification

Abstract

In recent years, with the improvement of cultivation technology and the rapid development of logistics industry,
pepper production has broken the regional and seasonal restrictions. According to data, the global pepper
planting area in 2020 is about 1.999 million hectares, with a year-on-year growth of 3.3%, and the global pepper
production is about 39.28 million tons. In the process of pepper cultivation and management, disease is an
important factor restricting its quality improvement and yield growth, and the correct diagnosis of disease is a
necessary prerequisite for effective prevention and control of diseases. At the same time, the traditional field
diagnosis of diseases is done manually. However farmers lack professional knowledge in the identification and
prevention of diseases and pests, and lack front-line guidance from expert. When diseases and pests occur, it is
easy to cause significant economic losses. Therefore, the research of intelligent pest detection technology is very
important to control the spread of diseases and pests.With the rise of artificial intelligence technology, automatic
recognition and diagnosis of crop disease images using computer vision has become a hot research topic at home
and abroad in recent years. In addition, the depth recognition model represented by convolutional neural network
has made an important breakthrough. In this paper, the deep learning algorithm of Mobilenet-V2 convolutional
neural network is used to extract the features 20% of the 10-000 healthy and diseased pepper leaf images taken,
from the real field environment, so as to overcome the problems of complex image background and low contrast,
and to achieve end-to-end image semantic segmentation. Finally, the detection and classification of pepper images
of four types of single diseases and insect pests and multiple diseases and insect pests are achieved. Meanwhile, in
order to compare with other neural network models, this paper uses the transfer learning method to apply the
VGG16, AlexNet, GoogLeNet, ResNet as well as MobileNet-V2 neural network structures which have been trained
in image recognition in advance to the recognition of pepper diseases, so as to improve the generalization
performance of the model and effectively reduce the time and space complexity of the convolution layer.The results
show that: the average accuracy of mobilenet-v2 model is 93.05%, compared with other comparative models, it
has the characteristics of high stability, low computational complexity and low memory consumption. At the same
time, the images obtained under the actual planting conditions play an important role in the development of
automatic detection, diagnosis and classification of pepper diseases and pests. It can be applied to the early
warning of pepper disease, to solve the problem of disease prevention and control in the absence of front-line
expert.