Comparative Study of Computer Vision Models for Insect Pest Identification in Complex Backgrounds

Abstract

Agriculture is considered the economic basis of countries around the globe, and the development of new technologies contributes to the harvesting efficiency. Autonomous vehicles are used in farms for seeding, harvesting and tasks like pesticide application. However, one of the main issues of any plantation is insect pest and disease identification, essential for pest control and maintenance of healthy plants. This work presents and compares three methods for insect pest identification using computer vision: Deep Convolutional Neural Network (DCNN), as a baseline; Hierarchical Deep Convolutional Neural Network (HD-CNN), in order to improve prediction of similar classes; and Pixel-wise Semantic Segmentation Network (SegNet). They were tested for two kinds of culture, soybean and cotton. SegNet outperformed both approaches by a wide margin: the methods had respective accuracies of 70.14% DCNN, 74.70% HD-CNN and 93.30% SegNet.