Semantic segmentation of tea geometrid in natural scene images using discriminative pyramid network

Abstract

A discriminative pyramid (DP) network-based method is presented in this paper, which aims to perform semantic segmentation on tea geometrid in natural scene images. The method uses image flipping, translation, mirroring, random zooming, and other techniques for training sample augmentation and adopts local histogram equalization to reduce nonuniform light influence on segmentation. Moreover, this method constructs a DP network to improve segmentation accuracy. The DP network contains two sub networks: pyramid attention and border networks. The pyramid attention network can capture the global context information of targets at different scales, increase the receptive fields to focus on small targets, and solve the problems of large shape change, small size, and difficult detection of tea geometrids. The border network can increase the differences between tea geometrids and tea tree stalks, diseased leaves, and other types of backgrounds with similar appearances by extracting and supervising semantic boundaries to help the network learn additional discriminative features. Experiment results show that the proposed method has outstanding accuracy for semantic segmentation of tea geometrid in natural scene images.