Self-adversarial Training and Attention for Multi-task Wheat Phenotyping

Abstract

Abstract. Phenotypic monitoring provides important data support for precision agriculture management. This study proposes a deep learning-based method to gain an accurate count of wheat ears and spikelets. The deep learning networks incorporate self-adversarial training and attention mechanism with stacked hourglass networks. Four stacked hourglass networks follow a holistic attention map to construct a generator of self-adversarial networks. The holistic attention maps enable the networks to focus on the overall consistency of the whole wheat. The discriminator of self-adversarial networks displays the same structure as the generator, which causes adversarial loss to the generator. This process improves the generator’s learning ability and prediction accuracy for occluded wheat ears. This method yields higher wheat ear count in the Annotated Crop Image Database (ACID) data set than the previous state-of-the-art algorithm. Keywords: Attention mechanism, Plant phenotype, Self-adversarial networks, Stacked hourglass.