AgriGAN: unpaired image dehazing via a cycle-consistent generative adversarial network for the agricultural plant phenotype.

Abstract

Artificially extracted agricultural phenotype information exhibits high subjectivity and low accuracy, while the utilization of image extraction information is susceptible to interference from haze. Furthermore, the effectiveness of the agricultural image dehazing method used for extracting such information is limited due to unclear texture details and color representation in the images. To address these limitations, we propose AgriGAN (unpaired image dehazing via a cycle-consistent generative adversarial network) for enhancing the dehazing performance in agricultural plant phenotyping. The algorithm incorporates an atmospheric scattering model to improve the discriminator model and employs a whole-detail consistent discrimination approach to enhance discriminator efficiency, thereby accelerating convergence towards Nash equilibrium state within the adversarial network. Finally, by training with network adversarial loss + cycle consistent loss, clear images are obtained after dehazing process. Experimental evaluations and comparative analysis were conducted to assess this algorithm's performance, demonstrating improved accuracy in dehazing agricultural images while preserving detailed texture information and mitigating color deviation issues.

Plant phenotyping relevance

農業植物フェノタイピング画像から特徴情報を抽出するための画像デヘイズ手法を開発し、性能比較・評価しており、フェノタイプ取得前処理が研究の中心である。

Code and data availability

The paper's Data availability statement explicitly commits the authors' cucumber hazy/haze-free image dataset and AgriGAN analysis code to a public GitHub repository (HZSUZJ/DLDF), which is listed in the allowed URLs. The dataset (407 haze-free / 479 hazy training images, 51/49 test images) and TensorFlow code are the纸