Intelligent identification of rice leaf diseases via improved faster-RCNN with multi-feature scale fusion.

Abstract

Many Artificial Intelligence and Machine Learning technologies have been applied to detect rice diseases. These approaches are either unable to identify the diseases or have a slow recognition speed. Therefore, an improved Faster-RCNN (Faster-RCNN-Pro) model is proposed to overcome these issues. First, SENet attention modules are embedded in the backbone of Faster-RCNN to enhance confidence of objects that are difficult to recognize by enhancing key image information and suppressing background information. Second, structure of the feature extraction network and RPN are improved by using multi-feature scale fusion to increase the utilization of micro-target features. Third, the quantization error introduced in the process of pooling the region of interest is then eliminated by ROI Align. Finally, a balanced L1 loss function is designed to effectively reduce the imbalance between samples with a large gradient that are difficult to learn, and samples with a small gradient that are easy to learn. The experiment results show that the improved model has a better detection accuracy and robustness in recognizing the fine features of rice leaf diseases. Therefore, the application of this model to the intelligent identification of rice leaf disease can significantly improve the accuracy and reduce the misjudgment rate.

Plant phenotyping relevance

イネ葉の病害状態を画像から識別する改良Faster-RCNNを開発・評価しており、植物病害表現型の取得手法が研究の中心である。

Code and data availability

The paper's rice leaf disease detection experiments were performed on a public Kaggle image dataset (rice blast, brown spot, hispa, healthy leaves), which the authors explicitly state is publicly available with a URL matching the allowed list. No author analysis code or trained model checkpoints are disclosed.