SMD: A Method for Semi-Supervised Maize Leaf Disease Detection

Chengcheng Chen, Weijia Zhang, Ronghao Fu, Tairan Wang, Tiantian Pang, Jiehong Wu, Zhiyong Zheng, Xianchang Wang

Abstract: Maize leaf diseases are critical determinants of maize yield and quality. With the decrease of population and arable land resources, deep learning-based machine vision methods provide obvious advantages in recognizing and detecting maize leaf diseases in precision and cost, and it has become a hotspot of current research. However, there are many types of maize leaf diseases, over 90 types of maize leaf diseases can be found in the worldwide. The diagnosis of one disease or several diseases cannot fulfill the needs of practical applications. In addition, deep learning model training requires a large amount of datasets and label, and it is very difficult to obtain comprehensive dataset samples and labels with existing supervised learning methods. A method that relies on limited data samples and label is desperately needed. In this paper, we propose a novel method for detecting maize leaf diseases using semi-supervised learning techniques, which incorporates a soft teacher model into the Faster R-CNN network framework. By leveraging a small set of labeled data, the model generates pseudo-labels for a large pool of unlabeled data. This innovative strategy significantly enhances the detection performance of the model, resulting in more accurate and reliable disease detection outcomes. Results of the experiment in the PlantVillage public dataset indicated that the method achieved a superior detection result under different labeling percentages of data. In particular, its performance reaches 57.2% when the labeled data is 50%, which is a 25.1% improvement compared to the supervised learning method.