Integrating masked generative distillation and network compression to identify the severity of wheat fusarium head blight

Abstract

Fusarium head blight (FHB) is a severe disease, with implications for both crop quality and safety. The inability to accurately and rapidly determine diseases severity has resulted in increasing grain loss and the pesticide expenses. Furthermore, the complexity of many current models presents challenges in their deployment and utilization. Thus, this study introduces an improved lightweight model for efficient and rapid assessment of FHB severity. Firstly, we collected 2650 wheat images with different severities in natural environments. Second, we refined and compressed RepGhostNet, replacing the original ReLU function with LeakyReLU and using the AdamW optimizer during training to enhance model accuracy. Third, using the strategy of masked generative distillation, we further improved the accuracy of SlimRepGhostNet while ensuring model lightweight. The MGD-SlimRepGhostNet achieved an accuracy of 94.58% and a frames per second (FPS) of 152.17. This represents a 4.34% increase in accuracy and a 21.17 increase in speed compared to the original RepGhostNet. Lastly, we have designed a WeChat mini program that achieves the performance of MGD-SlimRepGhostNet in real environments, highlighting its practicality. The proposed method effectively addresses the inaccuracies and labor-intensive associated with nature of traditional visual assessment methods deployed for evaluating FHB severity in wheat, while its rapid inference capability renders it highly suitable for deployment and application on mobile devices.