End-to-end plant disease detection using transformers with collaborative hybrid assignment training

초록

Plant diseases pose a significant threat to fruit production and quality if not detected and managed promptly. Precise and efficient recognition of these diseases is critical for ensuring plant health and maximizing fruit production. To tackle this issue, a range of image processing and deep learning techniques have been preferred for plant disease recognition due to their superior performance. This paper proposes an end-to-end transformer-based model that improves both the accuracy and detection rate of fruit diseases. The model is based on a state-of-theart transformer model and trained using the Collaborative Hybrid Assignment (Co-DETR) scheme. Moreover, several targeted modifications to the original model are conducted to optimize its performance. These modifications enable the model to detect six types of plant diseases with a mean average precision (mAP) of 0.89 while maintaining efficient training times. The proposed model consistently outperforms state-of-the-art detection models. In addition, the model offers interpretability through the visualization of feature discriminability scores to ensure that the prediction process is interpretable and understandable. Finally, the model demonstrates robust performance under challenging environmental conditions, such as poor lighting and image blurring, which are essential for real-world applications in disease management and precision agriculture.

키워드