EK-YOLOv5: An Enhanced Real-Time Deep Learning Model for Plant Disease Detection

Balkis TEJ, Soulef BOUAAFIA, Mohamed Ali HAJJAJI, Abdellatif MTIBAA

Abstract

Tomato and pepper plants are essential crops in Tunisia’s agricultural sector, yet their yield is often compromised by various foliar diseases. This study introduces EK-YOLOv5, an enhanced deep learning model derived from the YOLOv5 architecture, which was designed to improve the detection and localization of leaf diseases in tomato and pepper plants. The proposed EK-YOLOv5 model incorporates two major innovations: replacing the original SiLU activation function with an Exponential Linear Unit (ELU) in order to accelerate convergence and improve non-linearity handling, and employing the K-means++ algorithm for generating more representative anchor boxes for diverse lesion shapes and sizes. Both the baseline YOLOv5 model and the improved EK-YOLOv5 model were trained and tested on a custom dataset, and their performance was evaluated by using standard metrics such as precision, recall, the F1-score and mAP. The experimental results reveal that EK-YOLOv5 significantly outperforms the original model, achieving a mAP of 97.4% as compared with the value of 95.8% obtained by YOLOv5, along with marked improvements with regard to recall, precision, and the detection accuracy. These findings demonstrate the potential of EK-YOLOv5 as a reliable and efficient framework for smart agriculture applications in plant disease monitoring.

Keywords

Deep Learning, EK-YOLOv5 Network, Object Detection and Localization, Plant Leaf Disease.

View full article