Hyperparameter optimization of apple leaf dataset for the disease recognition based on the YOLOv8

Abstract

Apple leaf diseases have a major influence on apple productivity and quality, demanding a precise and efficient recognition system. Using the YOLOv8 family of object detection models, we created a disease recognition model for the apple leaf dataset in this study. The developed model was fine-tuned extensively by hyperparameter optimization to identify the best variant for practical deployment. Firstly, fine-tuning with different YOLOv8 series was conducted on an apple leaf dataset including various types of images. Among them, the YOLOv8s demonstrated the best balance with a fitness of 0.97171, a precision of 0.97082, a recall of 0.96837, a [email protected] of 0.98016, and an image processing speed of 1.58 ms. Further hyperparameter optimization was conducted using the One-Factor-At-a-Time (OFAT) and Random Search (RS) methods. In this case, the optimal settings determined as per the OFAT method were a batch size of 48, a learning rate of 0.01, a weight decay of 0.0005, a momentum of 0.963, and 200 epochs. These settings were adopted as the baseline for RS. RS then searched for 50 additional configurations; the best configuration, C34 (batch size of 48, learning rate of 0.0137, momentum of 0.9433, and weight decay of 0.0009), achieved a fitness score of 0.97688, a precision of 0.97797, a recall of 0.97295, and a [email protected] of 0.98257. The correlation analysis showed that learning rate and momentum significantly impacted the performance of the models. Overall, the C34 model demonstrates high accuracy, rapid processing speed, and robustness suitable for training real-time, large-scale apple leaf disease recognition.