Automatic Detection of Plant Leaf Diseases Based on Improved Yolov8 Algorithm

Abstract

In this paper, we introduce two-way weighted feature pyramid and convolutional block attention mechanism based on the single-stage target detection framework to enhance multi-scale feature fusion and highlight the lesion-related response. The experiment adopted the configuration of batch size of 8 and a total of 50 rounds of training, and carried out training and evaluation on the data division containing 1136 training images and 211 validation images. The validation set contained 769 labeled instances and covered more than ten types of pests and diseases. During the training process, the loss of bounding box, classification and distribution focus decreased overall, while the accuracy rate, recall rate and mAP index of validation set increased or fluctuated with the training progress. Ablation experiments show that adding attention module and feature pyramid structure to the baseline model can bring accuracy benefit, and the combination of the two can achieve better detection performance. At the same time, the number of parameters and calculation overhead increase, and the inference speed decrease slightly. The results show that the proposed improvement is beneficial to disease localization and identification in complex background and small target scenarios, and can provide a reference for visual monitoring in smart agriculture.

Code and data availability

The paper describes a YOLOv8+BiFPN+CBAM plant leaf disease detection model trained on a private 1347-image dataset, but contains no public dataset deposit, no author code/model release, and no availability statements or URLs for any paper-specific asset.

No evidence-backed public reproduction asset is currently recorded.