DaylilyNet: A Multi-Task Learning Method for Daylily Leaf Disease Detection.

Area of Science:

• Plant Pathology
• Computer Vision
• Agricultural Technology

Background:

• Accurate detection of daylily diseases is vital for crop yield.
• Existing detection models face challenges with complex backgrounds and small target identification, leading to reduced accuracy.
• Automated disease detection systems are needed to improve efficiency and timeliness in disease management.

Purpose of the Study:

• To develop an advanced object detection algorithm, DaylilyNet, for improved daylily disease detection.
• To enhance the accuracy and robustness of disease detection models, particularly for small diseased leaf targets.
• To evaluate the performance of DaylilyNet under varying data conditions, including information loss.

Main Methods:

• Proposed DaylilyNet, an object detection algorithm employing multi-task learning.
• Integrated a semantic segmentation loss function to focus on diseased leaf regions.
• Utilized a spatial global feature extractor and a feature alignment module to improve feature interactions and localization accuracy.
• Created 'sliding window' and 'non-sliding window' datasets to assess performance with different data processing techniques.

Main Results:

• DaylilyNet demonstrated superior performance over YOLOv5-L, achieving a 5.2% and 4.0% higher mean Average Precision (mAP@0.5) on the 'sliding window' and 'non-sliding window' datasets, respectively.
• The algorithm reduced computational parameters and time costs compared to existing models.
• DaylilyNet maintained a performance advantage even when trained on datasets with missing information.

Conclusions:

• DaylilyNet offers a significant advancement in automated daylily disease detection.
• The proposed model effectively addresses challenges posed by complex backgrounds and small targets.
• DaylilyNet's robustness to data variations makes it a promising tool for practical agricultural applications.