Early Diagnosis of Apple Leaf Pests and Diseases of Artificial Intelligence and Inspection Robots
🎯 Project Aim
- Design a lightweight deep learning pipeline that identifies early-stage apple leaf pests and diseases with high accuracy.
- Deploy the optimized model on Jetson TX2 and integrate it with an autonomous inspection robot for in-orchard monitoring.
📓 Project Content
- Collected field images covering eight early-stage apple leaf diseases and expanded the dataset ~12× through digital augmentation to support robust training.
- Adopted Mosaic augmentation and proposed the Apple-CSP lightweight module to improve recognition of small lesion targets while keeping the network compact.
- Introduced asymmetric dilated convolution and the LAD-Inception structure, leading to LAD-Net and ALS-Net models that balance accuracy, speed, and parameter count.
- Ported the models to Jetson TX2 and collaborated on a wheeled inspection robot with remote monitoring, autonomous navigation, and cloud data synchronization.
🪨 Project Difficulties
Key challenges encountered during the project included:
- Capturing sufficient early-stage disease imagery under natural light while maintaining dataset diversity and quality after augmentation.
- Preserving detection accuracy for tiny lesions despite aggressive lightweighting required for edge deployment.
- Achieving stable hardware-software integration on the resource-constrained Jetson TX2 platform, including navigation, perception, and cloud connectivity for the inspection robot.