Abstract:  In the autonomous operation tasks of smart agricultural machinery, real-time field road defect detection is considered essential to ensure the safe operation of the machinery. However, existing road defect detection technologies have been minimally explored in the agricultural domain, and a dedicated dataset for field roads has yet to be developed. To obtain a defect detection model specific to field roads, a model was initially trained on a conventional road pothole dataset, and then followed by transfer learning using a simulated field road pothole dataset. To address the issue of reduced detection accuracy after transfer learning, an attention mechanism was introduced into the YOLOv5s network architecture to enhance the network’s precision, achieving an accuracy of 83.15% for defect detection in field road scenarios and satisfying its accuracy requirements. To verify the edge performance of the defect detection model, it was deployed onto the Jetson Nano for simulation experiments. To meet the real-time detection requirements of the field road defect detection model on edge devices, TensorRT was used to optimize and compress the model, and the pothole detection speed was improved from 396 milliseconds per frame to 157 milliseconds per frame.

Key words:  defect detection, transfer learning, deep learning, lightweight model design, edge deployment