Shengnan Gao, Xiaoshun Li, and Yingying Liu
Agricultural robots, semantic segmentation, semantic recognition, deep learning, attention mechanism
The paper presents a deep learning-driven approach for semantic segmentation and recognition in agricultural robotics, enhancing the robots’ scene comprehension and operational efficiency. It addresses challenges in complex farm environments by introducing an innovative model that merges high-quality UAV-captured agricultural imagery with an enhanced encoder–decoder structure. This integration facilitates fine-grained image segmentation and introduces an attention mechanism plus a semantic recognition module, thereby bolstering the robot’s capability in crop identification, growth monitoring, and disease detection. To ensure the robustness of the model, a comprehensive dataset was meticulously compiled and balanced through extensive data gathering and preprocessing. The model design innovates by incorporating an advanced feature extractor with null space pyramid pooling and an attention mechanism. This design augments multi- scale feature representation and regional focus, tackling the varied and intricate nature of agricultural landscapes efficiently. Optimised with the Adam optimiser and trained using cross-entropy loss, the model underwent a meticulous training regimen to refine its performance. Evaluation outcomes highlight its excellence across key metrics: accuracy, recall, F1 score, and IoU, with additional gains observed post-application of test-time augmentation. These results affirm the method’s efficacy and practical utility in advancing agricultural robotics. Consequently, this research significantly contributes to the smart evolution of agricultural machinery and charts a promising trajectory for future automation and precision agriculture endeavors.
Important Links:
Go Back
