Enhancing weed detection through knowledge distillation and attention mechanism

Weeds pose a significant challenge in agriculture by competing with crops for essential resources, leading to reduced yields. To address this issue, researchers have increasingly adopted advanced machine learning techniques. Recently, Vision Transformers (ViT) have demonstrated remarkable success in various computer vision tasks, making their application to weed classification, detection, and segmentation more advantageous compared to traditional Convolutional Neural Networks (CNNs) due to their self-attention mechanism. However, the deployment of these models in agricultural robotics is hindered by resource limitations. Key challenges include high training costs, the absence of inductive biases, the extensive volume of data required for training, model size, and runtime memory constraints. This study proposes a knowledge distillation-based method for optimizing the ViT model. The approach aims to enhance the ViT model architecture while maintaining its performance for weed detection. To facilitate the training of the compacted ViT student model and enable parameter sharing and local receptive fields, knowledge was distilled from ResNet-50, which serves as the teacher model. Experimental results demonstrate significant enhancements and improvements in the student model, achieving a mean Average Precision (mAP) of 83.47%. Additionally, the model exhibits minimal computational expense, with only 5.7 million parameters. The proposed knowledge distillation framework successfully addresses the computational constraints associated with ViT deployment in agricultural robotics while preserving detection accuracy for weed detection applications.