BiAgriNet: Binarized Knowledge-Distilled Network for Real-Time Semantic Segmentation in Agriculture

Semantic segmentation is critical for agricultural applications such as crop-weed classification and precision farming. However, real-time segmentation in resource-constrained environments requires lightweight and efficient networks. This paper introduces BiAgriNet, a novel agricultural segmentation framework that combines full weight-and-activation binarization with a teacher–student knowledge-distillation paradigm. The proposed student network employs a 1bit ResNet18 encoder and a grouped dilated Atrous Spatial Pyramid Pooling bottleneck, enabling multiscale context capture while maintaining a lightweight design. Trained using advanced teacher-student knowledge distillation with ResNet18 +DeepLabV3 as the teacher, BiAgriNet achieves 85.6% mIoU, with a memory footprint of only 0.8 MB, which is 27× smaller than DeepLabv3 while maintaining competitive accuracy. With an inference speed of 180 FPS, nearly 5× faster than FCN8s (37 FPS), BiAgriNet demonstrates its practicality for real-time precision agriculture on embedded systems, offering a compelling balance between efficiency and performance.