LED-Net: A lightweight and efficient dual-branch convolutional neural network for high-performance fruit tree branches semantic segmentation on mobile devices

In response to the global labor shortage, fruit-picking robot technology has emerged as a critical solution for enhancing agricultural efficiency. Precise detection of fruit tree branches is essential for collision avoidance and optimal operation. However, the inconsistent performance of existing algorithms limits their application on edge devices. To address this, we propose LED-Net, a lightweight, efficient, dual-branch convolutional neural network designed for high-performance branch semantic segmentation in resource-constrained mobile environments. LED-Net is built on the Cascaded Efficient Spatial Pyramid Block (CESPB) and leverages dilated convolution in both its context and spatial branches, harnessing the advantages of dilated convolution while preserving detailed information. The network introduces two novel components: the Global Enhancement Transformer Block (GETB) and the Spatial Edge Attention Module (SEAM), which aggregate context information and enhance spatial edge features, respectively. To evaluate the proposed method, we constructed a semantic segmentation dataset of fruit tree branches during the apple maturity stage and compared LED-Net with eight state-of-the-art lightweight networks, including DDRNet and PIDNet. Experimental results demonstrate that LED-Net achieves the lowest parameter count of 1.661M and FLOPs of 9.206G. In the branch segmentation task, LED-Net achieves the highest IoU of 81.46 %, Accuracy of 90.13 %, and F1-score of 89.78 %. For 1280×720 resolution images, the network achieves a segmentation speed of 177.49 frames per second on an RTX 3090 GPU, meeting real-time processing requirements. In conclusion, LED-Net achieves a balance between precision, computational cost, parameter efficiency, and real-time performance, showcasing its exceptional potential for deployment in resource-constrained mobile harvesting robots. Source code and Dataset are available at https://github.com/ly27253/LED-Net.