Bunet: An effective and efficient segmentation method based on bilateral encoder-decoder structure for rapid detection of apple tree branches

Automatic apple harvesting robots have received much research attention in recent years to lower harvesting costs. A fundamental problem for harvesting robots is how to quickly and accurately detect branches to avoid collisions with limited hardware resources. In this paper, we propose a lightweight, high-accurate and real-time semantic segmentation network, Bilateral U-shape Network (BUNet), to segment apple tree branches. The BUNet consists mainly of a U-shaped detail branch and a U-shaped semantic branch, the former for capturing spatial details and the latter for supplementing semantic information. These two U-shape branches complement each other, keeping the high accuracy of the Encoder-decoder Backbone while maintaining the efficiency and effectiveness of the Two-pathway Backbone. In addition, a Simplified Attention Fusion Module (SAFM) is proposed to effectively fuse different levels of information from two branches for pixel-by-pixel prediction. Experimental results show (on our own constructed dataset) that BUNet achieves the highest Intersection over Union (IoU) and F1-score of 75.96% and 86.34%, respectively, with minimum parameters of 0.93M and 11.94G Floating-point of Operations (FLOPs) in branch segmentation. Meanwhile, BUNet achieves a speed of 110.32 Frames Per Second (FPS) with input image size of 1280x720 pixels. These results confirm that the proposed method can effectively detect the branches and can, therefore, be used to plan an obstacle avoidance path for harvesting robots.