Research on sugarcane stem counting method based on improved RT-DETR-seg.

Introduction: Sugarcane is an important sugar crop. In the life cycle management of sugarcane, counting serves as the core foundational data for yield estimation, precise water and fertilizer regulation, and the optimization of intelligent harvesting equipment. Therefore, accurate counting of sugarcane in the field facilitates yield prediction and fine-grained management of water and fertilizer resources. Methods: To address the issue of instance feature fragmentation caused by the elongated shape of sugarcane and complex occlusions, this paper presents an improved RT-DETR-seg (Real-Time Detection Transformer instance segmentation) model. The specific improvements include: adding asymmetric strip convolution (ASCs) to the backbone network to enhance the perception of elongated sugarcane targets through vertical feature aggregation, while also suppressing background noise and embedding a coordinate attention (CA) mechanism into the hybrid encoder to leverage long-range spatial dependencies, thereby bridging occluded regions and establishing feature associations to effectively mitigate repeated counts caused by steam leaf truncation. Furthermore, a joint loss function named the moment-gradient regularized (MGR) loss is designed, which integrates spatial moments and gradient field constraints and incorporates an aspect ratio penalty and edge smoothness constraints to significantly increase the segmentation accuracy under strong lighting and mottled shadow conditions. Results: Ablation experiments were conducted to individually verify the performance improvements associated with the three major enhancement modules: the ASC, CA mechanism, and MGR loss function. Compared with mainstream single-stage instance segmentation algorithms, the detection performance and counting advantages of the improved model in sugarcane scenes with dense occlusions are verified. The experimental results revealed that the improved model achieved a bounding box detection accuracy of 91.03% and a mask segmentation accuracy of 81.04% on the self-constructed multi-scenario sugarcane dataset, whereas the mean absolute error (MAE) of the counting results decreased significantly from 1.09 with the baseline model to 0.28 with the improved model. Discussion: This study presents a sugarcane counting method based on an improved RT-DETR-seg model to address the challenge of occlusion in complex sugarcane field environments, ultimately enabling high-precision, high-throughput automated monitoring of slender economic crops such as sugarcane. The enhanced algorithm balances robust perception and real-time inference under complicated field conditions, providing a solid algorithmic foundation for the development and deployment of edge-side counting systems.