RiceSeg-YOLO: A Multi-scale Attention-Based Instance Segmentation Model for Rice Leaf Rolling in Complex Paddy Environments

Yili Wang, Yiling Ding, Chengsheng Yuan, Qiulin Wu

Abstract: Precise instance segmentation of rice leaf rolling caused by Cnaphalocrocis medinalis is a critical task for intelligent agricultural monitoring. However, in complex paddy environments, this task is challenged by indistinct fine-grained features, target-background confusion, and scale inconsistencies across leaf structures. To address these issues, a novel instance segmentation model, RiceSeg-YOLO, is proposed to effectively identify rice leaf rolling. Firstly, a specialized dataset containing 10,570 augmented images is built to validate superior performance of the proposed scheme. Secondly, the Convolutional Additive Self-attention Residual Block (CASRB) and the Poly Kernel Inception Spatial Pyramid Pooling Network (PKI-SPPN) are integrated into the backbone network to enhance fine-grained feature extraction and robust multi-scale context aggregation for better target-background separation. Lastly, a novel neck network, the Multi-Scale Ghost Feature Pyramid Network (MSGFPN), is designed for effective multi-level feature fusion and multi-scale representation. Experimental results demonstrate that, compared to the baseline YOLOv8s-seg, RiceSeg-YOLO achieves a 9.3% improvement in box mAP50:95 and an 8.1% increase in mask mAP50:95. Moreover, the proposed scheme significantly outperforms YOLOv8m-seg across all evaluation metrics, while simultaneously reducing parameters by 5M and lowering computational overhead by 65.3 GFLOPs, demonstrating its efficiency and effectiveness.