基于动态可变形卷积与注意力融合的密集桑蚕分割算法

Dense Silkworm Segmentation Algorithm Based on Dynamic Deformable Convolution and Attention Fusion

摘要:

实现桑蚕自动化计数是蚕桑养殖智能化管理的关键环节。然而，由于桑蚕个体尺寸小、形态多变且在密集养殖环境下存在严重遮挡，人工肉眼计数的方法难以实现精确计数且费时费力，导致无法根据桑蚕数量精准投放桑叶和石灰粉以确保其精准进食量和消毒环境,桑叶投放过少影响桑蚕生长，过多则造成浪费与环境污染，石灰粉用量不足则消毒效果有限，过量又可能损害蚕体健康。为解决密集桑蚕计数困难并提高桑蚕检测与分割的准确性，本文提出了一种基于YOLO 11n改进的实例分割算法DLC-YOLO，首先，在主干网络设计了DysnakeConv与C3k2融合结构，通过动态可变形卷积核沿桑蚕轮廓进行自适应形变，增强对桑蚕细长弯曲形态的特征提取能力。其次，在颈部网络引入了基于CGAfusion的内容引导注意力融合机制改变颈部原本的特征融合方式，在检测头输入端融合主干网络与特征金字塔的多尺度跨层特征，有效提升了多尺度特征的融合效率。最后，设计了LSCSBD-Seg的分割检测头，通过局部语义约束和边界细化模块，有效改善了密集桑蚕个体的边缘分割精度。为验证性能，构建了密集桑蚕实例分割数据集，涵盖不同养殖密度和光照条件。实验结果表明，DLC-YOLO算法与YOLO 11n算法相比，检测精度mAP50、mAP50:95分别提高了2.3、5.1个百分点，分割精度mAP50、mAP50:95分别提高3.0、2.9个百分点，均优于原始YOLO 11n模型。

Abstract:

Automated counting of silkworms is crucial for intelligent sericulture management. However, due to small individual size, morphological variations, and severe occlusion in dense farming environments. This impeded precise distribution of mulberry leaves and lime powder for optimal feeding and disinfection. To address dense silkworm counting challenges and enhance detection and segmentation accuracy, DLC-YOLO, an improved instance segmentation algorithm was proposed based on YOLO 11n. Firstly, in the backbone network, a DysnakeConv and C3k2 fused structure was designed, utilizing dynamic deformable convolution kernels to adaptively trace silkworm contours, strengthening feature extraction for slender, curved morphologies. Secondly, in the neck network, a CGAfusion content-guided attention mechanism replaced the original feature fusion, integrating multi-scale cross-layer features from the backbone and feature pyramid at the detection head input. This significantly improved multi-scale feature fusion efficiency, particularly for dense small targets. Finally, an LSCSBD-Seg segmentation head incorporated local semantic constraints and boundary refinement modules to enhance edge segmentation precision in dense clusters. A dedicated dense silkworm instance segmentation dataset covering varied rearing densities and lighting conditions was constructed for validation. Experimental results demonstrated that DLC-YOLO outperformed YOLO 11n, with detection precision (mAP50, mAP50:95) increasing by 2.3 and 5.1 percentage points, and segmentation precision (mAP50, mAP50:95) improving by 3.0 and 2.9 percentage points, respectively.