基于YOLO 26n-seg和沙漏神经网络模型的草莓相对轮廓预测

Strawberry Relative Contour Prediction Based on YOLO 26n-seg and Hourglass Network Model

摘要:

针对非结构化农业环境下草莓果实常受枝叶、果柄及相邻果实遮挡,导致机器视觉系统难以获取完整几何轮廓,进而影响采摘机器人定位精度与抓取成功率的问题,本研究提出一种融合改进YOLO 26实例分割与沙漏网络的草莓相对轮廓级联预测方法。首先采用最新的YOLO 26n-seg模型作为视觉感知主干,通过引入轻量化GhostNetV2重构主干网络并嵌入注意力机制,在复杂背景下精准分割出果实可见区域的掩码。其次针对遮挡区域的形状推断难题,结合主成分分析旋转校正算法对轮廓特征进行标准化处理,并构建基于Hourglass Network的相对轮廓预测模型。该模型利用其对称的自底向上与自顶向下结构,有效整合果实可见部分的局部细节与全局语义特征,建立从可见区域特征到整体轮廓坐标的映射关系。试验结果表明,改进分割模型平均精度均值达到91.3%,参数量降至3×106;轮廓预测模型在4个遮挡方向的交叉测试中,测试集平均R2稳定在0.6500~0.6987,平均绝对误差（MAE）降至9像素左右,显著优于未校正及传统MLP（Multi-layer perceptron）模型。实现了对不同遮挡程度草莓果实精准形态补全,为果园智能采摘机器人避障规划与无损抓取提供了可靠的空间几何信息。

Abstract:

Addressing the issue that in unstructured agricultural environments,strawberry fruits are often obscured by leaves,stems,and adjacent fruits,making it difficult for machine vision systems to obtain complete geometric contours,which in turn affects the positioning accuracy and grasping success rate of picking robots,a relative contour cascade prediction method was proposed for strawberries that integrated an improved YOLO 26 instance segmentation with a Hourglass Network. This method first used the latest YOLO 26n-seg model as the visual perception backbone. By introducing a lightweight GhostNetV2 to reconstruct the backbone network and embedding an attention mechanism,it accurately segmented the mask of the fruit's visible area against complex backgrounds. Secondly,to address the challenge of shape inference in occluded areas,principal component analysis (PCA) rotation correction was applied to standardize contour features,and a relative contour prediction model based on the Hourglass Network was constructed. This model leveraged its symmetric bottom-up and top-down structure to effectively integrate local details of the fruit's visible parts with global semantic features,establishing a mapping from visible area features to the coordinates of the entire contour. Experimental results showed that the proposed improved segmentation model achieved an average precision (AP) of 91.3% with parameters reduced to 3×106,and the contour prediction model maintained an average R2 between 0.6500 and 0.6987 with a mean absolute error (MAE) of around 9 pixels across different occlusion directions on the test set,significantly outperforming uncorrected and traditional multi-layer perceptron (MLP) models. This method enabled accurate shape completion of strawberry fruits under various degrees of occlusion,providing reliable spatial geometric information for obstacle avoidance planning and non-destructive grasping by intelligent orchard picking robots.