基于轻量级检测网络与显著性分割网络级联的鲜食甘薯轮廓提取方法

Cascaded Lightweight Detection and Saliency Segmentation for Field Sweet-potato Contour Extraction

摘要:

鲜食甘薯是海南省重要的经济作物，采用机器人进行捡拾作业是提高其收获效率的有效手段。然而，在复杂的田间环境下，目标轮廓的准确提取仍存在鲁棒性较差和分割精度不高的问题。为此，提出了一种级联改进的检测 — 分割方法：以轻量化的 YOLO 11n 作为检测骨干，并与 BASNet (Boundary-aware salient Net) 级联，旨在提高田间轮廓识别的准确性与稳定性。首先用轻量化的 FasterNet 替换 YOLO11n 的原有主干，显著减少模型参数并提升推理速度;其次，在主干网络中引入多尺度、可变扩张率的 LSKA (Large separable kernel attention) 模块，以灵活扩大感受野，增强对小目标及被植被遮挡目标的响应能力;然后，在中小目标检测头之前插入 SENetV2 (Squeeze-and-excitation version2) 通道注意力模块，通过全局特征重标定进一步提升复杂背景下的目标信号提取能力;最后，将 BASNet 显著性分割网络级联至检测模块，对检测框区域进行像素级精细分割，有效剔除背景干扰。对比试验表明，改进后的 FLS-YOLO11n 在显著降低模型体积与参数量、提高推理帧率的同时，实现了召回率提升 4.1 个百分点、平均精度提升 1.8 个百分点。级联至 BASNet 后，分割的平均绝对误差 (MAE) 约降低 40%, 而最大 F 测度、最大结构相似性测度、最大增强对齐测度分别提升约 3.37%、3.31% 与 4.19%。最后，在机器人上进行试验，抓取成功率达到 90.6%。试验结果表明，该网络能够在复杂收获环境中实现较高的轮廓识别精度，为甘薯捡拾机器人在类似场景的工程化应用提供可行的技术路径。

Abstract:

Fresh sweet potato is an important economic crop in Hainan Province, and the use of robots for picking operations is an effective means to improve harvesting efficiency. However, in complex field environments, accurate extraction of target contours still suffers from poor robustness and low segmentation accuracy. To address this issue, a cascaded improved detection-segmentation method was proposed: lightweight YOLO 11n was used as the detection backbone, cascaded with boundary-aware salient net (BASNet) to enhance the accuracy and stability of contour recognition in the field. Firstly, the original backbone of YOLO 11n was replaced with lightweight FasterNet, significantly reducing model parameters and improving inference speed. Secondly, a large separable kernel attention (LSKA) module with multi-scale and variable dilation rates was introduced into the backbone network to flexibly expand the receptive field, enhancing the response capability to small targets and targets occluded by vegetation. Then, the squeeze-and-excitation version 2 (SENetV2) channel attention module was inserted before the small and medium target detection heads to further improve target signal extraction capability in complex backgrounds through global feature recalibration. Finally, the detector's bounding boxes were passed to BASNet for pixel-level salient segmentation, removing background noise and refining object contours. On benchmark comparisons, the improved FLS-YOLO 11n achieved a 4.1 percentage points increase in recall and a 1.8 percentage points gain in mAP while substantially reducing model volume and parameter count and improving inference FPS. After cascading with BASNet, segmentation MAE was reduced by roughly 40%, and max Fβ, maxEφ, and max Sm was increase by 3.37%, 3.31% and 4.19%, respectively. Field trials on a harvesting robot produced a 90.6% grasp success rate. Results demonstrated that the proposed pipeline attained high contour recognition accuracy in complex harvesting environments, offering a practical technical path for engineering deployment of sweet potato picking robots.