基于自适应卷积注意力机制的茶叶嫩芽分割与采摘点定位方法

Selective Kernel Attention Mechanism-based Method for Tea Buds Segmentation and Picking Point Localization

摘要:

为解决自然环境下茶叶嫩芽密集分布、多尺度特征以及小目标导致的识别与定位难题，本文提出融合改进注意力机制的茶叶嫩芽实例分割模型，并基于轮廓最低点查找算法（Contour bottom search algorithm, CBS）精准定位茶芽采摘点。首先，在YOLO v8n-seg主干网络引入改进的自适应卷积注意力机制（Selective kernel attention mechanism, SKAM），将空间金字塔快速池化（Spatial pyramid pooling-fast, SPPF）模块替换为感受野模块(Receptive field block, RFB)，通过扩展感受野，增强模型对多尺度茶芽特征的捕获能力。其次，在颈部网络设计浅层特征融合结构，采用动态上采样算子（Up-sampling by dynamic sampling, Dysample）保留小目标细节，提高密集嫩芽分割精度。最后，基于mask轮廓分析提出最低点定位算法。试验结果表明，本文模型边界框检测指标（准确率93.9%、召回率87.2%、平均精度均值(mAP50)93.6%）较基准模型分别提升3.7、6.5、4.0个百分点；掩膜分割指标（90.9%、84.1%、88.6%）相应提升4.9、7.4、6.0个百分点。在密集与尺度多变场景下，提出的茶芽采摘点定位算法平均定位准确率达到76.4%，验证了本文方法能够稳定识别密集多尺度茶芽并精确定位采摘点，为智能采茶装备开发提供了理论依据。

Abstract:

Aiming to address the challenges of recognition and localization caused by the dense distribution, multi-scale features, and small target characteristics of tea buds in natural environments, a tea bud instance segmentation model that incorporated an improved attention mechanism was proposed. Additionally, a contour bottom search algorithm (CBS) was employed to accurately locate the picking points of tea buds based on the analysis of the contour's lowest point. Firstly, an improved selective kernel attention mechanism (SKAM) was introduced into the backbone network of YOLO v8n-seg. The spatial pyramid pooling-fast (SPPF) module was replaced with the receptive field block (RFB) module. By expanding the receptive field, the model's ability to capture multi-scale features of tea buds was enhanced. Secondly, a shallow feature fusion structure was designed in the neck network. The up-sampling by dynamic sampling (Dysample) operator was adopted to preserve the details of small targets, thereby improving the segmentation accuracy of densely distributed tea buds. Finally, a lowest point localization algorithm was proposed based on the analysis of the mask contour. The experimental results showed that the bounding box detection indicators of the proposed model (precision of 93.9%, recall of 87.2%, and mAP50 of 93.6%) were improved by 3.7, 6.5, and 4.0 percentage points, respectively compared with that of the baseline model. The mask segmentation indicators (90.9%, 84.1%, and 88.6%) were correspondingly improved by 4.9, 7.4, and 6.0 percentage points. In scenarios with dense distribution and varying scales, the average localization accuracy of the proposed tea bud picking point localization algorithm reached 76.4%. This verified that the method can stably identify dense tea buds of multiple scales and accurately locate the picking points, providing a theoretical basis for the development of intelligent tea picking equipment.