基于改进SAMF-YOLO v10的火龙果目标检测方法

Pitaya Fruit Target Detection Method Based on Improved SAMF-YOLO v10

摘要:

为解决火龙果果实在自然生长环境中因尺寸差异大、数量密集导致的重叠遮挡问题，本研究提出了一种基于YOLO v10改进的协同注意多特征融合目标检测模型（SAMF-YOLO v10）。在特征提取中，使用多尺度特征融合（Multi-scale feature fusion, MFC）模块代替传统的卷积模块，以增强模型对不同尺度特征的表达能力并提升细节信息的捕获效果。同时，加入协同通道空间注意力（Synergistic effects between spatial and channel attention, SCSA）模块，通过对不同空间位置及不同方向特征之间的关系进行建模，实现特征之间的协同增强，从而提高对遮挡目标和小目标的特征表达能力。结果表明，SAMF-YOLO v10模型在火龙果数据集上的检测精确率为90.18%，召回率为86.28%，平均精度均值（mAP@0.5）为93.03%。与YOLO v10模型相比，SAMF-YOLO v10在精确率、召回率和mAP@0.5上分别提升3.17、4.93、3.87个百分点;所提模型能够有效整合多层次和关键信息，从而提升火龙果图像目标检测性能。

Abstract:

Aiming to tackle the issues of mutual overlap and occlusion of pitaya fruits under complex natural growth environments, which are primarily induced by drastic scale variation and dense distribution of fruit targets, an improved collaborative attention-based multi-feature fusion object detection model named SAMF-YOLO v10 was developed on the basis of the original YOLO v10 architecture.In the feature extraction process, the multi-scale feature fusion (MFC) module was used instead of the traditional convolution module, which can enhance the model's ability to express features of different scales and improve the capture of detailed information.Meanwhile, the synergistic effects between spatial and channel attention (SCSA) module was incorporated, it can capture the relationships between features at different spatial positions and in different directions to achieve collaborative enhancement of features, thereby improving the feature expression ability for occluded targets and small targets.The results showed that the SAMF-YOLO v10 model achieved a precision of 90.18%, a recall of 86.28%, and an mAP@0.5 of 93.03% on the pitaya fruit dataset.Compared with the original YOLO v10 model, SAMF-YOLO v10 improved the precision, recall, and mAP@0.5 indicators by 3.17, 4.93, and 3.87 percentage points, respectively.The experimental results demonstrated that the newly proposed model was able to effectively integrate multi-level feature information and critical semantic content, which further enhanced and optimized the overall target detection performance for pitaya fruit images.