基于YOLO-DCL的复杂环境油茶果遮挡检测与计数研究

Camellia oleifera Fruits Occlusion Detection and Counting in Complex Environments Based on Improved YOLO-DCL

摘要:

为解决复杂环境中油茶果因遮挡造成的检测与计数难题，提出了一种基于双主干网络（Dual-backbone）和连续注意力特征融合模块（Consecutive attention feature fusion，CAFF）的检测模型。该模型结合了两种不同主干网络的优势，实现了对不同特征的高效提取。此外，设计了双输入单输出的连续注意力特征融合模块，取代了传统的拼接操作（Concat），优化了多尺度特征信息的融合策略。为了在精度与模型内存占用量之间取得平衡，采用了幻影卷积模块（Ghostconv），并去除了空间金字塔池化层（Spatial pyramid pooling fast，SPPF），加快了训练速度，减少了参数量。改进后的YOLO-DCL（YOLO dual-backbone & consecutive attention feature fusion & lightweight）模型在各类遮挡检测任务上表现优秀，平均精度均值达到92.7%，精确率为90.7%，召回率为84.9%，而模型内存占用量仅为5.7 MB。相较YOLO v8n模型分别上升4.0、8.6、2.3个百分点，内存占用量下降9.5%。该模型还具备油茶果遮挡类别的自动计数功能，可降低人工统计的劳动成本，适合在野外复杂环境中部署应用。

Abstract:

To solve the challenges of detecting and counting Camellia oleifera fruits with multiple occlusions in complex environments, a detection model was proposed based on a dual-backbone network and a consecutive attention feature fusion module (CAFF). The dual-backbone network combined the advantages of two different backbone networks to achieve efficient extraction of different features. In addition, a dual-input single-output CAFF module was designed. This CAFF module replaced the traditional concat operation and optimizes the fusion strategy for multi-scale feature information. In order to strike a balance between model precision and size, the ghost convolution (Ghostconv) module was used, and the spatial pyramid pooling fast (SPPF) layer was removed. It accelerated training time and reduced the number of parameters. The improved YOLO dual-backbone & consecutive attention feature fusion & lightweight (YOLO-DCL) model performed well on all kinds of occlusion detection tasks, with a mean average precision (mAP) of 92.7%, precision of 90.7%, and recall of 84.9%, while the model size was only 5.7 MB. Compared with the YOLO v8n model, it increased 4.0 percentage points of mAP, 8.6 percentage points of precision, and 2.3 percentage points of recall. At the same time, the model size was decreased by 9.5%. Besides, the model incorporated the ability to automatically count Camellia oleifera fruits with occlusion categories, which can reduce labor costs and improve the accuracy of yield estimation. It was very suitable for deployment in complex environments.