基于改进YOLO v5s的轻量化植物识别模型研究

为方便调查宁夏全区荒漠草原植物种类及其分布，需对植物识别方法进行研究。针对YOLO v5s模型参数量大，对复杂背景下的植物不易识别等问题，提出一种复杂背景下植物目标识别轻量化模型YOLO v5s-CBD。改进模型YOLO v5s-CBD在特征提取网络中引入带有Transformer模块的主干网络BoTNet(Bottleneck transformer network),使卷积和自注意力相结合，提高模型的感受野；同时在特征提取网络融入坐标注意力(Coordinate attention, CA),有效捕获通道和位置的关系，提高模型的特征提取能力；引入SIoU函数计算回归损失，解决预测框与真实框不匹配问题；使用深度可分离卷积(Depthwise separable convolution, DSC)减小模型内存占用量。实验结果表明，YOLO v5s-CBD模型在单块Nvidia GTX A5000 GPU单幅图像推理时间仅为8 ms,模型内存占用量为8.9 MB,精确率P为95.1%,召回率R为92.9%,综合评价指标F1值为94.0%,平均精度均值(mAP)为95.7%,在VOC数据集...

Lightweight Plant Recognition Model Based on Improved YOLO v5s

In ordered to facilitate the investigation of desert grassland plant species and their distribution in the whole Ningxia region, plant identification methods need to be studied. To address the problems of large number of parameters in YOLO v5s model, it is not easy to recognize plants in complex backgrounds, and a lightweight model of plant target recognition in complex backgrounds, YOLO v5s-CBD, was proposed. The improved model YOLO v5s-CBD introduced the BoTNet with Transformer module into the feature extraction network, to combine convolution and self-attention to improve the feeling field of the model. At the same time, coordinate attention was incorporated into the feature extraction network to effectively capture the relationship between channel and position and improve the feature extraction ability of the model. In terms of loss calculation, the SIoU function was introduced to calculate the regression loss to solve the problem of mismatch between the prediction box and the real box. Using depthwise separable convolution to reduce model volume. The experimental results showed that the model YOLO v5s-CBD infers a single image in only 8ms, a model volume of 8.9MB, a precision of 95.1%, a recall of 92.9%, a F1 value of 94.0%, and a mean average precision of 95.7% in a single Nvidia GTX A5000 GPU, and a mean average precision of 80.09% in the VOC dataset. Compared with YOLO v3-tiny, YOLO v4-tiny and YOLO v5s, the improved models reduced model volume and improved mean average precision. The model YOLO v5s-CBD had good robustness in both public dataset and Ningxia desert grassland plant dataset, faster inference speed and easy to deploy. It was applied in Ningxia desert grassland mobile plant image recognition APP and fixed ecological information observation platform, which can be used to investigate the species and distribution of desert grassland plants in the whole region of Ningxia, and long-term observation and tracking of Dashuikeng, Huangjichang, Mahuangshan and other places, Yanchi County, Ningxia.