基于YOLOv4的多目标花卉识别系统

Multi-target recognition system of flowers based on YOLOv4

[目的]花卉多目标识别定位是自动化作业的基础,大型的目标检测网络识别效果好,但由于复杂度高难以满足实时检测需求,本文提出了基于YOLOv4的轻量化目标检测算法。[方法]采用MobileNetV3替换原有的主干特征提取网络,自下而上融合网络的浅层和深层特征,简化路径聚合网络进一步减少计算量。结合优化K-means聚类获得的预选框参数来提高算法对特定目标的检测精度,并比较训练策略对模型性能的影响,将优化后的算法集成到用户交互界面,实现花卉识别定位。[结果]该系统实现了操作简单的花卉多目标的识别,具有实时反馈和较高准确率等优势,改进算法在余弦退火方式下训练得到的模型对图像的处理速度为每秒31.85帧,最高检测精确率达96.43%。[结论]这种基于YOLOv4的轻量级目标检测模型具有较高的识别率,系统对多目标花卉的检测具有可行性,为自动化作业提供技术支撑。

[Objectives]Multi-target recognition and positioning of flowers is the basis of automatic operations.Although large target detection network has great identification result,its high complexity makes it difficult to meet the requirements of real-time detection.In this paper,a lightweight target detection algorithm based on YOLOv4 was proposed.[Methods]MobileNetV3 was used to replace the original backbone feature extraction network,the shallow and deep features of the network were fused from the bottom up,and the Path Aggregation Network was simplified to further reduce the computation.Using the optimized anchor parameters obtained by K-means clustering,the detection accuracy of the algorithm for specific targets was improved,and comparing the impact of training strategies on the performance of the model,the optimized algorithm was integrated into the user interface to realize the identification and positioning of flowers.[Results]The system realized the multi-target recognition of flowers with simple operation and had the advantages of real-time feedback and high accuracy.The image processing speed of the model trained by the improved algorithm in cosine annealing mode was 31.85 frames per second,and the highest detection accuracy was 96.43%.[Conclusions]This lightweight target detection model based on YOLOv4 had a high recognition rate,and the system can be useful for multi-target flower detection,which can provide technical support for automatic operation.