基于深度学习YOLO模型的植物图像识别算法研究

Research on recognition algorithm of plant image based on deep learning YOLO model

为提高YOLO模型进行植物图像快速识别的准确率,对基于深度学习YOLO模型算法进行研究。在传统网络模型的基础上,引入了可变形卷积,将损失函数加入分类函数层中,有效的提高了模型的泛化性能,结合cuDNN库进行并行计算,提高了算法的效率。采用改进的YOLOv3网络模型结合公开数据集进行测试,实现了102种植物的有效识别,平均检测时间为1.275 s,在Top-2、Top-3准确率上均达到96%,与传统YOLO算法相比,识别的准确率和效率均有提高。结果表明,所建立的基于深度学习YOLO模型,在复杂多种类植物识别方面有推广应用价值。

In order to improve the accuracy of YOLO model in fast plant image recognition,the algorithm was studied based on the deep learning YOLO model.Based on the traditional network model,the deformable convolution was introduced,and the loss function was added to the classification function layer,which improved effectively the generalization performance of the model.Parallel computing was carried out by the algorithm combining with cuDNN library,which improved the efficiency of the algorithm.Improved YOLOv3 network model was used to test combining with the public data set,102 kinds of the plants were identified effectively.Average detection time was 1.275 s,and the accuracy of top-2 and top-3 was 96%.Compared with the traditional YOLO algorithm,the recognition accuracy and the efficiency were improved.Results showed that the established YOLO model had popularization and application value based on deep learning in the complex and multi species plant recognition.