基于改进YOLOv7-tiny算法的多种类不均衡样本水稻害虫检测

Detection of of Rice Pests in Multi-Species Unbalanced Samples Based on Improved YOLOv7-Tiny Algorithm

为实现基于机器视觉的田间水稻害虫检测,本研究结合IP102农业害虫数据集及网络资源,建立了含有26类标签的不均衡样本水稻害虫数据集;改进YOLOv7-tiny单阶段目标检测算法,以部分卷积PConv作为主要卷积核,结合极化自注意力机制(Polarized Self-Attention),将提取到的特征进行复杂双向多尺度特征融合,建立了适合多种类不均衡样本的水稻害虫检测模型。结果表明,在加入迁移学习和多尺度训练的条件下,改进后的YOLOv7-tiny检测算法在自建水稻害虫数据集的平均检测精度达到96.4%,单张图片的检测时间为8.8 ms,模型大小为9 055 kb,可实现对田间水稻害虫的快速准确识别,为水稻害虫的智能化检测和防治提供了技术支持。

In order to realize the detection of rice pests in the field based on machine vision,this study combined the IP102 agricultural pest dataset and network resources,and established an unbalanced rice pest dataset containing 26 kinds of labels.Improved YOLOv7-tiny single-stage target detection algorithm,using partial convolution as the main convolution kernel,combined with the Polarized Self-Attention mechanism,and carried out complex bidirectional multi-scale feature fusion for the extracted features.A rice pest detection model suitable for multi-species unbalanced samples was established.The results show that under the conditions of adding transfer learning and multi-scale training,the average detection accuracy of the improved YOLOv7-tiny detection algorithm in the self-built rice pest data set is 96.4%,the detection time of a single image is 8.8ms,and the model size is 9055 kb,which can realize the rapid and accurate identification of rice pests in the field.It provided technical support for the intelligent detection and control of rice pests.