[1]张剑飞,柯 赛.基于 YOLOX-s 的农业害虫检测研究[J].计算机技术与发展,2023,33(05):208-213.[doi:10. 3969 / j. issn. 1673-629X. 2023. 05. 031]
　ZHANG Jian-fei,KE Sai.Detection of Agricultural Pests Based on YOLOX-s[J].,2023,33(05):208-213.[doi:10. 3969 / j. issn. 1673-629X. 2023. 05. 031]

点击复制

基于 YOLOX-s 的农业害虫检测研究()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年05期

页码:

208-213

栏目:

新型计算应用系统

出版日期:

2023-05-10

文章信息/Info

Title:

Detection of Agricultural Pests Based on YOLOX-s

文章编号:

1673-629X(2023)05-0208-06

作者:

张剑飞; 柯 赛

黑龙江科技大学 计算机与信息工程学院,黑龙江 哈尔滨 150022

Author(s):

ZHANG Jian-fei; KE Sai

School of Computer & Information Engineering,Heilongjiang University of Science & Technology,Harbin 150022,China

关键词:

目标检测; YOLOX-s; 害虫检测; Swin-Transformer; 注意力机制

Keywords:

object detection; YOLOX-s; pests detection; Swin-Transformer; attention

分类号:

TP520. 2040

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 05. 031

摘要:

针对现有目标检测算法难以应对现代农业环境下多种类害虫高精度检测的问题, 提出了一种基于 Swin -Transformer 和 YOLOX-s 改进的 ST-YOLOX-s 目标检测模型,实现对 30 类常见害虫的有效目标检测工作。 为解决 YOLOX-s 模型对小型目标害虫检测效果不佳的问题,在 YOLOX-s 模型基础上添加 P2 特征尺度,提升模型对小型目标害虫的检测能力;为弥补卷积神经网络对通道信息关注薄弱的问题,将高效通道注意力模块添加到 YOLOX-s 的 CSPLayer,强化卷积神经网络的特征提取能力; 为探索高效自注意力机制下的模型全局特征学习能力, 将添加图属性的层次化 Swin -Transformer 结合到网络模型,弥补卷积神经网络忽视全局特征的问题;最后通过 琢-CIoU 回归定位损失来实现高精度检测定位。 实验表明,ST-YOLOX-s 在多种类害虫检测上具有更好的检测性能,最终 AP50 以及 AP50 -95 检测结果分别达到92. 27% 和 67. 32% ,相比较 YOLOX-s 模型检测精度分别提高了 2. 01% 和 1. 91% 。 同时 ST-YOLOX-s 检测模型与其他主流模型相比检测精度也有显著优势。

Abstract:

To solve the problem that the existing target detection algorithms cannot meet the high precision detection of many kinds ofpests in a modern agricultural environment,we present an?

improved ST-YOLOX-s target detection model based on Swin-Transformerand YOLOX-s,which achieves effective target detection for 30 kinds of common pests. To solve the problem?

that the YOLOX-s modelis not effective in detecting tiny target pests,a P2 feature scale is added on the basis of YOLOX-s model to improve the detection abilityof the model on small?

target pests. To compensate for the weakness of channel information in convolution neural networks,an efficientchannel attention module is added to the CSPLayer of YOLOX-s to?

enhance the feature extraction ability of convolution neural networks.To explore the learning ability of global features of models under an efficient self-attention mechanism,a hierarchical Swin-Transformerwith graph attributes is incorporated into the network model to compensate for the neglect of global features by convolutional networks.Last Passed 琢-CIoU regression locates loss to achieve high-precision detection and positioning. Experiments show that ST-YOLOX-shas better detection performance on many kinds of pests. The final detection results?

of AP50 and AP50-95 reach 92. 27% and 67. 32% ,respectively. Compared with the YOLOX-s model,the detection accuracy is improved by 2. 01% and 1. 91% ,respectively. At the sametime,　the ST-YOLOX-s detection model also has a significant advantage in detection accuracy compared with other mainstream models.

常用功能