Multiple paddy disease recognition methods based on deformable transformer attention mechanism in complex scenarios

Abstract

AbstractPaddy disease recognition presents challenges in the agricultural industry, and existing algorithms struggle to accurately identify diseases in complex scenarios. In this paper, we propose a precise object detection framework to address the challenges of severe overlap, multi-disease detection, morphological irregularities, multi-scale object classification, and complex scenarios in real-world environments in paddy disease detection. The proposed model is based on an improved version of the DEtection TRansformer (Detr) algorithm. The enhanced network architecture fuses multi-scale features by adding a feature fusion module after the backbone network, which is able to retain more original information of the images and greatly improves the detection accuracy; the use of deformable attention module greatly reduces the computational complexity of the model. To evaluate the PDN, a dedicated paddy disease detection dataset with 1200 images is created. Experimental results demonstrate that the proposed model obtained a precision value of 100%, a recall value of 89.3%, F1-score of 94.3%, and a mean average precision (mAP) value of 60.2%. The model outperforms the existing state-of-the-art detection models in detection accuracy.KEYWORDS: Paddy disease recognitionTransformermachine vision detection Disclosure statementNo potential conflict of interest was reported by the author(s).Additional informationFundingThis work was supported in part by the Jiangsu Basic Science (Natural Science) Research Projects in Higher Education Institutions (No.23KJB460034), Jiangsu province Youth Fund Project (No.BK2023040059), the China Postdoctoral Science Foundation Funded Project (No. 2022M721185), Jiangsu Agriculture Science and Technology Innovation Fund (No. CX(21)3145).Notes on contributorsXinyu ZhangXinyu Zhang is currently a master's student in mechanical engineering at the School of Mechanical Engineering, Yangzhou University. His research interest is machine learning.Hang DongDr. Hang Dong is a lecturer at Yangzhou University. He received his PhD degree in Mechanical Manufacture and Automation from Dalian University of Technology (2019). His research interests include the deep learning, machine learning, and robotics. Hang Dong is the corresponding author and can be contacted at hdong@yzu.edu.cn.Liang GongLiang Gong was born in Maanshan City, Anhui Province, China on October 26, 1999. He received his bachelor's degree from Anhui Polytechnic University in 2021. He is currently a master's student in mechanical engineering at the School of Mechanical Engineering, Yangzhou University. His research interests are machine vision and machine learning.Xin ChengXin Cheng was born in Lian Yungang, China, in 2002.He is currently a student in Yangzhou University.His research interests include computer vision,natural language processing.Zhenghui GeZhenghui Ge is currently an associate professor at Yangzhou University, China. He received his PhD degree from Nanjing University of Aeronautics and Astronautics, China, in 2018. His researches mainly focus on electrochemical machining.Liangchao GuoDr. Liangchao Guo is a distinguished research fellow at Yangzhou University. He received his PhD degree in Mechanical Manufacture and Automation from Dalian University of Technology (2019). His research interests include the fabrication and application of gas sensing, and storage devices.