基于改进Tiny-YOLOv4的棉花顶芽轻量化识别方法  

Research on Cotton Terminal Bud Recognition Based on Improved Tiny-YOLOv4 Algorithm

作　　者：亚森江·白克力 周显锞 刘锋 岳勇 赵子祺 Yasenjiang Baikeli;ZHOU Xian-ke;LIU Feng;YUE Yong;ZHAO Zi-qi(College of Mechanical and Electrical Engineering,Xinjiang Agricultural University,Urumqi 830052,China;Institute of Computer Innovation and Technology,Zhejiang University,Hangzhou 310000,China;Department of Laboratory and Base Management,Xinjiang Agricultural University,Urumqi 830052,China)

机构地区：[1]新疆农业大学机电工程学院,乌鲁木齐830052 [2]浙江大学计算机创新技术研究院,杭州310000 [3]新疆农业大学实验室与基地管理处,乌鲁木齐830052

出　　处：《新疆农业大学学报》2023年第5期414-420,共7页Journal of Xinjiang Agricultural University

基　　金：新疆维吾尔自治区青年科学基金项目(2022D01B91);新疆维吾尔自治区自然科学基金面上项目(2021D01A101)。

摘　　要：为了解决既要提高棉花顶点识别效率,又要简化复杂模型以便在边缘计算设备上部署的问题,本研究提出了一种基于Tiny-YOLOv4的改进方法。对Tiny-YOLOv4的backbone网络进行了结构优化,通过用更高效的Bottleneck模块替换原有的第一个BottleneckCSP模块,简化了模块内的连接,直接将部分输入特征图传递至输出,有效减少了模型的参数数量和计算复杂度。针对棉花顶芽的检测还引入了微型CSP-Spatial Pyramid Pooling模块,旨在提升对微小物体的检测准确性并进一步降低计算负担。经验证,改进后的模型相比Tiny-YOLOv4模型,在计算复杂度上减少了33%的浮点运算,处理时间缩短至0.071 s。结果表明,该方法在确保高效识别的同时,显著降低了计算复杂度,展现了其在轻量化部署方面的优势。In order to solve the problem of both improving the efficiency of cotton vertex recognition and simplifying the complex model for deployment on edge computing devices,this paper proposes an improved approach based on Tiny-YOLOv4.The backbone network of Tiny-YOLOv4 was structurally optimized,and the numbers of parameters and computational complexity of the model were effectively reduced by replacing the original first BottleneckCSP module with a more efficient Bottleneck module,simplifying the connections within the module,and passing part of the input feature maps directly to the output.In addition,a miniature CSP-Spatial Pyramid Pooling module was introduced for the detection of cotton terminal buds,which intended to improve the detection accuracy of tiny objects and further reduce the computational burden.It was verified that the improved model reduced the computational complexity of floating-point operations by 33%and the processing time to 0.071 seconds compared to the original Tiny-YOLOv4 model.Results showed that the method could significantly reduce the computational complexity while ensuring efficient recognition,demonstrating its advantages in lightweight deployment.

关 键 词：Tiny-YOLOv4 轻量化识别 棉花顶芽 特征提取 

分 类 号：S562[农业科学—作物学] TP391.41[自动化与计算机技术—计算机应用技术]