基于改进YOLOv4-tiny的节肢动物目标检测模型  

Arthropod Object Detection Model Based on Improved YOLOv4-tiny

作　　者：余咏 吴建平[1,2] 何旭鑫[1] 韦杰 高雪豪 YU Yong;WU Jian-ping;HE Xu-xin;WEI Jie;GAO Xue-hao(School of Information Science&Engineering,Yunnan University,Kunming 650504,China;Yunnan Provincial Electronic Computing Center,Kunming 650223,China)

机构地区：[1]云南大学信息学院,云南昆明650504 [2]云南省电子计算中心,云南昆明650223

出　　处：《计算机技术与发展》2024年第1期114-120,共7页Computer Technology and Development

基　　金：国家自然科学基金项目(62172354)。

摘　　要：针对自然环境下节肢动物背景复杂、形态万千、遮挡目标和目标尺度多样等因素,导致模型检测效率不高、边界框定位不准确的情况,提出一种基于改进YOLOv4-tiny的节肢动物目标检测模型。首先,结合空间、通道卷积注意力机制(CBAM),抑制背景噪声;其次,引入可变形卷积(DCN)以及改进的加权双向特征金字塔,重塑卷积和特征融合方式进行多尺度预测;最后,在FPN网络中引出一层Feat@3,嵌入空间金字塔池化结构,有效提取节肢动物的各种显著特征,使模型泛化能力更强,将改进后的模型命名为YOLOv4-tiny-ATO。实验结果表明,该模型在大小仅为54.6 Mb的前提下,很好地平衡了检测速度和检测精度,检测精度为0.725,检测速度达到89.6帧·s-1,召回率为0.585,较改进前相比YOLOv4-tiny模型,检测精度提高0.426,模型在模型大小、检测速度上更适用于移动端部署,模型检测精度也能达到应用标准,满足对节肢动物的检测需求。Aiming at the situation that the model detection efficiency is not high,and the bounding box prediction is wrong caused by the complex background,variety of morphology,occlusion target and diverse target scale of arthropods in the natural environment,an arthropod target detection model based on improved YOLOv4-tiny is proposed.Firstly,combining spatial and channel convolutional attention mechanism(CBAM),the background noise is suppressed.Secondly,deformable convolution(DCN)and an improved weighted bidirectional feature pyramid are introduced to reshape the convolution and feature fusion methods for multiscale prediction.Finally,a layer of Feat@3 is extracted in the FPN network,and a spatial pyramid pool structure is embedded to effectively extract various significant features of arthropods,so as to enhance the generalization ability of the model.The improved model is named YOLOv4-tiny-ATO.The experimental results show that the proposed model balances detection speed and accuracy well with a size of only 54.6 Mb.The detection accuracy is 0.725,the detection speed reaches 89.6 frames per second,and the recall rate reaches 0.585,which is 0.426 higher than that of the YOLOv4-tiny model before the improvement.The model is more suitable for mobile deployment in terms of model size and detection speed,and the model detection accuracy can also meet the application standards to meet the detection needs of arthropods.

关 键 词：节肢动物 目标检测 可变形卷积 YOLOv4-tiny 双向特征金字塔 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]