一种基于级联R-CNN的水稻害虫检测算法  

Rice Insects Detection Algorithm Based on Cascaded R-CNN

作　　者：刘凯旋 黄操军[1] 李亚鹏 佟尚谕 Liu Kaixuan;Huang Caojun;Li Yapeng;Tong Shangyu(College of Information and Electrial Engineering,Heilongjiang Bayi Agricultural University,Daqing 163319)

机构地区：[1]黑龙江八一农垦大学信息与电气学院,大庆163319

出　　处：《黑龙江八一农垦大学学报》2021年第5期106-111,134,共7页journal of heilongjiang bayi agricultural university

摘　　要：为了解决水稻害虫识别目标小而导致算法识别精度相对较低的问题,通过多种优化算法改进级联R-CNN的结构,提出了一种基于级联R-CNN的水稻害虫检测算法。首先通过特征金字塔优化提取的小目标特征,其次用ROI Align替换特征提取模型中的ROI池化,减少小目标特征的丢失,最后用Soft-NMS减少重叠目标对小目标丢失的影响。实验表明,该方法能够有效的识别和检测复杂背景下的水稻害虫,且多种害虫准确率平均值mAP达到94.15%。In order to solve the problem of relatively low recognition accuracy of the algorithm due to the small recognition target of rice insects,it used a variety of optimization algorithms to improve the structure of Cascaded R-CNN,and proposed a rice insects detection based on cascaded R-CNN algorithm.Firstly,the feature pyramid was used to optimize the extracted small target features.Secondly,ROI Align was used to replace the ROI pooling in the feature extraction model to reduce the loss of small target features.Finally,Soft-NMS was used to reduce the impact of overlapping targets on the loss of small targets.Experiments showed that this method could effectively identify and detect rice insects in a complex background,and its average accuracy AP reached 94.15%.

关 键 词：级联R-CNN 水稻害虫 空间金字塔池化 图像分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]