基于YOLO算法的多类目标识别  

Multi-object Recognition Based on YOLO algorithm

作　　者：于秀萍[1] 吕淑平[1] 陈志韬 YU Xiuping;Lü Shuping;CHEN Zhitao(College of Automation, Harbin Engineering University, Harbin 150001, China)

机构地区：[1]哈尔滨工程大学自动化学院,哈尔滨150001

出　　处：《实验室研究与探索》2019年第3期34-36,76,共4页Research and Exploration In Laboratory

基　　金：黑龙江省教育科学"十三五"规划重点课题(GBB1317028)

摘　　要：针对传统目标识别方法需要人工设计特征工程,费时费力,泛化性能差的缺点,以YOLO算法和tiny-yolo模型为基础,在tiny-yolo的基础上增加了3×3卷积层和NIN(Network in Net Work)卷积层,设计了一个包含15个卷积层的神经网络模型m-yolo。在voc2007和voc2012数据集上的实验结果表明,m-yolo模型提高了识别的准确性和定位的精确性,并且保证了在识别速度上与tiny-yolo基本保持一致,平均识别时间仅上升了0. 6 ms。Artificial design feature engineering is required to improve the traditional target identification methods which are time-consuming and laborious, and has poor generalization performance. A neural network model with 15 convolutional layers is designed based on YOLO algorithm and tiny-yolo model. On the basis of tiny-yolo, a 3×3 convolution layer and an NIN convolution layer are added. Experimental results on voc2007 and voc2012 data sets show that the new model improves the average mean average precision and recall. The recognition speed is the same as tiny-yolo, but the average recognition time only increases by 0.6 ms.

关 键 词：YOLO算法 目标识别 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]