液晶与显示, 2020, 35 (12): 1291, 网络出版: 2020-12-28
基于Faster R-CNN的仪表识别方法
Instrument recognition method based on Faster R-CNN
摘要
针对仪表识别系统背景复杂, 对小目标不敏感, 检测精度低等问题, 本文提出一种特征融合金字塔(FPN)和Faster R-CNN网络结合的仪表自动识别方法。首先使用FPN和Faster R-CNN网络的RPN结合定位表盘和指针区域, 并对多类仪表进行分类; 此外为了平衡仪表图像的正负样本, 提高检测准确性, 引入Focal Loss损失函数与RPN网络结合进行训练数据集; 其次对指针区域进行基于FPN的图像分割, 将FPN网络与反卷积结合, 提高指针区域分割准确性; 最后拟合指针获取指针偏转角度, 得到仪表读数。实验结果表明, 提出的方法准确率达到94.25%, 与传统算法相比, 提出的方法不仅检测精度高, 而且实用性更强。
Abstract
In view of the complex background of the instrument recognition system, insensitive to small targets, and low detection accuracy, an automatic instrument identification method combining Feature Fusion Pyramid (FPN) and Faster R-CNN network is proposed in this paper. First, FPN and the RPN of the Faster R-CNN network are used to combine the positioning of the dial and the pointer area, and to classify multiple types of meters. In addition, in order to balance the positive and negative samples of the meter images and improve the detection accuracy, the Focal Loss loss function is combined with the RPN network to train the data set. The FPN-based image segmentation is performed on the pointer area. The FPN network and deconvolution are combined to improve the accuracy of the pointer area segmentation, and finally the pointer is fit to obtain the pointer deflection angle and obtain the meter reading.The experimental results show that the accuracy of the proposed method reaches 94.25%. Compared with the traditional algorithm, the proposed method not only has high detection accuracy, but also has stronger practicability.
李娜, 姜志, 王军, 董兴法. 基于Faster R-CNN的仪表识别方法[J]. 液晶与显示, 2020, 35(12): 1291. LI Na, JIANG Zhi, WANG Jun, DONG Xing-fa. Instrument recognition method based on Faster R-CNN[J]. Chinese Journal of Liquid Crystals and Displays, 2020, 35(12): 1291.