基于Coarse-to-fine注意力机制的指针式仪表读数识别

Reading Recognition of Pointer Instrument Based on Coarse-to-fine Attention Mechanism

为了解决变电站指针式仪表读数识别过程中所涉及到待检测目标尺度分布范围广泛,而导致特征金字塔在不同尺度上不一致性的问题,提出一种基于注意力机制的特征融合的改进的YOLOv3算法。通过利用注意力机制自动调整各尺度特征映射融合的空间权重,该模块包括两步:同比例放缩和基于注意力机制的特征融合。采用改进后的YOLOv3算法检测图片中仪表的表盘位置,然后根据Coarse-to-fine的思想,使用Mask R-CNN对表盘指针采取实例分割操作,并使用简单线性回归对仪表中的指针进行线性拟合,计算出指针斜率,最终计算仪表读数。实验结果表明,改进后的YOLOv3算法对指针式仪表的识别精度达到了91.85%,对小目标检测效果有较高的提升,使模型具有更好的鲁棒性。同时,实例分割算法与线性回归模型的结合也为指针式仪表读数的自动识别提供了新思路。

In order to solve the problem that the target scale distribution range to be detected is wide in the process of reading recognition of pointer instrument in substation,resulting in the inconsistency of feature pyramid on different scales,an improved yolov3 algorithm based on feature fusion based on attention mechanism is proposed in this paper.By using the attention mechanism to automatically adjust the spatial weight of each scale feature fusion,the module includes two steps:feature resizing and feature fusion based on attention mechanism.This paper uses the improved YOLOv3 algo-rithm to detect the dial position of the meter in the picture.Then according to the thought of Coarse-to-fine,this paper use Mask R-CNN to segment the pointer in the dial,and use the simple linear regression algorithm to linearly fit the meter pointer,calculate the pointer slope,Finally calculate the instrument reading.The experimental results show that the im-proved YOLOv3 algorithm has a recognition accuracy of 91.85%for pointer meters,and the detection effect of small tar-gets is significantly enhanced.It has higher robustness.At the same time,the combination of instance segmentation and linear regression also provides a new idea for automatic identification of pointer meter readings.