基于快速区域卷积神经网络识别变形仪表图像的二次矫正方法

为解决工业中摄像头在俯视、仰视等角度拍摄仪表时导致表盘变形影响读数的问题,提出一种基于Faster R-CNN识别变形仪表图像的二次矫正方法。利用ResNeXt50作为Faster R-CNN的主干网络,结合特征金字塔FPN(Feature Pyramid Network)生成特征层,并融合SENet(Squeeze and Excitation Networks)模块将仪表特征更为突出,便于定位仪表区域并裁减,再由Harris角点检测找到表盘的四个顶点；利用二次矫正方法对变形仪表还原成正视角度仪表；最后计算示数。实验结果表明：该方法使得mAP值由基本模型的75.51%提升至94.45%,且仪表在变形情况下,仍能得到比较好的读数结果,误差率为0.83%。

Second correction Method for deformable instrument image Recognition based on Faster R-CNN

In order to solve the problem that the dial deformation affects the reading when the camera shoots the instrument from the overhead and elevation angles in industry, a secondary correction method based on Faster R-CNN is proposed to identify the deformation instrument image. ResNeXt50 is used as the backbone Network of Faster R-CNN, and the Feature layer is generated by combining Feature Pyramid Network (FPN). Moreover, SENet (Squeeze and Excitation Networks) module is integrated to make the meter features more prominent, so it is easy to locate and trim the meter area. The four vertices of the dial are located by Harris corner detection. Using the second correction method, the deformation instrument is reduced to the front Angle instrument. Finally, calculate the indicator number. The experimental results show that the mAP value is increased from 75.51% of the basic model to 94.45% by this method, and the instrument can still get a good reading result under the deformation condition, and the error rate is 0.83%.