Research On Pointer Meter Reading Recognition Method Based On Deep Learning

As the prevalence of digitization and intelligentilization in industry,pointer meters,due to their advantages in reliability and stability,are still broadly applied for condition monitering in substations.However,most of the existing deep learning-based methods explicitly detect the pointer and scales by semantic segmentation,thus demanding high quality meter images.Therefore,these methods cannot well adapt to environmental interferences,and are not capable of providing satifactory performance when the meter images are blurred,yellowing,or stained,etc.This thesis proposes a new deep learning-based method for pointer meter detection and reading recognition,which evades detecting meter pointers and scales marks ex-plicitly.Instead,it predicts the pointer meter orientation directly by combining pointer detection,scale recognition and interference reduction implicitly,and finally calculate the meter reading.The main research content are as follows:1.Meter detection method based on improved Faster R-CNN and meter image rec-tification method based on G-RMI algorithm.Most of the deep learning-based object detection tasks require large scale dataset,and training on numerous data is time-consuming.Therefore,this method proposes to fine-tune the detection model with an COCO-pretrained backbone on a subset of virtual substation scene dataset,which leads to an mAP of 99.7%with 2000 training samples.Then,to solve the problem that the meter detection model lacks transferability on real data,this method proposes to exert constraint on the image feature,forcing the model to pay attention to not only the shell but also the dial region of the meter.Exper-imental results shows that through feature constraint,the performance of meter detection on real data is promoted by 1.7%.After that,the meter keypoints are rectified by the G-RMI algorithm,and the meter image is rectified.Experiments shows that this method can rectify the meter images effectively.2.A meter reading recognition method based on deep regression.Given a detected meter,this method employs ResNet50 to extract image feature,and obtain the pointer orientation using an orientation regression module,and computes the me-ter reading at last.Compared with existing meter reading recognition method based on Mask R-CNN,this method circumvents the difficulty of detecting meter keypoints,therefore improves the reading recognition performance.On the vir-tual meter dataset,this method achieves an average reading precision of 97.2%,which is higher than the Mask R-CNN-based method by 7.4%.Quantitative and qualitative results demonstrates that this method has stronger robustness against meter interferences such as blurring,yellowing and stains,etc.The method for pointer meter detection and reading recognition proposed by this thesis,can accurately detect,rectify and recognize the meters.In comparison with ex-isting deep learning-based methods,this method is more accurate and robust in reading recognition.