| Because of the low cost and the readings are not susceptible to environment factors,the pointer meter is widely used in petroleum,chemical,aerospace and so on.However,because the analog meter has no digital interface,there are inefficiencies in using humans to read the data.Therefore,it is of great significance to intelligently read the readings of the pointer meter.At present,the traditional pointer-type meter reading recognition algorithm can only operate in a specific environment or a fixed location,and has no high reliability.With the development of deep learning,this dissertation combines digital image processing technology and deep learning related technologies to conduct in-depth research on the intelligent recognition of reading of pointer meter,which mainly includes the following three aspects:(1)The positioning and segmentation of the dial.This dissertation proposed an automatic recognition method for pointer-type instrument panels based on the improved Mask R-CNN algorithm,which locates and segments dial for pictures containing the natural environment,and reduces the amount of computation in subsequent image processing.Experiments show that the target detection accuracy of the improved Mask R-CNN algorithm is improved by2.1%,and the instance segmentation accuracy is improved by 1.9%.Compared with the traditional method of dial positioning algorithm,this algorithm has the characteristics of high accuracy and strong robustness.(2)Preprocessing of the dial.This dissertation conducts in-depth research on the characteristics of the collected instrument images,and finds that it is difficult to ensure that the camera and the instrument panel are always on the same horizontal plane during image acquisition,and the instrument panel area will be tilted to different degrees.Therefore,the instrument panel is tilt-corrected using a perspective transformation.On this basis,the images are preprocessed such as grayscale,binarization,and connected domain labeling to enhance the features of the dial image.(3)The positioning and fitting of the pointer and the reading of the meter.The pointers are first extracted and refined using ordinary least squares.Then the projection method is used to segment the numbers and tick marks on the dial.On this basis,the improved Le Net-5convolutional neural network is used to identify the numbers on the dial.The experimental results show that the recognition accuracy of the improved Let Net-5 reached 98.94%.Finally,combined with the angle method to calculate the meter number.Compared with manual readings,the algorithm readings are found to be more accurate and efficient.The pointer meter detection and recognition method based on deep learning proposed in this dissertation realized the accurate detection,image correction and reading recognition of the meter.Compared with the traditional Mask R-CNN algorithm,the proposed improved Mask R-CNN algorithm in this dissertation has higher reading recognition accuracy and stronger robustness,and is suitable for automatic meter reading of pointer-type meters in various environments,providing a feasible idea for target detection and meter reading of pointer meters. |
