| With the rapid development of machine vision,the intellectualization of pointer type dial instrument reading recognition has become an important research topic,and it is widely used in industry and military.Traditional instrument panel data acquisition mainly relies on manual writing,which has low efficiency and high error rate.Therefore,it is of great significance to intelligently identify and read the pointer dial instrument panel.This thesis of face detection and face alignment algorithm of multitasking cascade convolution network(Multi-task Cascaded Convolutional Networks,MTCNN)and FaceBoxes face detection algorithm is analyzed,and designed based on the multitasking cascade convolution network positioning method of dial detection and dial feature points,proposed a high accuracy realtime multiscale dial pointer detection method.The work and innovation of this paper mainly include:Firstly,a multi-task concatenated convolution network based dial detection and dial feature point location method is designed to detect the dashboard area in the whole picture and detect and locate the feature points in the dial,algorithm adopts the multi-task learning mechanism,makes full use of the relationship between subtasks,and realizes the dual tasks of detecting the dashboard area and locating the dial feature points in parallel.In addition,algorithm uses a new online hard sample mining strategy,and the training effect is better than manual sample selection.Through the training data set to generate the corresponding training model,using the training model of the test set images for dial detection and dial feature points localization,realizes the accurate detection of the dial area on the different background and dial feature points accurately positioning and lightweight convolution Neural Network(Convolutional Neural Network,CNN)design can achieve the effect of real-time detection.Secondly,a high-precision real-time multi-scale dial pointer detection method is proposed.This method has a lightweight but powerful network structure,which can achieve real-time effects and process dial Pointers of different scales,contains only a full convolutional neural network and are end-to-end training.In addition,it has a new anchor encryption strategy that allows different types of anchor points to have the same density on the input image,significantly increasing the recall rate of small dial pointers.First to test the dial pointer area based on the FaceBoxes network,then according to the detected the dial pointer frame combination of MTCNN positioning to dial plate heart feature points location preliminary judgement is made on pointer pointing to the information,further to detect the dial pointer box using the adaptive hoff straight line detection,and combined with the line length on both sides of the dish heart to finishing pointer pointing to the merit integrated two methods to judge the direction of the pointers,not only greatly improve the efficiency of the needle detection and accuracy,and the algorithm is more stable and better robustness.Finally,due to the error of the indicator of tilting dial in computer vision recognition,the mathematical principle of correcting the deviation based on rotation matrix was studied,and a dial recognition reading correction method based on affine transformation and threedimensional modeling was proposed to accurately calculate the compensation Angle,and then the dial recognition reading was corrected autonomously.OpenGL was used to build the dial model and the correction algorithm was used to fit the data.The correction algorithm was used to compensate the reading error of the tilted dial,so that the reading result was more accurate.Experimental results show that the algorithm can improve the accuracy of the dial reading recognition. |
