Research On Industrial Instrument Detection And Automatic Reading Method Based On Deep Learning

With the continuous promotion of intelligence and automation,industrial instruments have been widely used in various fields,and the research on industrial instrument detection and automatic reading methods has gradually emerged.However,due to the complexity of industrial field environment,the traditional industrial instrument detection and automatic reading methods are not highly robust,less applicable in complex environments such as lighting changes,instrument image distortion,and coexistence of multiple types of instruments,and the reading steps are relatively cumbersome.In this paper,we propose a deep learning-based method for detecting and automatically reading industrial instruments,which has high robustness and generality in complex environments,can simultaneously identify instruments of different phenotypes through the superior performance of deep learning in the field of image recognition,and simplifies the reading steps to a certain extent.The main research of this paper has the following points:(1)An Improved YOLOv4 industrial instrument detection method is proposed to address the problems of difficult detection of industrial instruments under complex situations such as dark light,similar instruments,and blurred images,and the lack of robustness of traditional detection methods.Firstly,the CBAM attention mechanism is introduced in the detection neck of YOLOv4 network to make the network adaptively enhance the important channel domain and spatial domain information in instruments detection,and then the classification and regression branches in the detection head of YOLOv4 network are decoupled to form two sets of parallel branches,so that the classification and regression branches can independently extract the feature information suitable for the corresponding tasks,thus improving the network learning capability.The experimental results show that the Improved YOLOv4 has faster training convergence and better instruments detection performance,and the mAP value increases from 98.41%to 99.34%of YOLOv4 network,which is an improvement of 0.93%.(2)Firstly,the SIFT-RANSAC algorithm is introduced to conduct the pointer instruments image distortion correction for the pointer instrument tilt and distortion problem,and then,an Improved DeepLabv3+based automatic reading method for pointer instruments is proposed to address the problems that existing automatic reading methods for pointer instruments do not have high robustness and small applicability in complex environments.Considering the problems of large number of parameters and slow recognition speed of DeepLabv3+network,the backbone network of DeepLabv3+is replaced by MobileNetv2 network to achieve fast segmentation and extraction of pointer region and scale region.Compared with DeepLabv3+,the MobileNet-DeepLabv3+network is about 1x faster in segmentation and the number of parameters in the network is reduced by about 9/10,and the mIOU decreases slightly but does not affect the accuracy of pointer instruments reading.Meanwhile,the experimental results show that the automatic reading method of pointer instruments proposed in this paper can also achieve better reading accuracy in dark light,blur,reflection,distortion and other complex environments,and can carry out reading recognition of pointer instruments of different phenotypes,and the average relative error of the method in this paper is 2.13%,and the average referencing error is only 0.46%.(3)To address the problems of tedious steps and low robustness of traditional digital instruments automatic reading methods,a deep learning-based digital instrument automatic reading method is proposed in this paper,firstly,treating digital display area localization as a rotating frame target detection task,adding angle regression branch in CenterNet network to realize precise localization of digital display area and tilt angle regression,then using tilt angle directly to achieve the tilt correction of digital display area and crop the digital character image,and finally use CRNN network to achieve the end-to-end recognition of digital characters.The experimental results show that the method in this paper can accurately locate the digital display area and correctly identify the readings even under the complex environment such as instrument tilt,reflection and dark light,and its average accuracy of reading recognition reaches 93.0%with an average execution time of 47.6ms,which can effectively realize the automatic reading of digital instruments.