Research On Reading Recognition Algorithm Of Industrial Instrument Based On Deep Learning

With the development of industrial intelligence and digitalization,the tasks of measuring and identifying artificial instruments in substations are gradually replaced by power inspection robots.In terms of industrial instrument identification,most inspection robots can only complete the acquisition of instrument data,and the subsequent identification work also requires manual participation in recording readings,which has problems such as high risk,low work efficiency and poor immediacy of identification.At the same time,in the traditional instrument recognition method based on machine vision,a large number of algorithms such as low efficiency,error detection error detection and other shortcomings seriously affect the recognition work.Therefore,this paper proposes a new method of substation instrument reading recognition based on deep learning,which can be applied to two types of instrument reading detection and recognition tasks of single pointer instrument and digital instrument,and realize the highly intelligent power inspection.After a large number of experimental verification and comparative analysis,the algorithm finally realized the reading recognition effect of industrial instrument with high precision and low error.This paper mainly studied and realized the following contents:(1)In the aspect of industrial instrument detection,this paper proposes an industrial instrument detection algorithm based on improved YOLOv5.The Ghost network module is introduced to reduce network parameters,and the model compression and channel pruning are used to further compress the model and extract multi-scale features.In order to compensate for the loss of precision,the characteristic distillation method based on FSP matrix is used to carry out characteristic distillation for the network.Experimental results show that the algorithm improves the detection accuracy of the instrument detection data set,and greatly reduces the model complexity,which provides a guarantee for the deployment of robot platform.(2)In terms of pointer meter reading recognition,this paper proposes a pointer meter reading recognition algorithm based on improved Deep Lab V3+ after region positioning of pointer meter according to YOLOV5 instrument detection algorithm.The algorithm improves the segmentation accuracy of the pointer and scale in the positioning area of the instrument.Then a perspective transform method based on segmentation information is proposed to correct the attitude of the instrument image.Finally,PCA linear detection and Angle method are used to complete the reading recognition of the pointer instrument.Experiments on experimental data set show that the algorithm improves the instrument segmentation accuracy and shortens the experiment time,and effectively solves the problem that traditional pointer instrument is affected by natural scene factors.(3)In the aspect of digital instrument reading recognition,this paper proposes a digital instrument reading recognition algorithm based on improved DBnet and CRNN.Firstly,dbNET text detection algorithm is used to detect the digital reading area,and then CRNN algorithm with attentional mechanism is proposed to recognize the reading.The feature sequence of the instrument reading image is extracted by the cyclic layer network,which is input to the Bi LSTM network for encoding,and then transmitted to the LSTM bidirectional decoder containing the Attention mechanism for decoding.Finally,the predicted reading is output.The test results of the experimental data set show that the algorithm can not only identify the pure digital reading in the digital instrument with high precision,At the same time,it effectively solves the problem of character conglutination and inaccurate identification of decimal point.