Research On Industrial Instrument Indication Recognition In Complex Environment Based On Deep Learning

Industrial instruments are important equipment to ensure industrial production.Accurately identifying the values of industrial instruments is crucial to ensure the stability and efficiency of production.The recognition scheme based on deep learning technology has good robustness and generalization ability,which can effectively complete instrument recognition tasks.However,its disadvantage is that the model occupies a large amount of parameters and computation,making it difficult to deploy in resource limited devices.This article takes the actual instruments of a certain factory as the research object,proposes an industrial instrument indication recognition algorithm based on deep learning technology,and ensures the practicality and feasibility of the proposed algorithm through experiments.The main research content of this article is as follows:1.Aiming at the positioning and classification of multiple types of instruments in actual factories,a lightweight instrument positioning model based on pruning is proposed.Firstly,a meter positioning dataset is constructed by combining meter images and data augmentation methods.Secondly,combined with multi-scale feature information and the characteristics of instrument images in industrial environments,the model structure of Yolov5 is reduced,such as removing large-scale downsampling and feature fusion operations.Thirdly,the channel pruning and soft pruning methods are studied.Based on Yolov5,pruning methods such as batch normalization layer weight constraints and convolutional layer weight constraints are used to prune models of different detection scales,the performance of models with different pruning methods and pruning ratios in instrument positioning tasks is compared and analyzed.The experimental results show that the method based on the weight constraint of the convolutional layer has the best performance under a large pruning ratio.Compared with the original Yolov5 model,its parameter size is reduced by 88.03%,the amount of floating-point operations is reduced by60.75%,and external memory is reduced by 86.98%.The detection precision of each category can be maintained above 99%.2.Aiming at the accuracy and real-time requirements of digital display instrument indication recognition in the actual factory environment,a digital display instrument recognition method based on the lightweight Yolov5 model is proposed.Firstly,obtain digital instrument images through localization and perform data augmentation to construct a dataset.Secondly,based on Yolov5,the model is simplified by introducing the bneck module in Mobile Netv3,and construct a depth-separable convolution module(bneck-CA-E)based on the CA attention mechanism,and then propose a lightweight Yolo recognition model.The experimental results show that the improved lightweight Yolo model can achieve 98.92% recognition accuracy of logarithmic display instrument characters.Compared with the original Yolov5 model,this model has achieved a greater degree of optimization in terms of parameter size,the amount of floating-point operations,and external memory,which are respectively reduced by 91.97%,81.25%,and 90.18%,and the detection speed is increased by 56.85 frames per second to 84.35 frames per second.3.A pointer instrument recognition method based on Unet model is proposed for the task of pointer instrument recognition in actual factory environments.First,the Unet model is used to realize the division of the instrument plate and pointer.By introducing the batch normalization layer in the VGG backbone network,and adding the CBAM attention mechanism in front of each pooling layer,the detection performance is improved.The experimental results show that the precision of the improved Unet model reaches 98.72%,the recall rate is 96.42%,and the m IOU index is 93.02%,which proves the accuracy and effectiveness of the improved model.Secondly,the edge detection-based projective transformation matrix calculation method is used to rectify the mask image of the instrument plate,and the average error of the indicator recognition after correction is only0.076.Finally,the PCA method is used to fit the straight line to the corrected pointer mask image,and the angle method is used to calculate the indication.In the experimental results compared with the least squares method,the average error of the PCA method is only0.0446.