Research On Digital Display Instrument Recognition Algorithm Based On YOLOv5s And Transformer

Currently,digital display meters are widely used in the industrial field,but some meters cannot provide external data interfaces,requiring manual reading of meter data.This traditional approach is prone to human error and has low execution efficiency.In order to address this issue,this paper investigates digital display meter auto-identification technology within a real-world industrial setting,employing deep learning methodologies with the goal of enhancing the precision and real-time capabilities of character recognition.Firstly,in order to tackle the identification of the digital display meter area and the instrument character region,this article introduces a lightweight digital display meter detection model built on YOLOv5 s.Initially,the Shuffle-Block module is employed in the YOLOv5 s backbone network to reduce model parameter size.Subsequently,the SECA attention mechanism is added to the neck network to enhance the model’s feature expression capability,and the C3 Ghost module is utilized for further model lightweighting.The Light-Bi FPN structure is adopted to strengthen multi-scale feature fusion capabilities and reduce computational costs,and the loss function is reconstructed using the novel,efficient SIOU bounding box loss term.A variety of popular object detection models are chosen for performance comparison experiments using public datasets as well as custom industrial datasets.The outcomes indicate that the model successfully enhances detection precision while preserving its lightweight nature,and strikes a balance in terms of inference speed.Moreover,the model meets the accuracy and real-time requirements in actual production scenarios.Secondly,to address the issue of instrument character recognition,this paper designs a digital display meter character recognition network incorporating STN and Transformer.First,the STN spatial transformation network is used to perform spatial transformations on input images,correcting them horizontally and vertically.Next,the Swin Transformer encoder network is employed to perform multi-scale feature encoding on the corrected character images.Finally,the Transformer decoder is used to decode the encoded content and output the character prediction sequence.This model undergoes training and evaluation using both public datasets and bespoke industrial datasets,and is contrasted with multiple well-established character recognition algorithms.The findings demonstrate that this method attains outstanding performance in character recognition precision,while maintaining a balance in recognition speed.Lastly,this paper designs a digital display meter intelligent reading system,including system architecture,hardware platform,and software design,for automated reading of meter data.The system is deployed in the metrology calibration workshop of Shaanxi Provincial Metrology Institute and subjected to on-site functionality testing.Experiments demonstrate that the system meets the requirements for recognition accuracy and detection speed,possessing practical application value in actual production scenarios.