Research on pointer meter reading algorithm based on YOLO and OCR

Abstract Improving the reading accuracy and applicability of pointer-type meters is crucial to the reliability of industrial production. To address the problems of low efficiency and error-prone of traditional manual reading, as well as the insufficient applicability of the existing algorithms, this study proposes a meter reading algorithm based on YOLO and OCR, which constructs a twice localization framework. Firstly, the Oriented Bounding Box model is used to complete the first localization of the pointer and scale, and then the Hough detection is used to accurately locate the pointer features, and the MLFFA-YOLO based scale detection model is proposed for the scale features. The model adopts the improved Enhanced Repvgg (Enh_Repvgg) module to enhance the cross-scale feature extraction capability of the backbone network. Designed the Split-Channel Multi-Scale Mean Fusion Module (SCMS-MFM) combined with improved Enhanced Mixed Local Channel Attention (Enh_MLCA) module for the Multi-layer Feature Fusion Attention (MLFFA) module to realize multi-scale feature enhancement of neck network. The original loss function is replaced by Powerful-IoU (PIoU) loss function. Finally, a post-processing algorithm is designed in this study to reduce the misrecognition problem of the Optical Character Recognition model through scale value logic checking. The experimental results show that the average localization accuracy of the proposed model reaches more than 97% in various environments on pointer-type meters with an accuracy class of 1.6, and the average relative error of the readings after post-processing is 0.551%, which meets the high-precision requirements of industrial meters detection.