Automatic reading recognition algorithm for multi-category pointer meters based on lightweight network

Abstract The automatic reading recognition algorithm is widely used for its efficiency in pointer meter measurement. However, current algorithms lack adaptability when dealing with pointer meters of different ranges and have lower accuracy when dealing with low-quality images such as tilted, blurry, and poorly lit images. To address these issues, this research proposes a multi-category pointer meters reading recognition network(MMRNet). Firstly, the improved RetinaNet is used to achieve multi-category detection of pointer meters in complex environments, solving the classification problem of pointer meters with different ranges. Secondly, the augmentation perspective transformation module (APM) corrects perspective distortion, enhances the details of pointer meter dials, and improves the accuracy of reading low-quality images. Finally, the digital reading module (DRM) employs an improved hough transform to detect the pointer and scale lines, and subsequently utilizes an angle-based method to achieve accurate reading recognition. To validate the effectiveness and robustness of our approach, we constructed a multi-category dataset of pointer meters and conducted extensive experiments. Experimental results show that MMRNet outperforms state-of-the-art algorithms, achieving the highest detection accuracy of AP 50 at 0.993 and the average errors of the pointer meter readings are no more than 0.661%.