Adaptive range parameter reading method for pointer instruments based on cascaded multi-models

Abstract Aiming at the issues of weak universality and poor performance in complex scenarios of common automatic reading methods for pointer instrument, a self-adaptive range reading method based on cascaded multi-Models, which includes an improved EA-DETR model, U 2 -Net model and PP-OCRv4 model, is proposed in the paper. The improved EA-DETR model is used to identify target areas such as instrument panel, pointer, pointer rotation center, and scales in the image, and to encode the scale target areas based on their anchor box positions. The U 2 -Net model is used to extract the key points such as scale lines and pointer from the instrument panel image, and the endpoint positions of the long scale lines are determined by using the progressive circle exploration method. The PP-OCRv4 model is used to identify textual information such as scale area values and scale units. Based on the output information of three models, the dynamic local angle method is used for instrument reading. To improve the target recognition accuracy and efficiency of EA-DETR model, the improving methods of backbone network reconstitution, attention-based intra-scale feature interaction model optimization and new loss function introduction are explored for EA-DETR model. Ablation and comparative experiments show that the optimized model comparing with the original model, while the mAP@ 0.5–0.95 is improved to 0.736, reduces the number of parameters by 24.86%, the computational cost by 11.58%, the model weight by 25.43%, and achieves an FPS of 218.4 frames per second. Instrument reading experiments show that the mean relative error and the mean global error using the proposed method are 2.069% and 0.241% respectively, verifying the superiority, effectiveness, and generalizability of the proposed method.