A mechanical water meter reading detection method based on improved Yolov8n

To address challenges arising from dirty water dials, uneven lighting conditions, and varied shooting angles—which reduce the accuracy and recognition rates of mechanical water meter identification, we propose an intelligent reading method for old water meters based on improved YOLOv8n. Firstly, the C2f module is augmented with an efficient multi-scale attention mechanism employing dimensional decomposition and cross-channel correlation strategies to enhance discriminative feature representation. Secondly, the bidirectional feature pyramid network is used in the Concat module. The neck network also integrates shallow-level feature maps and a specialized prediction module to maintain high detection accuracy under interference conditions. Finally, the normalized wasserstein distance loss is combined with CIoU as a location regression loss function to reduce location bias sensitivity. The results show that the mAP₅₀ of the improved algorithm reaches 98.4%, and the mAP₅₀ and recall rate R are increased by 4.6% and 1.7%, respectively. It still has strong robustness in the face of complex condition interference.