Owing to the disorderly arrangement of water meter pipelines and the random rotation angles of their mechanical numeral wheels, the captured images of water meters reveal issues such as tilt, blurriness, and missing characters. It is evident that traditional optical character recognition fails to meet the detection requirements, and the two-stage detection method, positioning first and then recognizing, needs to be revised. In this study, water meter reading recognition is treated as an object detection task, whereby extracting detection box information output by the object detection algorithm facilitates the acquisition of water meter readings.YOLOv8n is chosen as the baseline model, and a target detection algorithm, GMS-YOLO, based on group multi-scale convolution, is put forward. Firstly, by replacing convolution in the Bottleneck module with group multi-scale convolution, the model achieves different scales of receptive fields, thereby enhancing its feature extraction capability. Secondly, the large-kernel separable attention (LSKA) is incorporated into the SPPF module to augment the perceptual ability of fine features. Lastly, the ShapeIoU bounding box loss function is opted to replace CIoU, enhancing the model's positioning ability and expediting its convergence speed. Evaluated on a self-compiled water meter image dataset, GMS-YOLO attained a [email protected] of 92.4% and a precision rate of 93.2%, representing increments of 2.0% and 2.1% over YOLOv8n and demonstrating significant superiority over the baseline model. The average detection time of GMS-YOLO per image is ten milliseconds, a capability well-suited to practical detection tasks.