This paper proposes a lightweight residual network for highly accurate recognition of the numbers of wheel dials. The proposed residual network is based on an optimized ResNet-18 network with the depthwise separable convolution and Dropout function. In addition, a case study on automatic water meter readings based on captured images is presented. Compared to existing methods, the results demonstrate improved recognition accuracy as high as 99.8% and increased computation speed with reduced model parameter size.