Rapidly recognizing water meter readings is crucial for intelligent water management systems. Despite the widespread availability of smart water meters, the lower cost of word-wheel water meters means they continue to be used in most cases. As a result, manual reading and data review processes persist, hindering efficient management of water resources. Traditional recognition methods have been hampered by complex algorithms and insufficient robustness. This article proposes a deep-learning-based detection and recognition method for word wheel water meters, which involves dividing the reading process into three stages: 1) detection; 2) correction; and 3) recognition. We have targeted algorithm design to suit the unique environment where the water meter is located. We then made specific refinements and improvements to the recognition method to improve the performance. The method achieved an impressive segmentation accuracy of 98.2% and a recognition accuracy of 98.7% on a self-built data set collected throughout Hangzhou, China. Additionally, it boasts a small model size and a short inference time, showcasing excellent efficiency. By streamlining manual meter reading and data review processes, our approach holds great potential for facilitating effective water resource management.