CAM-YOLO v8: A model for water meter word wheel region segmentation

The use of mobile phone APP meter reading reduces the probability of human intervention and misidentification, but due to the complex location and geographical conditions of the buried water meter, as well as the base code and gear in the water meter dial, there are problems of low accuracy and rough segmentation results when the water meter word wheel region is divided, so a YOLO v8 segmentation method based on coordinate attention mechanism is proposed. The backbone network incorporates the AM (Attention-MobileNet_block) module to apply the Coordinate Attention (CA) mechanism, which facilitates the extraction of multi-scale features, reduces model parameters and computational complexity, and enhances computing efficiency. In the feature fusion stage, the C2f_CA mechanism, based on the Bottleneck_CA module, effectively mitigates background interference while concentrating on the water meter's target region. Experimental results indicate that, compared to other segmentation models, our model achieves an F1-score of 98.61% and a mean Pixel Accuracy (mPA) of 99.11% for word wheel region segmentation, establishing a solid foundation for the precise identification of water meter code words in the future.