To address the challenges of character adhesion, strong background interference, and recognition ambiguity in pointer-type power metering instrument data identification, this study proposes a unified framework that synergizes image preprocessing with deep recognition networks. By integrating a character detection module and a semantic comprehension module, we develop a robust method tailored for complex scenarios. The approach employs the Paddle-based optical character recognition (OCR) algorithm to automatically acquire the measurement range of power meters from images. Leveraging the U²-Net network model, it precisely segments the pointer and scale lines on the meter dial. The Hough line transform and circle center fitting methods are utilized to determine the rotational center of the pointer. The positional mask images are then transformed into rectangular coordinates, vertically stacked to form a 1D array, and analyzed through dual data-comparison strategies to calculate the relative position of the pointer on the scale lines, thereby completing meter data extraction. Experimental validations under real-world conditions demonstrate that the proposed method achieves an average recognition accuracy of 99% across multiple types of power metering. Compared to deep learning network approaches, it exhibits superior adaptability in complex environments, with a significant reduction in recognition errors.