Optical holographic data storages (HDS) with high theoretical-capacity have been researched for more than two decades. Among them, coaxial HDS receives the most attention. Amplitude-based coding in coaxial HDS systems is fundamental and generally employs a 3:16 modulation scheme that sets merely 3 non-adjacent bright pixels within a 4*4 pixel-block, called as Three-Level Format (TLF). Unfortunately, TLF data frame recorded upon the disc is difficult to be retrieved accurately in the practice. In our previous work, we experimentally analyzed inaccurate TLF data frames recognition and its corresponding factors. The core issue is how to accurately and fast locate TLF data frame in encoded image. Thus, we have to use some image processing techniques such as Gaussian blur to assist in locating TLF data frame. However, using image processing techniques will inevitably introduce long processing-time. Therefore, we propose an efficient two-stage decoding scheme. In the first stage, we locate TLF data frame in encoded images and calculate scaling multiplier (about 1.9s). In the second stage, we calculate the coordinates of all data points and read them (about 0.08s). For the first image of a batch, we execute complete two-stage process, but the subsequence images are no longer to perform the first stage and only fine-tuning TLF data frame location. The experimental results show that the average latency of reading an image is reduced by 1.9s. The average data point error rate is 3.1%, and the average data block error rate is 7.8%, consistent with the results of executing complete process.