Textual Image Compression for Maintaining or Improving the Recognition Performance

The present study investigates compression of textual images with high compression ratios while preserving or improving the general quality, readability, and optical character recognition of the compressed textual images. A novel textual image compression/decompression approach is proposed in which the compression path includes dynamic range reduction, wavelet transform, and set partitioning in hierarchical trees (SPIHT) encoding. The decompression path employs SPIHT decoding, then inverse wavelet transform, and then the proposed image enhancement technique. The compression and recognition performances of the proposed approach are evaluated using quantitative and qualitative measures that are then compared to those of the JPEG2000, DjVu, and multi-dimensional multi-scale parser approaches. In addition to the conventional rate-distortion curve, mean opinion score (MOS) is used and the novel measures of "breakdown point" and "downfall slope" are defined. The quantitative and qualitative results of the proposed approach have achieved results similar to those of the peak signal-to-noise ratio, but considerably outperformed the other two approaches for average MOS, average recognition rate, and the newly defined measures.