Abstarct: Today, most art, historical, and scientific products are produced on paper. Most libraries round the world are converting their resources into digital format for economical reasons. Storing such huge amount of digital information requires using efficient compression methods. In this thesis, we aim to propose a MRC-baxsed compression method for text images. Text images are usually decomposed into three laxyers of foreground, background, and mask. The mask laxyer is a binary image which determines the textual regions. The foreground laxyer shows color of textual regions and finally, the background laxyer includes graphical regions and line drawings. In thesis we created a text image databaxse which includes about 1449 text images. These images have various types of scxripts, qualities, and text-graphic combinations. We proposed and used a gray-scale-text-image binarization method to produce the mask laxyer which is baxsed on two ideas of combining classifiers (using neural networks) and multi-scale analysis (using filter bank). This proposed method has four advantages including adaptive computing a local threshold, using multi-scale analysis, combining classifiers, and low run time. The proposed method has a run time up to 50% lower than that of the Otsu's method. Also, due to MOS measure values collected using 50 human observers, the mean MOS value of the proposed method is 4.2 and for Otsu is 3.8. We proposed two methods for compression of grayscale text images produced after employing the proposed text binarization. First method is baxsed on KLT transform and the second one uses a proposed preprocessing and JPEG compression. The first compression method has compression ratio of 2.2 times (in average) relative to the JPEG and 1.6 relative to JPEG2000 in equal SNRs. The second proposed compression method produces better quality for reconstructed image than standard JPEG in equal bits per pixel. Also, in equal quality, this method has compression ratio of 1.5 times relative to the JPEG.