Methods

Our goal is to segment a 512x512 pixel image into three classes: text, picture, and background.  The image is divided into 256 32x32 blocks, and each of these blocks are classified.  If the 32x32 blocks cannot be classified, the block size is changed to achieve accurate segmentation.

For each 32x32 block, we transform the block using a Daubechies wavelet to one level.  We make a vector of the wavelet coefficients in the LH and HL bands to classify the block.  If the variance of this vector is zero, the block is classified as background.  If the variance approaches extreme values, it is classified as text or picture; otherwise, the block is marked as undetermined.

The undetermined block is then divided into four 16x16 subblocks.  We then try to classify each of these blocks individually as either text, background, picture, or undetermined.  If the subblock is undetermined, we use context information to classify the subblock.  If the subblock has adjacent text and picture blocks, the block is at a boundary.  Classification cannot be improved by merging the block with surrounding blocks since the merged big block would contain mixed types. In this case, we classify the subblock as either as text or picture.  If the subblock doesn't have adjacent text and picture blocks, we assume that it is unclassifiable because the sample size is too small.  Thus, we merge the eight surrounding 32x32 blocks with the block and produce a block of size 96x96.  The subblock is then classified according to the variance of the 96x96 block.