Unrecognized lines using psm 3

[fuzzy7k]

Ever so frequently I will get a page where one line on the whole page is not recognized. I think I've tracked the problem to blob recognition, but don't know where to go from here. The attached images are of an index page and they are obtained using textord_tabfind_show_images. The line that is not recognized is the large red box in the one image, and the skinny orange box in the second image. What do these different color boxes mean and what can I do to make them big blue boxes?

[fuzzy7k]

I found the function that puts everything on the table, with regard to the scrollview blob debug window...
ccstruct/blobbox.cpp:

ScrollView::Color BLOBNBOX::TextlineColor(BlobRegionType region_type,
                                          BlobTextFlowType flow_type) {
  switch (region_type) {
    case BRT_HLINE:
      return ScrollView::BROWN;
    case BRT_VLINE:
      return ScrollView::DARK_GREEN;
    case BRT_RECTIMAGE:
      return ScrollView::RED;
    case BRT_POLYIMAGE:
      return ScrollView::ORANGE;
    case BRT_UNKNOWN:
      return flow_type == BTFT_NONTEXT ? ScrollView::CYAN : ScrollView::WHITE;
    case BRT_VERT_TEXT:
      if (flow_type == BTFT_STRONG_CHAIN || flow_type == BTFT_TEXT_ON_IMAGE)
        return ScrollView::GREEN;
      if (flow_type == BTFT_CHAIN)
        return ScrollView::LIME_GREEN;
      return ScrollView::YELLOW;
    case BRT_TEXT:
      if (flow_type == BTFT_STRONG_CHAIN)
        return ScrollView::BLUE;
      if (flow_type == BTFT_TEXT_ON_IMAGE)
        return ScrollView::LIGHT_BLUE;
      if (flow_type == BTFT_CHAIN)
        return ScrollView::MEDIUM_BLUE;
      if (flow_type == BTFT_LEADER)
        return ScrollView::WHEAT;
      if (flow_type == BTFT_NONTEXT)
        return ScrollView::PINK;
      return ScrollView::MAGENTA;
    default:
      return ScrollView::GREY;
  }
}

and some detailed description on what it all means...

// The possible region types of a BLOBNBOX.
// Note: keep all the text types > BRT_UNKNOWN and all the image types less.
// Keep in sync with kBlobTypes in colpartition.cpp and BoxColor, and the
// *Type static functions below.
enum BlobRegionType {
  BRT_NOISE,      // Neither text nor image.
  BRT_HLINE,      // Horizontal separator line.
  BRT_VLINE,      // Vertical separator line.
  BRT_RECTIMAGE,  // Rectangular image.
  BRT_POLYIMAGE,  // Non-rectangular image.
  BRT_UNKNOWN,    // Not determined yet.
  BRT_VERT_TEXT,  // Vertical alignment, not necessarily vertically oriented.
  BRT_TEXT,       // Convincing text.
  BRT_COUNT       // Number of possibilities.
};

// BlobTextFlowType indicates the quality of neighbouring information
// related to a chain of connected components, either horizontally or
// vertically. Also used by ColPartition for the collection of blobs
// within, which should all have the same value in most cases.
enum BlobTextFlowType {
  BTFT_NONE,           // No text flow set yet.
  BTFT_NONTEXT,        // Flow too poor to be likely text.
  BTFT_NEIGHBOURS,     // Neighbours support flow in this direction.
  BTFT_CHAIN,          // There is a weak chain of text in this direction.
  BTFT_STRONG_CHAIN,   // There is a strong chain of text in this direction.
  BTFT_TEXT_ON_IMAGE,  // There is a strong chain of text on an image.
  BTFT_LEADER,         // Leader dots/dashes etc.
  BTFT_COUNT
};

So, it thinks there is an image in there somehow and all I did to fix it was to bypass an if statement.

diff --git a/textord/colfind.cpp b/textord/colfind.cpp
index ea5d73d..3b4246e 100644
--- a/textord/colfind.cpp
+++ b/textord/colfind.cpp
@@ -309,7 +309,7 @@ int ColumnFinder::FindBlocks(PageSegMode pageseg_mode, Pix* scaled_color,
   stroke_width_->GradeBlobsIntoPartitions(
       pageseg_mode, rerotate_, input_block, nontext_map_, denorm_, cjk_script_,
       &projection_, diacritic_blobs, &part_grid_, &big_parts_);
-  if (!PSM_SPARSE(pageseg_mode)) {
+  if (!PSM_SPARSE(pageseg_mode) && 0) {
     ImageFind::FindImagePartitions(photo_mask_pix, rotation_, rerotate_,
                                    input_block, this, &part_grid_, &big_parts_);
     ImageFind::TransferImagePartsToImageMask(rerotate_, &part_grid_,

I think the `&& 0` should be replaced with an `&& init_var` and mainlined. Something like textord_imagefind. Any comments, suggestions?