Text detection pipeline

1. Partition image pixels using k-means
2. Extract connected components
3. For each connected component extract the following features:
1. Color histogram  and a measure of consistency
2. Stroke Width Transform histogram and a measure of consistency
3. Number of holes in a component
4. For every pair of connected components compute the following similarity metrics:
1. Color Similarity
2. Stroke Width Similarity
3. Pixel distance in x and y directions separately
4. Graph distance between every pair of elements (this means how many times you need to move from one component to a neighboring component to reach from component A to component B).
5. Merge component according to the similarity metrics. We have a hand-learned criteria.
6. Classify each component according to whether it looks like text or not. (We have a hand-set criteria.) Use the following features for text classification:
1. Area (the number of pixels)
2. Stroke Width Consistency
3. Width/Height
4. relative Stroke Length: Area/(Stroke_Width^2)
5. Sparsity
7. Display text regions together with their bounding boxes

Here are the outputs:
https://dl.dropboxusercontent.com/u/20022261/reports/text_segmentation_benchmark.html