Document layout analysis using recursive morphological transforms
Su Chen · 1996
The aim of this study is to apply solid statistical methods to systematically model and extract various layout structures on document images, such as words, text lines and text blocks. We first establish the computation theory of the recursive morphological transforms, namely the recursive erosion transform, the recursive dilation transform, the recursive opening transform, and the recursive closing transform. The transforms serve as a set of powerful tools for the document image shape analysis. Then we describe our efforts to construct a series of carefully ground-truthed document image databases, such as the UW English document image database (I). The database offers a platform based on which we can develop, train and evaluate our document layout analysis system. We present three sub-components of our document layout analysis system. They are the text skew estimation, the word segmentation, and the object spatial analysis: The text skew estimation finds the text skew angle of a document image. We develop an automatic text skew estimation algorithm using the recursive opening and closing transforms. It computes the estimated text skew angles which are within 0.5$\sp\circ$ of the true text skew angles with a probability of 0.95 on real images. The word segmentation detects all the words on a document image. We describe a word segmentation algorithm that utilizes the recursive closing transform. We derive the quantitative measures, such as the rates of miss, false, correct, splitting, merging and spurious detections, to evaluate its performance. The results show that the algorithm correctly detects the words on a document image at a rate of about 95%. The object spatial analysis treats the detected words as atomic and employs a probabilistic linear displacement model (PLDM) and an augmented PLDM model to model and extract the text lines and text blocks in a document image. By gathering statistics from a large population of document images, we are able to validate our models and determine the proper model parameters. The correct text line and text block detection rates are about 92% and 81% respectively.