This repository serves to consolidate all of the scripts required to operationalize the digitization process of New York City's 19th century directories.

As of February 18, 2020, a workflow for the preprocessing of the images has been developed with the crop_for_columns.py script. The present workflow will crop out images to just the columns and then perform an affine transformation which effectively deskews and rotates the image.

Some directories were scanned in as single-channeled grayscale images. This will cause issues when converting from PIL images to OpenCV images directly as a batch under the assumption that all scanned images have three channels. The image pre-processing step has accounted for this when deskewing.

Using the other processing scripts created by @nmwolf, TSVs containing entries have been generated. We are currently at the final post-processing stage where entries are being cleaned and Tesseract text-recognition errors corrected.

-psm stands for Page Segmentation Modes. This page explains in detail how it works but basically there are 14 options as listed below:

Page Segmentation Modes:
0 Orientation and script detection (OSD) only.
1 Automatic page segmentation with OSD.
2 Automatic page segmentation, but no OSD, or OCR.
3 Fully automatic page segmentation, but no OSD. (Default)
4 Assume a single column of text of variable sizes.
5 Assume a single uniform block of vertically aligned text.
6 Assume a single uniform block of text.
7 Treat the image as a single text line.
8 Treat the image as a single word.
9 Treat the image as a single word in a circle.
10 Treat the image as a single character.
11 Sparse text. Find as much text as possible in no particular order.
12 Sparse text with OSD.
13 Raw line. Treat the image as a single text line, bypassing hacks that are Tesseract-specific.

Changing any of the Page Segmentation Nodes options will not signficantly improve the performance of the Tesseract. It appear that -psm = 5 is designed with Asian languages such as Chinese, Japanese, and Korean in mind.
Most of the options describe text formats that are unlike those in the directories. OSD will not be helpful because deskewing and rotation will already have taken place and the language of the directory is English. There might be non-English last names such as Núnez for example, but they are not significantly different from English to warrant changing the psm modes.

-oem stands for OCR Engine Mode. Tesseract's own readme is not particularly helpful. This page provides more information on the OEM options available.

OCR Engine Modes:
0 Legacy engine only.
1 Neural nets LSTM engine only.
2 Legacy + LSTM engines. (default)
3 Default, based on what is available.

Of these four options, the default -oem option appears to be the best.

In other words, Tesseract's psm and oem options might not be suitable in solving the issue of columns being skipped. The deskew and image rotation might have solved the problem by aligning the starting point of each line.

1. Pre-crop processing: Edge detection, dilation etc.

2. Post-crop processing: Deskew and image rotation.

3. Tesseract -oem and -psm values: 2 and 3 respectively

4. hocr-detect columns?

5. CRF extracted features and amount of training data

1. qc_json_check.py

Compares the json-format labeled output of the cropped jpeg - to - labeled entries workflow using validation json located in the "hand-corrections" folder of each NYC directory folder.

The output is a score between 0 and 1, with 1 an identical match in a txt diff of two entry pages and 0 having no matches.

Input parameters are -in and -test, where "in" is the directory (with trailing /) containing all _labeled.json files and "test" is the "hand-corrections" directory containing all handmade validation json files.

2. crop_for_columns.py

Crops the downloaded original jpegs in preparation for Tesseract.

Uses three parameters, -type, -in, and -out, where "in" is the directory containing the jpegs, "out" is the directory where cropped jpegs should be placed (for us, "cropped" folder), and -type can be minimal of full, depending on whether an output of the cropped jpeg should be created for inspection.