To collect handwritten Urdu word images, we use forms containing well separated boxes. Each form has 20-25 pairs of boxes. One box contains a reference word image and another box to write the shown word image in each pair. The writers (who can read and write the language) with different educational backgrounds and ages can write the shown words inside the white box. The writers are free to use pens of their choice and write the words in their natural style. There is no constraint to write except “write within the white box.” There are 8 writers (who can read and write Urdu) to write the words. The collected pages are digitized using a flatbed scanner with a resolution of 600 DPI in color and stored in JPEG format. Using morphological operations, we can extract and label the handwritten word images. Through this procedure, we create this dataset having handwritten Urdu word images and corresponding ground truth transcriptions. This dataset consists of 1,020,630 word images and their corresponding ground truth transcriptions. We divide this dataset into Training, Validation and Test Sets consisting of 71,207, 13,906, and 15,517 word images and their corresponding ground truth transcriptions, respectively. There are 11,936 unique Urdu words in the training set.

Training Set:

train.zip contains folder named “images” with 71,207 word level images, “train_gt.txt” file containing image name and ground truth transcription separated by “Tab space”, and “vocabulary.txt” contains list of 11,936 unique words in the Training set.

Validation Set:

val.zip contains folder named “images” with 13,906 word level images, and “val_gt.txt” containing image name and ground truth text separated by “Tab space”.

Test Set:

test.zip contains folder named “images” with 15,517 word level images, and “test_gt.txt” containing image name and ground truth text separated by “Tab space”.