Methods used to create the “A combined photo-index for the Defense Intelligence Agency photo archive”


The DIA compiled photo-indices by hand drawing the ground track for individual missions on one or more 1 x 1 degree map sheets (2 x 1 degree map sheets above 65 degree north). These sheets were subsequently grouped by latitude-longitude pair, and captured on microfilm along with additional materials indicating the location of the index , a table summarizing the group contents and ancillary materials. The microfilm was then scanned. Index sheets follow a basic format:

with mission information in a header, ground tracks with a subset of photos drawn in,  and four fiducials that indicate the corners of the 1 x 1 (or 2 x 1) subset. The position of the fiducials vary in a predictable way by latitude but their positions are also a function of the individual drawing the map and the scanning setup. A combination of template matching identification of the fiducials and statistical modelling of the positions of the fiducials as a function of other identifiable points and the latitude of the index made it possible to create approximate locations for each fiducial.

Individual indices have their location indicated in the header, but these are hand written and of varying format. To associate each index with a latitude/longitude pair, we needed to identify all of the “title sheets” – typewritten cover pages for each latitude-longitude group- and determine which of the subsequent sheets were indices. Summary texture indices were generated from image thumbnails of the scans and a training dataset was used to classify each into one of several types. Classification accuracy was 98% but there are 285,000 sheets which suggests that we could expect roughly 5700 mis-classifications. Therefore, every sheet was reviewed to confirm its classification.

Example sequence of the dataset 


After reviewing the classification, all the photo-indices for a given lat-lon pair were combined into a single summary image using the fiducial positions as a guide, and the image was cropped, scaled and mosaiced into a single image.