Domain Specific Video Compression

Common lossy video compression methods have already reached a very high performance level and are about to reach their natural limit. However, their compression efficiency is still insufficient for certain domains where a plethora of video data should be archived.
In the domain of medical endoscopy, where entire surgeries are recorded for documentation with high visual quality, we achieved important steps to lower the bitrate substantially.

Currently, doctors can only record segments of endoscopic surgeries. This leads to losing potentially important data, which has not been recorded, but could be significant in the future.

We aim to record entire surgeries, develop a compression method to lower the used bitrate and store in a more efficient way. With the defined four dimensions, we can improve the domain specific video compression.


In endoscopic surgery videos, the content is inside a circle area and surrounded by a noisy border. The circle of the content is dynamic in size (zoom in and out) and position.
We developed an algorithm to detect this circle by using lines to locate points on the border to the noise. After that, we use an RANSAC-Algorithm that takes all combinations of three points and calculates possible circles. Each point is getting a score depending on the probability that it will be located on the noisy border. Finally, the three points with the highest score will be used to calculate the final circle.

The original video has a noisy border and a size of 15,0 MB.

The border of the modified video is covered with black and has a size of 9,84 MB.

Relevance Segmentation

Within endoscopic surgery videos, there are many scenes, which are irrelevant for later
evaluation (e.g. recording still on but endoscope is turned off; out of patient recordings; dark or blurry frames). We aim to find those irrelevant scenes and present them to the doctor, which then can decide to delete, keep or store them in lower quality. To detect out of patient scenes we use the HSV (hue, saturation, value) – color space. For Blurriness detection, we use the “Difference of Gaussian” with two different parameters for the filters.


We made a study including 37 participants (18 resident surgeons and 19 surgeons) to investigate first how strong we can compromise videos without doctors recognize a quality loss. We did it by giving doctors two videos without letting him know which one is the original. Switching between the videos is possible.
In the second part of our study, we investigated how strong we can compromise videos until the quality is just sufficient. This time the doctors had one video and rated the quality on a scale.
Finally, we found out that the compression is information lossless until 8Mbit (instead of 20), good quality of compressed videos can be downscaled to 2.5Mbit and sufficient video quality to 1.4 Mbit.

Long Term

The older the videos are, the less important they get. Older videos can be classified in a chronologic manner and stored at the beginning in full quality. After a set amount of time, we can reduce the quality and safe important storage place for newer operation videos. If we do so, using our above described compressing methods, we can downscale the needed amount of data volume from 100% to 8%.