Especially in scientific applications, the amount of image data is large and retention times are long, increasingly straining the IT systems and server load of research institutions. "Big Data is a current problem that is widely discussed in the field of scientific imaging. It is fueled by three developments: Higher frame rates, higher resolution, and the demand to take images that are as three-dimensional as possible," confirms Dr. Gerhard Holst, Head of Science & Research at PCO. "In general, image data can be compressed, but in approaches pursued so far, only after the image has been generated, either lossy or with a low compression rate." As a manufacturer of high-end camera systems, PCO found Swiss-based Dotphoton, which specializes in image compression for critical applications and AI, while exploring new ways to compress image data.
Dotphoton's Jetraw solution starts before the image is created and uses knowledge of the noise behavior of the camera detectors to efficiently compress image data. The origins of Swiss image data compression go back to research questions in quantum physics. "In experimental setups with CCD/CMOS sensors to quantify entropy and the relationship between signal and noise, it was found that even with very good detectors, most of the entropy is noise. For a 16-bit sensor, we typically detected 9bits of entropy that was purely due to noise and only 1bit that came from signal," explained Bruno Sanguinetti, CTO and co-founder of Dotphoton. "One insight from our observations is that good detectors literally 'zoom in' on the noise."
Dotphoton provides evidence that when they compress by up to a factor of ten, the image data suffers no loss of information. Specifically, Dotphoton uses the detector's own temporal as well as spatial noise information for their noise compression. The specific measurement values of the camera are therefore an input prerequisite, which from Gerhard Holst's point of view results in a welcome synergy effect. PCO is a long-standing supporter of the image processing standard EMVA 1288, which is used to determine quality parameters of cameras in order to compare them with cameras from other manufacturers. In this context, the measurement data required for EMVA 1288 largely match the necessary parameters of the Dotphoton software and are accordingly already available on every PCO camera anyway.
Currently, the compressed image data in the PCO system is stored in the recorder module of the SDK and decompressed again by the Dotphoton software for processing. A full integration of the Dotphoton software into the camera would be a logical integration step, with the camera transferring only the already compressed data. "The ultimate benefit would be the integration of the compression into the FPGA," enthuses Gerhard Holst.
According to Dotphoton, this is not far off. Initial feedback from customers of PCO has been quite positive, and further inquiries, for example from the field of particle image velocimetry, are currently being examined. An expansion to high-speed imaging areas outside the purely scientific environment with equally high requirements for data archiving, such as crash tests, is also conceivable.
Regardless of this, Dotphoton's development of Jetraw has also been driven by the increased usage of image data in AI applications. For example, Bruno Sanguinetti points out that in recent years, the number of images generated purely for analysis purposes has increased dramatically.
The camera calibration information contained in Jetraw images is currently used only for compression. However, they could be used to improve other image processing tasks at the same time, such as allowing machine learning to work efficiently with images from different sources. Currently, Dotphoton is integrating various software packages and programming languages so that users can load and save Jetraw images directly.