Piotr Fryc
supervisor: Barbara Siemiątkowska, Anna Ostaszewska-Liżewska
Television Image Sensor as the input element of a TV camera system needs to be driven in a way that achieves best output signal quality. There are few parameters to be driven e.g exposition time. Achieving correct sensor parameters is no trivial task. There are numerous algorithms in use that base on image statistics.
In my study I explore the use of AI methods to drive the image sensor parameters. The goal is to achieve correct image in numerous scene and light conditions. For this, the strict definition of the image correctness should be adopted. The work is conducted using the Python PyTorch package. PyTorch capabilities were extended and published as open source for the purpose of dealing with video datasets. To gain understanding of a sequence of events in the video, the extension extracting optical flow from video will be prepared. To achieve suitable and repeatable conditions for testing, the TV camera simulator is being prepared as a Python package. AI methods such as 3D-ConvNet, Two-Stream (with Optical Flow data) are to be evaluated.
The plan is to publish a paper about using fully functional software TV camera simulator as a testbench for family of algorithms driving image sensor parameters. The simulator itself will be published as an open source Python package. Series of known algorithms (conventional and AI based) will be evaluated.
The next step is to develop my own, AI based methods of driving image sensors, evaluate them using the simulator and publish.