A dynamic masking technique for particle image velocimetry using convolutional autoencoders

Particle image velocimetry is an established technique for optical flow measurement, in which the local fluid velocity is derived from the motion of passive tracer particles. This motion is typically computed by a cross-correlation of small interrogation windows from subsequent image frames. Masking of objects within a velocimetry image is required to avoid correlation bias from interrogation windows overlapping the object. Manual masking efforts quickly become unpractical in the presence of moving or deforming objects and for high-speed recordings. Three methods are proposed to perform automatic dynamic image masking based on convolutional autoencoders, a type of artificial neural network. Promising results are achieved which suggests that neural-network-based PIV masking may be a valuable addition to the existing techniques.