An investigation of image-based task performance prediction


Human task performance with imaging sensors is characterized by perception experiments involving ensembles of observers viewing an ensemble of task relevant images from real sensors. Summary statistics from perception experiments are used, along with detailed descriptions of the sensors and early human vision processes to build predictive models such as NV-IPM. Use of these models typically requires knowledge of more than 100 specific parameters regarding the sensor, the viewing conditions, and the task. In this research we seek to do a blind prediction of task performance using task relevant image ensembles and image processing operations that produce statistically similar outputs to those obtained in real human perception experiments. We restrict our investigation to the task of identifying tracked vehicles. The data we seek to replicate through image processing are similarity matrices derived from the confusion matrices of actual perception experiments. This paper updates our work to date examining primarily the correspondence between several image processing approaches and perception data. © 2013 SPIE.

Publication Title

Proceedings of SPIE - The International Society for Optical Engineering