Camera system performance is a prominent part of many aspects of imaging science and computer vision. There are many aspects to camera performance that determines how accurately the image represents the scene, including measurements of colour accuracy, tone reproduction, geometric distortions, and image noise evaluation. The research conducted in this thesis focuses on the Modulation Transfer Function (MTF), a widely used camera performance measurement employed to describe resolution and sharpness. Traditionally measured under controlled conditions with characterised test charts, the MTF is a measurement restricted to laboratory settings. The MTF is based on linear system theory, meaning the input to output must follow a straightforward correlation. Established methods for measuring the camera system MTF include the ISO12233:2017 for measuring the edge-based Spatial Frequency Response (e-SFR), a sister measure of the MTF designed for measuring discrete systems.
Many modern camera systems incorporate non-linear, highly adaptive image signal processing (ISP) to improve image quality. As a result, system performance becomes scene and processing dependant, adapting to the scene contents captured by the camera. Established test chart based MTF/SFR methods do not describe this adaptive nature; they only provide the response of the camera to a test chart signal. Further, with the increased use of Deep Neural Networks (DNN) for image recognition tasks and autonomous vision systems, there is an increased need for monitoring system performance outside laboratory conditions in real-time, i.e. live-MTF. Such measurements would assist in monitoring the camera systems to ensure they are fully operational for decision critical tasks.
This thesis presents research conducted to develop a novel automated methodology that estimates the standard e-SFR directly from pictorial natural scenes. This methodology has the potential to produce scene dependant and real-time camera system performance measurements, opening new possibilities in imaging science and allowing live monitoring/calibration of systems for autonomous computer vision applications.
The proposed methodology incorporates many well-established image processes, as well as others developed for specific purposes. It is presented in two parts.
Firstly, the Natural Scene derived SFR (NS-SFR) are obtained from isolated captured scene step-edges, after verifying that these edges have the correct profile for implementing into the slanted-edge algorithm. The resulting NS-SFRs are shown to be a function of both camera system performance and scene contents. The second part of the methodology uses a series of derived NS-SFRs to estimate the system e-SFR, as per the ISO12233 standard. This is achieved by applying a sequence of thresholds to segment the most likely data corresponding to the system performance. These thresholds a) group the expected optical performance variation across the imaging circle within radial distance segments, b) obtain the highest performance NS-SFRs per segment and c) select the NS-SFRs with input edge and region of interest (ROI) parameter ranges shown to introduce minimal e-SFR variation. The selected NS-SFRs are averaged per radial segment to estimate system e-SFRs across the field of view. A weighted average of these estimates provides an overall system performance estimation.
This methodology is implemented for e-SFR estimation of three characterised camera systems, two near-linear and one highly non-linear. Investigations are conducted using large, diverse image datasets as well as restricting scene content and the number of images used for the estimation. The resulting estimates are comparable to ISO12233 e-SFRs derived from test chart inputs for the near-linear systems. Overall estimate stays within one standard deviation of the equivalent test chart measurement. Results from the highly non-linear system indicate scene and processing dependency, potentially leading to a more representative SFR measure than the current chart-based approaches for such systems. These results suggest that the proposed method is a viable alternative to the ISO technique.