Abstract
Models of spatial contrast detection have a wide range of applications in vision science and machine vision, including display measurements, feature detection, letter identification, Standard Spatial Observer, and retina implants. This wide range of applications has always inspired vision scientists to develop more accurate models. In 2005, Watson and Ahumada introduced a computational model which was the most accurate model at its time. Their model could achieve the RMS error of 0.79 dB using Gabor Filter Banks. However, its execution time (20 seconds/43 stimuli of ModelFest Dataset) and computational load were so high that Watson and Ahumada proposed the omission of spatial frequency channels for reducing the model's required runtime for practical purposes. While this measure decreases runtime, it causes a significant increase in the RMS error. Further, the existence of frequency channels is mandatory for developing a more advanced and accurate model using Frequency Dependent Aperture Effect (FDAE). This need led us to develop a new model based on Wavelet Filter Banks that gives us the advantage of speed and capability to process each frequency channels' outputs. Using Wavelet Filter Banks and FDAE which were unprecedented on this dataset, we could achieve the as yet lowest RMS error as low as 0.68 and timing performance of 1 sec/43 stimuli. This model benefits from simplicity, efficiency and accuracy, which makes it suitable for cheaper, lighter and smaller hardware implementation; hence, it can be used in wide range of practical uses, such as online tests, real-time processes, an improved Standard Observer, and retina prostheses. Moreover, our model has also shown a better timing and error performance compared to one of the most recent models in this field, "Retina-V1". We also introduced a new formula for modeling contrast sensitivity function, with better timing performance and less error.
Meeting abstract presented at VSS 2017