Abstract
Researchers have been seeking to develop a computational model that accounts for lightness/brightness perception in humans. Spatial filtering models have been accepted as a promising approach (e.g., Blakeslee & McCourt, 1999), and a novel lightness model, Markov illuminance and reflectance (MIR), was recently developed to address this issue (Murray, 2020). MIR is a Bayesian model that employs a probabilistic method called a Markov random field to represent observers’ prior beliefs about illuminance and reflectance. This model successfully accounts for some lightness phenomena that extant models fail to predict correctly (e.g., an effect of articulation on simultaneous lightness contrast). It thus appears a promising approach to this research topic. The present study seeks to improve the performance of MIR. First, we modified the inference process of MIR, implemented using the belief propagation method. Because some links in the Markov random field were left unused in the original implementation, we modified it to fully utilize all prepared links. This modification allowed the model to perform a more efficient search for solutions and improved its predictive performance for finely articulated images (i.e., checkerboard assimilation and pixel-wise noised images). Moreover, we also modified a constraint for X junctions. Although the original MIR considered all X junctions to be evidence of illumination edges, they are unlikely to cue an illumination change when two neighboring edges at an X junction show opposing contrast polarities. We find that a modification to the constraint regarding X junctions enabled the model to precisely identify a shadow on a checkerboard. The modified models also replicated all successful predictions using the original model that were reported by Murray (2020). The present study shows high potential for MIR and its ability to be extended to accommodate further improvements.