Identifying specular highlights: Insights from deep learning

Specular highlights are the most important image feature for surface gloss perception. Yet, recognizing whether a bright patch in an image is due to specular reflection or some other cause (e.g., texture marking) is challenging, and it remains unclear how the visual system reliably identifies highlights. There is currently no image-computable model that emulates human highlight identification, so here we sought to develop a neural network that reproduces observers’ characteristic successes and failures. We rendered 179,085 images of glossy, undulating, textured surfaces. Given such images as input, a feedforward convolutional neural network was trained to output an image containing only the specular reflectance component. Participants viewed such images and reported whether or not specific pixels were highlights. The queried pixels were carefully selected to distinguish between ground truth and a simple thresholding of image intensity. The neural network outperformed the simple thresholding model—and ground truth—at predicting human responses. We then used a genetic algorithm to selectively delete connections within the neural network to identify variants of the network that approximated human judgments even more closely. The best resulting network shared 68% of the variance with human judgments—more than the unpruned network. As a first step toward interpreting the network, we then used representational similarity analysis to compare its inner representations to a wide variety of hand-engineered image features. We find that the network learns representations that are similar not only to directly image-computable predictors but also to more complex predictors such as intrinsic or geometric factors, as well as some indications of photo-geometrical constraints learned by the network. However, our network fails to replicate human response patterns to violations of photo-geometric constraints (rotated highlights) as described by other authors.