Using the 5th dimensions of human visual perception to inspire automated edge and texture segmentation: A fuzzy spatial-taxon approach

Abstract

With the recent stunning success of machine learning, artificially intelligent machine vision research falls (roughly) into two camps: the big data camp and cognitive informatics camp. Big data uses statistical methods to discover latent structures that emerge from the co-occurrences of relevant features when sampling over enormous quantities of data. The cognitive informatics methods design computer vision systems to mimic human cognition. Though some visual latent features that emerge from deep learning networks, mimic mammalian visual detectors, as of yet the information processing mechanisms (analogous to human psychophysical mechanisms) remain hidden within the complexity of the deep nets. Furthermore, the sampling requirements of big data systems require limiting samples to pre-processed sets, such as SHIFT (shift invariant feature transform (Lowe 1999)). Techniques, such as the ones introduced in this paper, provide fast cognitively relevant methods for selecting samples and reducing the number of candidate features. The approach described in this paper live squarely in the camp of designing computer vision A.I. to mimics human cognitive processes. I introduce a novel definition of edges, based on human hierarchical scene perception. Hierarchical scene perception views vision within the 5 dimensions of horizontal & vertical position, depth, time and scene abstraction level (spatial-taxon). Fuzzy inference selects candidate edge elements using the Gestalt psychology principal of good curvilinear continuation, proximity and edges attachment. Spatial-taxon inference infers an edge outline for each level of abstraction within the scene architecture. The system was tested on 60 natural images and the results provide edges more aligned with human intuition of what edges should look like. ROC plots indicate solid performance, with the majority of human subjects rating the edge detection as high quality. The inferred edges are consistent with the finding of neurons responsive to proto-object boundaries in the visual cortex.