Experimental studies of visual models in automatic image annotation
HCII'11 Proceedings of the 14th international conference on Human-computer interaction: design and development approaches - Volume Part I
Texture segmentation based on neuronal activation degree of visual model
ICONIP'12 Proceedings of the 19th international conference on Neural Information Processing - Volume Part V
| Hi-index | 0.00 |
How does the visual system form stable, coherent representations of image structure (edges, textures, shapes) from the coarse and noisy patterns of light collected at the retina? A common view is that neurons in the visual pathway act as feature detectors, with a hierarchy of increasingly complex features represented in V1, V2, and higher cortical areas. This approach has defined most experimental and modeling work to date (and inspired much computer vision research). However, it fails when applied to natural scenes, where object boundaries do not always produce clear edges, and surface properties like texture are defined by their intrinsic variability rather than fixed configurations of shapes. Models formulated in terms of feature processing also fail to account for a large number of subtle behaviors exhibited by neurons in the visual cortex. In this dissertation we develop an alternative theory: rather than encoding preferred features, neurons describe entire distributions over their inputs, and thus capture the patterns of variability that underlie textures, contours, and other image elements. This allows the neural code to represent more abstract aspects of the image and remain invariant across fixations within local regions. We develop hierarchical models that implement this idea, and show that they yield better statistical descriptions of natural images than standard unsupervised learning techniques. The proposed models use distributed representations of image structure, a strategy likely employed in the brain. Although we do not fit the models to neural data, they exhibit a number of classical properties of "complex cells" in V1, as well as more subtle effects observed in V2 and V4. These results thus provide the first functional account for several previously unexplained neural behaviors. Finally, we demonstrate how model encoding of natural images can be used to analyze data from physiological experiments and predict neural responses to novel stimuli.