Computer Science Building, Room 151
Computer vision researchers design intelligent machines capable of perceiving visual information. One of many approaches is to model the fascinating visual processing done by primates. The eyes, the brain, and their interconnections provide a rich set of visual perceptions without conscious effort. How primates "see" has long been an intriguing research question in neuroscience, biology, physiology, and psychology. Based on discoveries in these fields, computer scientists have attempted to apply biologically plausible models to the design of visual machines. While still inferior to primates, the performance of existing systems has been demonstrated to be quite promising. It appears that a machine based on biological principles can be simple yet achieve high accuracy and generalize well.
This thesis takes a closer look at an artificial vision system based on biological principles, specifically with respect to object recognition and categorization. An existing system will be investigated, followed by various modifications leading to a new and improved system. The proposed model employs unsupervised feature learning, simulating a hypercolumn of the primary visual cortex, a hierarchical feed-forward framework, mimicking simple and complex cells, and finally neural network classification, based on a computational model of interconnected neurons. Compared to existing approaches, the proposed system is not only more biologically inclined, but also more effective. With significantly shorter running times it achieves good accuracies on several data sets compared to other state-of-the-art systems. Another key feature is good generalizability. The proposed approach does not rely on delicate domain specific segmentation procedures often employed by traditional vision systems. There are few parameters and a single set of parameters can be applied to images from different domains without significant loss of accuracy.
In order to evaluate the system, experiments are conducted both on a commonly used collection of natural scenes as well as a challenging collection of realistic underwater marine images. The latter data is a rather uncommon data source in the computer vision community. Although automated labeling of various species of tiny planktonic organisms has recently attracted the attention of a number of researchers, biologically inspired vision systems had not previously been applied to this kind of data. Despite domain specific difficulties, such as low image quality, and high diversity of shapes and motions, the potential of the proposed system is shown to be quite promising. Given the demonstrated performance, there is a high likelihood that such a system could substantially facilitate research of the planet's ocean ecosystems. As plankton plays an important role in the carbon cycle, high-volume automated labeling of planktonic specimens may lead to new insights concerning global climate changes. The encouraging experimental results in the marine domain bode well for future application of the proposed approach to other domains currently not considered by mainstream computer vision applications.
Advisor: Allen Hanson