|Ph.D Student||Berkovich Erez|
|Subject||Biologically Inspired Object Recognition|
|Department||Department of Biomedical Engineering||Supervisors||Professor Moshe Gur|
|Professor Emeritus Hillel Pratt|
|Full Thesis text|
Despite a remarkable progress in pattern recognition algorithms in recent years, these algorithms still do not compare with the human ability for object recognition. This recognition, performed intuitively and spontaneously, is one of the main functions of the human visual system. Mimicking the physiological solutions takes advantage of millions of years of evolution, natural selection, refining and fine-tuning of object recognition.
In this work, we first tackle the object categorization problem (e.g., face or clock? airplane or motorbike?) by using a biologically motivated recognition model. Low-level biologically motivated features of the image are extracted by appropriate filters, and converted to neuronal spike train rates. These spike trains are fed into a recognition module composed of a biological neural network which implements a neural microcircuit model, and connected to an artificial neural network which performs the classification. The experimental setup produces good categorization results while being consist with physiological mechanisms.
We then deal with the object identification problem (e.g., whose face?). We propose a new face recognition algorithm based on novel biologically motivated image features and a new learning algorithm, the Pseudo Quadratic Discriminant Classifier (PQDC). Our algorithm uses a complete ensemble of features: intensity, orientation and opponent-color channels shaped by appropriate filters at various scales and orientations. On the ensemble of these features we apply, iteratively, biologically-inspired Boosting. The biologically motivated algorithm yields better recognition results than those of the benchmarks, and the most informative features chosen for recognition are in agreement with physiology.
Finally, we demonstrate that our biologically motivated face recognition algorithm replicates the cross-race effect (the ability of people to recognize faces of their own race more easily than faces of other races). The results demonstrate that, like its biological inspiration, our algorithm's training process enhances the features most often experienced (same-race specific features). Consequently, the representational space is race adapted with improved same-race recognition. However the representation of other-race faces is affected, these faces are spanned by a sub optimal feature space which limits their recognition.