Real-time Neuromorphic Visual Pre-Processing and Dynamic Saliency

The human brain is by far the most computationally complex, efficient, and reliable computing system operating under such low-power, small-size, and light-weight specifications. Within the field of neuromorphic engineering, we seek to design systems with facsimiles to that of the human brain with means to reach its desirable properties. In this doctoral work, the focus is within the realm of vision, specifically visual saliency and related visual tasks with bio-inspired, real-time processing. The human visual system, from the retina through the visual cortical hierarchy, is responsible for extracting visual information and processing this information, forming our visual perception. This visual information is transmitted through these various layers of the visual system via spikes (or action potentials), representing information in the temporal domain. The objective is to exploit this neurological communication protocol and functionality within the systems we design. This approach is essential for the advancement of autonomous, mobile agents (i.e. drones/MAVs, cars) which must perform visual tasks under size and power constraints in which traditional CPU or GPU implementations to not suffice. Although the high-level objective is to design a complete visual processor with direct physical and functional correlates to the human visual system, we focus on three specific tasks. The first focus of this thesis is the integration of motion into a biologically-plausible proto-object-based visual saliency model. Laurent Itti, one of the pioneers in the field, defines visual saliency as ``the distinct subjective perceptual quality which makes some items in the world stand out from their neighbors and immediately grab our attention.'' From humans to insects, visual saliency is important for the extraction of only interesting regions of visual stimuli for further processing. Prior to this doctoral work, Russel et al. \cite{russell2014model} designed a model of proto-object-based visual saliency with biological correlates. This model was designed for computing saliency only on static images. However, motion is a naturally occurring phenomena that plays an essential role in both human and animal visual processing. Henceforth, the most ideal model of visual saliency should consider motion that may be exhibited within the visual scene. In this work a novel dynamic proto-object-based visual saliency is described which extends the Russel et. al. saliency model to consider not only static, but also temporal information. This model was validated by using metrics for determining how accurate the model is in predicting human eye fixations and saccades on a public dataset of videos with attached eye tracking data. This model outperformed other state-of-the-art visual saliency models in computing dynamic visual saliency. Such a model that can accurately predict where humans look, can serve as a front-end component to other visual processors performing tasks such as object detection and recognition, or object tracking. In doing so it can reduce throughput and increase processing speed for such tasks. Furthermore, it has more obvious applications in artificial intelligence in mimicking the functionality of the human visual system. The second focus of this thesis is the implementation of this visual saliency model on an FPGA (Field Programmable Gate Array) for real-time processing. Initially, this model was designed within MATLAB, a software-based approach running on a CPU, which limits the processing speed and consumes unnecessary amounts of power due to overhead. This is detrimental for integration with an autonomous, mobile system which must operate in real-time. This novel FPGA implementation allows for a low-power, high-speed approach to computing visual saliency. There are a few existing FPGA-based implementations of visual saliency, and of those, none are based on the notion of proto-objects. This work presents the first, to our knowledge, FPGA implementation of an object-based visual saliency model. Such an FPGA implementation allows for the low-power, light-weight, and small-size specifications that we seek within the field of neuromorphic engineering. For validating the FPGA model, the same metrics are used for determining the extent to which it predicts human eye saccades and fixations. We compare this hardware implementation to the software model for validation. The third focus of this thesis is the design of a generic neuromorphic platform both on FPGA and VLSI (Very-Large-Scale-Integration) technology for performing visual tasks, including those necessary in the computation of the visual saliency. Visual processing tasks such as image filtering and image dewarping are demonstrated via this novel neuromorphic technology consisting of an array of hardware-based generalized integrate-and-fire neurons. It allows the visual saliency model's computation to be offloaded onto this hardware-based architecture. We first demonstrate an emulation of this neuromorphic system on FPGA demonstrating its capability of dewarping and filtering tasks as well as integration with a neuromorphic camera called the ATIS (Asynchronous Time-based Image Sensor). We then demonstrate the neuromorphic platform implemented in CMOS technology, specifically designed for low-mismatch, high-density, and low-power. Such a VLSI technology-based platform further bridges the gap between engineering and biology and moves us closer towards developing a complete neuromorphic visual processor