Biologically Inspired Vision Architectures: a Software/Hardware Perspective

Even tough the field of computer vision has seen huge improvement in the last few decades, computer vision systems still lack, in most cases, the efficiency of biological vision systems. In fact biological vision systems routinely accomplish complex visual tasks such as object recognition, obstacle avoidance, and target tracking, which continue to challenge artificial systems. The study of biological vision system remains a strong cue for the design of devices exhibiting intelligent behaviour in visually sensed environments but current artificial systems are vastly different from biological ones for various reasons. First of all, biologically inspired vision architectures, which are continuous-time and parallel in nature, do not map well onto conventional processors, which are discrete-time and serial. Moreover, the neurobiological representations of visual modalities like colour, shape, depth, and motion are quite different from those usually employed by conventional computer vision systems. Despite these inherent difficulties in the last decade several biologically motivated vision techniques have been proposed to accomplish common tasks. For example Siagian & Itti [14] developed an algorithm to compute the gist of a scene as a low-dimensional signature of an image, in the form of an 80-dimensional feature vector that summarizes the entire scene. The same authors also developed a biologically-inspired technique for face detection [13]. Interesting results have also been reported in generic object recognition and classification (see for example [15] [16] [12] [11]). Also on the sensor side the biological vision systems are amazingly efficient in terms of speed, robustness and accuracy. In natural systems visual information processing starts at the retina where the light intensity is converted into electrical signals through cones and rods. In the outer layers of the retina the photoreceptors are connected to the horizontal and bipolar cells. The horizontal cells produce a spatially smoothed version of the incoming signal while the bipolar cells are sensitive to the edges in the image. Signals output from the cells are then used for higher level processing. Several architecture have been proposed to mimic in part the biological system and to extract information ranging from low to high level. For example Higgins [10] proposed a sensor able to perform an elementary visual motion detector. Other researchers proposed sensor to detect mid-level image features like corners or junctions [4] or even to perform higher level tasks such as tracking [6] or texture classification [5]. Robotics represents a typical field of application for hardware implementations of biologically inspired vision architectures.