Data-driven multiresolution camera using the foveal adaptive pyramid

Abstract

There exist image processing applications, such as tracking or pattern recognition, that are not necessarily precise enough to maintain the same resolution across the whole image sensor. In fact, they must only keep it as high as possible in a relatively small region, but covering a wide field of view. This is the aim of foveal vision systems. Briefly, they propose to sense a large field of view at a spatially-variant resolution: one relatively small region, the fovea, is mapped at a high resolution, while the rest of the image is captured at a lower resolution. In these systems, this fovea must be moved, from one region of interest to another one, to scan a visual scene. It is interesting that the part of the scene that is covered by the fovea should not be merely spatial, but closely related to perceptual objects. Segmentation and attention are then intimately tied together: while the segmentation process is responsible for extracting perceptively-coherent entities from the scene (proto-objects), attention can guide segmentation. From this loop, the concept of foveal attention arises. This work proposes a hardware system for mapping a uniformly-sampled sensor to a space-variant one. Furthermore, this mapping is tied with a software-based, foveal attention mechanism that takes as input the stream of generated foveal images. The whole hardware/software architecture has been designed to be embedded within an all programmable system on chip (AP SoC). Our results show the flexibility of the data port for exchanging information between the mapping and attention parts of the architecture and the good performance rates of the mapping procedure. Experimental evaluation also demonstrates that the segmentation method and the attention model provide results comparable to other more computationally-expensive algorithms.