Biological vision systems inspire processing methods in computer vision applications. This paper employs the insights of vision systems in hardware and presents a pixel-parallel, reconfigurable, and layer-based hierarchical architecture for smart image sensors. The architecture aims to bring computation close to the sensor to achieve high acceleration for different machine vision applications while consuming low power. We logically divide the image into multiple regions and perform pixel-level and region-level processing after removing spatiotemporal redundancy. Those processors use bio-inspired algorithms to activate the regions with region of interest of a scene. The hierarchical processing breaks the traditional sequential image processing and introduces parallelism for machine vision applications. Also, we make the hardware design reconfigurable even after fabrication to make the hardware reusable for different applications. Simulation results show that the area overhead and power penalty for adding reconfigurable features stay in an acceptable range. We emphasize to maximize the operating speed and obtain 800 MHz. Besides, the design saves 84.01% and 96.91% dynamic power at the first and second stages of the hierarchy by removing redundant information. Furthermore, the sequential deployment of high-level reasoning only on the selected regions of the image becomes computationally inexpensive to execute a complex task in real time.

生物视觉系统激发了计算机视觉应用中的处理方法。本文利用了视觉系统在硬件中的见识，并提出了一种用于智能图像传感器的像素并行，可重新配置和基于层的分层体系结构。该架构旨在使计算接近传感器，从而在消耗低功耗的同时为不同的机器视觉应用实现高加速。我们在逻辑上将图像划分为多个区域，并在去除时空冗余后执行像素级和区域级处理。这些处理器使用受生物启发的算法来激活具有场景感兴趣区域的区域。分层处理打破了传统的顺序图像处理，并为机器视觉应用引入了并行性。还，即使在制造后，我们也使硬件设计可重新配置，以使硬件可用于不同的应用程序。仿真结果表明，增加可重配置功能所需的面积开销和功率损失均在可接受的范围内。我们强调最大程度地提高工作速度并获得800 MHz。此外，该设计通过删除冗余信息，在层次结构的第一阶段和第二阶段节省了84.01％和96.91％的动态功耗。此外，仅在图像的选定区域上进行高级推理的顺序部署对于实时执行复杂任务在计算上不昂贵。我们强调最大程度地提高工作速度并获得800 MHz。此外，该设计通过删除冗余信息，在层次结构的第一阶段和第二阶段节省了84.01％和96.91％的动态功耗。此外，仅在图像的选定区域上进行高级推理的顺序部署对于实时执行复杂任务在计算上不昂贵。我们强调最大程度地提高工作速度并获得800 MHz。此外，该设计通过删除冗余信息，在层次结构的第一阶段和第二阶段节省了84.01％和96.91％的动态功耗。此外，仅在图像的选定区域上进行高级推理的顺序部署对于实时执行复杂任务在计算上不昂贵。