Lapped convolutional neural networks for embedded systems | IEEE Conference Publication | IEEE Xplore

Lapped convolutional neural networks for embedded systems


Abstract:

Convolutional neural network (CNN) has achieved numerous breakthroughs in many artificial intelligent applications. However, its complexity is quite high and usually requ...Show More

Abstract:

Convolutional neural network (CNN) has achieved numerous breakthroughs in many artificial intelligent applications. However, its complexity is quite high and usually requires expensive GPU or FPGA implementation, which is not cost-effective for many embedded systems. In this paper, we develop a novel lapped CNN (LCNN) architecture that is suitable for resource-limited embedded systems. Our architecture follows the divide-and-conquer principle. The CNN is designed such that it can be decomposed into two or more stages, each can be implemented by a hardware module with a low-resolution input and very low complexity. The original input image is divided into some subimages of the same size, with properly designed overlaps with each other. These subimages are sequentially processed by the hardware module that implements the first stage of the CNN. The outputs from different subimages are then merged and processed by the next stage low-cost hardware CNN module. The result is exactly identical to that of applying a larger-scale CNN to the entire image with higher resolution. Therefore, by reusing low-cost hardware CNN modules, a low-cost and larger-scale CNN system can be achieved. The performance of the proposed scheme is demonstrated by experimental results.
Date of Conference: 14-16 November 2017
Date Added to IEEE Xplore: 08 March 2018
ISBN Information:
Conference Location: Montreal, QC, Canada

References

References is not available for this document.