Abstract
The volume of information collected from various sensors gives an unprecedented challenge for both data communications and storage space. This is particularly challenging for edge computing applications which need to consider the hardware constraints for implementation. For data communications, transmitting a whole image takes up a lot of hardware resources which is an inefficient method. The Compressed Sensing (CS) technique provides a method to acquire a sparse representation of original data and reconstruct the data from its sparse representation. In traditional CS, the compression process causes heavy computation, due to sampling and sparse representation. Additionally, the whole process needs complex design. To deal with these challenges, the Deep Neural Networks (DNNs) have been proposed into CS (e.g., CSNet). The DNN method can adaptively fit an optimal matrix for sampling and sparse representation to give more information for better performance in reconstruction process. In this paper, we propose an end-to-end Binary Convolutional Neural Network (BCNN) to present the traditional CS process for edge computing applications. Our binary computation approach can further optimize the resource footprint and accelerate the computation process. Furthermore, because the proposed BCNN is an end-to-end network, the sampling, sparse representation and reconstruction can be optimized jointly during the training. In our work, we also implement the sampling process on FPGA boards to demonstrate low usage and consumption of hardware resources for edge computing and edge AI applications.