Abstract
Convolutional Neural Network (CNN) is a good candidate for computer vision applications. CNN is well known for its great classification accuracy at image recognition tasks. The cost of CNN is its large power consumption as it needs a lot of multiplication and addition operations. Approximation can reduce the power consumption. CNNs can be implemented by CPUs, GPUs or FPGAs. In this paper, the proposed CNN is implemented on Xilinx XC7Z020 FPGA and is trained to recognize MNIST dataset This CNN is approximated through quantization which reduces the accuracy only by 0.53% while using 7-bits for the implementation. A reduction of 2.7X is achieved in energy consumption compared to the conventional design which uses 16-bits. Dynamic Partial Reconfiguration (DPR) reconfigures the FPGA during the run time with appropriate power consumption design if the battery level decreases. This enables recognition applications to run with low power budget but with sacrificing minor accuracy instead of termination.