Optimizing the Convolution Operation to Accelerate Deep Neural Networks on FPGA