FPGA based ANN classifier for Bengali Handwritten Digit Recognition

In order to accurately and quickly recognize Bengali handwritten digits and characters, this paper suggests an FPGA-based hardware accelerator design of ANN for handwritten Bengali character recognition applications. To discover the best FPGA design for the various categorization rates, a study of network designs is first undertaken. Sigmoid as well as ReLU neurons integrated by Xilinx Vivado workspace and an axi DMA controller created in Verilog Hardware Description Language (HDL) make up the majority of its components. It is clear that sigmoid-based function implementation exceeds ReLU-based implementation for very tiny data widths (such as 4 and 8 bits). The maximum detection accuracy for Sigmoid implementation is 94.00%, and the maximum for ReLU is 87.00%. Finally, this hardware upgrade considerably increased processing speed by 4 times compared to software implementation and about 27 times faster in terms of existing CPU-based hardware implementation, in addition to maintaining recognition accuracy the same as software equivalence. Also, the power consumption is reduced by more than half of its existing solution as discussed before. As a result, the fast and accurate recognition of Bengali handwritten numbers with low power consumption was made possible by our specialized ANN models with FPGA capabilities.