Convolution (5x5)

Feature Map

28 x 28 x 6

Convolution (5x5)

Feature Map

14 x 14 x 6

Convolution (5x5)

Feature Map

10 x 10 x 16

Dataset

Raw texture images for training, validation, testing

Software Implementation

- LeNet model (Python + NumPy)

- Training & testing

- Store results

Hardware Implementation

- Convert model (Python to C++)

- Vivado HLS: C++ to RTL

- Vivado HLx: synthesis & bitstream

- Spartan-7 FPGA

Performance Comparison

• Accuracy (CPU vs FPGA)

• Training Time (CPU vs FPGA)

• Hardware Acceleration Validation