Verification

The following outlines verification of the software-hardware emulation. Hardware emulation is a tool for hardware/software co-verification and integration testing.

Baby Yoda was the first image successfully tested (using 512x512 dimensions).

Overview

 

Emulation

Originally the emulation that was built had packaging issues, so a more efficient method was created in python.

This design offers users a Python interface; simply give the PC an image and kernel size, MATLAB is called. The output is acquired from MATLAB and the actual value can be calculated.

Conv_tester

Originally, this was written in MATLAB; however, it was rewritten in Python.

This MATLAB-based implementation is called and our FPGA Convolution algorithm performs image filtering. This script loads an image and a kernel, preprocesses them, and then calls the MATLAB-based implementation of our FPGA Convolution algorithm to perform image filtering. The results are validated using SciPy, and asserts.

For thorough testing, we loaded in 1000 images from the CIFAR-10 image dataset.

Import Libraries

 

Conversion to Greyscale

The user provides the script with dimensions and metrics, it takes in an image, which is converted to greyscale.

 

Verify with MATLAB Output

Computes the ground truth correlation and checks the MATLAB output for similarity based on a tolerance.

 

Load Image, Preprocesses and Zero Pads

 

Exporting to MATLAB

The first lines generates the edge detection kernel. The image and array is saved to text files. Finally, the image and kernel arrays are loaded into MATLAB.

 

Test the FPGA Convolution Algorithm

The script will call FPGA_Runner.m from MATLAB to test the algorithm.

 

Pass/Fail Test

"True" when the convolution passes, and "False" when the convolution from MATLAB does not match the SciPy convolution.

The system will check if it matches a true convolution; if it outputs valid the test passed, while invalid indicates that the test failed. Note that if the script fails at any of the images, the loop will stop.

Project Website

Check out our website here for more information about our project.