Convolutional filter(s) for images

Introduction

This project is a simple multi-thread convolutional filter for images. It reads a batch of images from a file, applies a filter to each image, and writes the output to a different file. The filter is defined in a text file. The program uses OpenMP to parallelize the convolutional layer. Here we use a 5x5 filter with 1/273 normalization factor. The filter is applied to each color channel of the image. 5 threads are used to parallelize the convolutional layer and each thread processes a different color channel of the image. The stride(which is 1) and zero-padding of the filter can be set in the convolutional layer. The output layer dimensions are calculated automatically.

Note: The program can only run on a linux machine, as it uses OpenMP for parallelization and windows does not support the compiler.

• Images are stored in folder input/grey. All images should be of the same size (In our case 300x300) and in the same folder for the program to work.

• Python Imaging Library (PIL) was used to convert images into RGB matrices and to convert filtered matrices back to images. PIL was preferred over other C++ libraries due to ease of use.

Code structure

• Filter.hpp: Defines a 3-D convolutinal kernel class with a bias term. It contains some helper functions to allocate memory to tensors and to normalize them.
• Convolution.hpp: Defines a convolutional layer. One can set the stride and zero-padding of the filter in this. Also, dimensions of the output layer are calculated automatically.
• conv2d method takes as argument a 3-D data volume and a list of filters (one filter generates one activation map).

For example, applying a 5 x 5 x 3 filter on a 300 x 300 x 3 image (with 1 zero padding and 1 stride) will generate an 2-D output layer of 300 x 300.

• List of filters would make the output layer. Shape of output layer as well as the data block is returned by the function conv2d.

• Main.cpp runs some example filters on a batch of 3 images. It generates 3 filters, one as an edge detector for each color channel (see push_filter). Then defines a convolution layer with given params and applies the layer to each of the images. It then writes the output to a different file. *Inside helpers directory, we have make_mats.py and load_img.py that are used to generate images-matrices and vice versa.

for (int id = 0; id < num_images; ++id) {
  ...
  auto output = clayer.conv2d(input, filters);
  ...
}

Steps to run

• First and foremost install the required libraries:

sudo apt update
sudo apt install python3-pip
pip install -r requirements.txt

• Compile using make. Build file for convolutional filter demo program is 'main'.

• The run.sh file allows us to run both makefile and the rest of python commands that's going to execute the overall project. To run the project:

1. First of all make sure that run.sh has excutable permissions:

chmod +x run.sh

2. Run the run.sh file:

./run.sh

rm *.o main
g++  -std=gnu++11 -O2 filter.hpp -o filter.o
g++  -std=gnu++11 -O2 conv2d_layer.hpp -o conv2d_layer.o
g++  -std=gnu++11 -O2 main.cpp -o main

• List of images to use is in file make_mats.py. In the demo it uses a batch of 3 300 * 300 * 3 (color) images.

python3 make_mats.py img_mats/out.dat

• Run the convolutional filter (read from standard input)

./main img_mats/out.dat img_mats/filter_out.dat < filter.txt

• Make output images from matrices

python3 load_img.py img_mats/filter_out.dat out_mats

Results

You can checkout the image results in the out_mats directory. The output images are generated by applying a filter to the input images. The filter is defined in filter.txt, which is the given kernel with 1/273 normalization factor. So the Matrix is stored on the file after normalization.

The output images are stored in out_mats inside output directory.

The output images are stored in out_mats inside output directory.

Check out a couple of images with the filters applied: