cnn-forward

Using the CNN trained by Torch7 to classify images in: 1). C++ with the support of OpenCV; 2). Matlab.

#Dependency

1. OpenCV is required to read the images.
2. Torch7 is required to convert the cnn result(eg. gstrb.net) to txt file.
3. Matlab is required run the Maltab version of the code.

#How to run the program The porgram is writted and test in Debian GNU/Linux, it is easy to run in other system

#for c++
cd forward-cnn
cmake .
make
./run

#for matlab
cd forward-cnn
run('cnnRun.m')

#The structure of the cnn in torch

model:add(nn.SpatialConvolutionMM(1, 16, 5, 5))
model:add(nn.Tanh())
model:add(nn.SpatialLPPooling(16,2, 2, 2, 2, 2))

model:add(nn.SpatialConvolutionMM(16, 32, 5, 5))
model:add(nn.Tanh())
model:add(nn.SpatialLPPooling(32,2,2, 2, 2, 2))

model:add(nn.Reshape(32*5*5))
model:add(nn.Linear(32*5*5, 256))
model:add(nn.Tanh())
model:add(nn.Linear(256, 43))

#About the files

• ./data/torch-net-to-txt.lua is used to convert the weight and bias in gtsrb.net to txt file.
• ./data/png:some png images are here for testing
• ./data/gtsrb.net is trained by Torch7
• cnn.cpp cnn.h: forward cnn class written in c++
• run.cpp: test program
• *.m: matlab file