Step 1: Graph construction

First we need two tensors for the input data.

data_tf <- tf$placeholder(tf$float32,shape(2,2))
size_tf <- tf$placeholder(tf$int32)

Construct the transition matrix with the input value of size. We need operation tensors doing this.

seq_tf <- tf$linspace(start=1.0, stop=0.0, num=size_tf+1L)
flipseq_tf <- tf$linspace(start=0.0, stop=1.0, num=size_tf+1L)
mat1_tf <- tf$concat(list(list(seq_tf),list(flipseq_tf)),axis=0L)
mat2_tf <- tf$transpose(mat1_tf)

Two more operation tensors for matrix multiplication.

h1_tf <- tf$matmul(data_tf,mat1_tf)
out_tf <- tf$matmul(mat2_tf,h1_tf)

Define the output, actually you could include more if you want.

tf_output <- out_tf

Step 2: Data loading

The package JPEG allows us to load in jpg image file into numerical type.

library(jpeg)
chicken <- readJPEG("example.jpg")
plot(c(0,1),c(0,1))
rasterImage(chicken,0, 0, 1, 1)

The chicken image has a ground truth image with size \(600\times600\). To test our algorithm implementation, we downsampled it into size \(50\times50\).

Step 3: Graph computing

Define functions for graph computing.

blinear_interpolation <- function(original_pic,times){
  orignal_size <- dim(original_pic[,,1])[1]
  a <- orignal_size
  b <- times
  new_pic <- array(,c(a*b-b+1,a*b-b+1,3))
  sess <- tf$Session()
  for(d in 1:3){
    original_channel_data <- original_pic[,,d]
    new_channel_data <- matrix(,nrow=a*b-b+1, ncol = a*b-b+1)
    for (i in 1:(a-1)){
      for (j in 1:(a-1)){
        start_x <- i
        end_x <- i+1
        start_y <- j
        end_y <- j+1
        this.data <- original_channel_data[start_x:end_x,start_y:end_y]
        out <- sess$run(tf_output,feed_dict = dict(data_tf = this.data,size_tf = b))
        new_channel_data[(b*i-b+1):(b*i+1),(b*j-b+1):(b*j+1)] <- out
      }
      cat("row",i,"finished!","\n")
    }
    cat("channel",d,"finished!","\n")
    new_pic[,,d] = new_channel_data
  }
  sess$close()
  return(new_pic)
}

Suppose we are going to make the picture 8 times the size as before.

new_pic <- blinear_interpolation(chicken,8)

Step 4: Output image

Then we could take a look at the change.

par(mfrow=c(1,2))

plot(c(0,1),c(0,1),type=n)
rasterImage(chicken,0, 0, 1, 1)

plot(c(0,1),c(0,1),type=n)
rasterImage(new_pic,0, 0, 1, 1)

Actually we could try different values of size. Here’s what we get using a list of values.

The code below helps you save your processed image to local file.

writeJPEG(new_pic,target = "linear_interpolation_8.jpg")

Our tutorial ends here.