Skip to content

Latest commit

 

History

History
73 lines (64 loc) · 3.8 KB

paramtransform.md

File metadata and controls

73 lines (64 loc) · 3.8 KB
Raw

Parameterized transformations

This section includes functions for performing transformations on images requiring parameter Tensors like a warp field or a convolution kernel.

[res] image.warp([dst,]src,field,[mode,offset,clamp_mode,pad_val])

Warps image src (of sizeKxHxW) according to flow field field. The latter has size 2xHxW where the first dimension is for the (y,x) flow field. String mode can take on values lanczos, bicubic, bilinear (the default), or simple. When offset is true (the default), (x,y) is added to the flow field. The clamp_mode variable specifies how to handle the interpolation of samples off the input image. Permitted values are strings clamp (the default) or pad. When clamp_mode equals pad, the user can specify the padding value with pad_val (default = 0). Note: setting this value when clamp_mode equals clamp will result in an error. If dst is specified, it is used to store the result of the warp. Otherwise, returns a new res Tensor.

[res] image.convolve([dst,] src, kernel, [mode])

Convolves Tensor kernel over image src. Valid string values for argument mode are :

  • full : the src image is effectively zero-padded such that the res of the convolution has the same size as src;
  • valid (the default) : the res image will have math.ceil(kernel/2) less columns and rows on each side;
  • same : performs a full convolution, but crops out the portion fitting the output size of valid; Note that this function internally uses torch.conv2. If dst is provided, it is used to store the output image. Otherwise, returns a new res Tensor.

[res] image.lcn(src, [kernel])

Local contrast normalization (LCN) on a given src image using kernel kernel. If kernel is not given, then a default 9x9 Gaussian is used (see image.gaussian).

To prevent border effects, the image is first global contrast normalized (GCN) by substracting the global mean and dividing by the global standard deviation.

Then the image is locally contrast normalized using the following equation:

res = (src - lm(src)) / sqrt( lm(src) - lm(src*src) )

where lm(x) is the local mean of each pixel in the image (i.e. image.convolve(x,kernel)) and sqrt(x) is the element-wise square root of x. In other words, LCN performs local substractive and divisive normalization.

Note that this implementation is different than the LCN Layer defined on page 3 of What is the Best Multi-Stage Architecture for Object Recognition?.

[res] image.erode(src, [kernel, pad])

Performs a morphological erosion on binary (zeros and ones) image src using odd dimensioned morphological binary kernel kernel. The default is a kernel consisting of ones of size 3x3. Number pad is the value to assume outside the image boundary when performing the convolution. The default is 1.

[res] image.dilate(src, [kernel, pad])

Performs a morphological dilation on binary (zeros and ones) image src using odd dimensioned morphological binary kernel kernel. The default is a kernel consisting of ones of size 3x3. Number pad is the value to assume outside the image boundary when performing the convolution. The default is 0.