drml: Dimension Reduction in Image Classification
In matloff/imageClassR: Quick and Easy Functions for Image Classification

Description Usage Arguments Details Value Author(s) Examples

Functions for the "dimension reduction + machine learning" approach to image classification.

   drmlTDAsweep(data, yName, qeFtnName, opts = NULL, RGB = FALSE, pixAug = 0, 
       tdasAug = 0, holdout = floor(min(1000, 0.1 * nrow(imgs))), 
       nr = 0, nc = 0, thresh = c(50, 100, 150), intervalWidth = 2) 
   drmlPCA(data, yName, qeFtnName, opts = NULL, dataAug = NULL, 
       holdout = floor(min(1000, 0.1 * nrow(data))), pcaProp) 
   drmlUMAP(data, yName, qeFtnName, opts = NULL, dataAug = NULL, 
       holdout = floor(min(1000, 0.1 * nrow(data))), nComps = 25) 
   drmlDCT(data, yName, qeFtnName, opts = NULL, dataAug = NULL, 
       holdout = floor(min(1000, 0.1 * nrow(data))), nFreqs) 
   drmlRLRN(data, yName, qeFtnName, opts = NULL, RGB = FALSE, pixAug = 0, 
       holdout = floor(min(1000, 0.1 * nrow(imgs))), nr = 0, nc = 0, 
       thresh = c(50, 100, 150))

`data`	Data frame, one image per row, pixels within an image being stored in row-major. For color images, 3 sets of columns, for the 3 primary colors.
`yName`	Name of the column within `data` that stores the image labels, an R factor.
`qeFtnName`	Name of the function from the qeML to be used in the "ML" portion of "DR+ML."
`opts`	Options for `qeFtnName`.
`RGB`	TRUE for color, FALSE for grayscale.
`pixAug`	Number of images to add via data augmentation, between the DR and ML stages.
`holdout`	Size of holdout set.
`nr`	Number of pixel rows within an image.
`nc`	Number of pixel columns within an image.
`thresh`	Vector specifying the threshold values. If this is a negative scalar -m, then then m threshold values will be generated, partitioning [0,255] into m+1 equal parts.

Dimension reduction is done on the pixel data, after which the ML method is applied. If data augmentation is requested, this is performed on the dimension-reduced data, before applying ML. This should yield a speedup over doing data augmentation before dimension reduction. Half the augmented images are horizontal flips, half vertical.

If holdout is nonzero, the data are first randomly partitioned into training and validations sets, and overall misclassification rate is reported in the testAcc component of the return value.

New cases can be classified with the generic predict function (only for TDAsweep as of now).

Norm Matloff, Yu-Shih Chen, Melissa Goh

## Not run: 

data(hm)  # histology MNIST, built-in dataset
tdasOut <- drmlTDAsweep(hm,'label','qeRF',nr=28,nc=28,thresh=-7)
tdasOut$testAcc
# 0.216, 22


## End(Not run)