inst/caseStudies/IMG_1236_crop.R

In duncantl/Rtesseract: Interface to the tesseract OCR system

library(Rtesseract)
img = "IMG_1236_crop.png"

# These are the true values determined manually.
trueWords = scan("IMG_1236_crop.txt", what = "")
truth = matrix(trueWords, , 4, byrow = TRUE)

# syms = apply(truth, 1, function(x) strsplit(paste(x, collapse = " "), "")[[1]])


# Now do the OCR
library(Rtesseract)
w = ocr(img, "word", opts = c("tessedit_char_whitelist" = "0123456789."))


# compare w and truth by word, breaking down characters.

miss = compareWords( names(w), trueWords)

tt = table(true = miss$true, ocr = miss$ocr)
class(tt) = "ConfusionMatrix"
plot(tt)


if(FALSE) {
    # Another approach to displaying the confusion matrix
library(ggplot2)
plot = ggplot(as.data.frame(tt))
plot + geom_tile(aes(x=ocr, y=true, fill=Freq)) + scale_x_discrete(name="Actual Class") + scale_y_discrete(name="Predicted Class")
    + scale_fill_gradient(breaks=seq(from=-.5, to=4, by=.2)) + labs(fill="Normalized\nFrequency")
}


mc = nrow(miss)/sum(nchar(truth))
# 51% misclassification 

# Note that we get a lot of the 0's after the decimal place wrong.

mainDigits =
 do.call(rbind, by(miss, miss$wordIndex,
                   function(x) {
                       n = nchar(gsub("\\..*", "", x$trueWord))
                       x[ x$position <= n, ]
                   }))

nrow(mainDigits)/sum(nchar(truth))  # Down to 31% on these digits.


library(png)
i  = readPNG(img)

# Substract off the background?
#i = i - mean(i[1:10, 1:10, 1:2])  # min(i)
#i[  i < 0 ]  = 0

# Plot the image  Do it with the plot method for a TesseractAPI, but no need to do this.
# Can do it directly with plot() and rasterImage(), but have to get the coordinate system
# right so easier to do it this way for now.
api = tesseract(image = img, "tessedit_char_whitelist" = "0123456789.")
Recognize(api)
par(mfrow = c(1, 1))
plot(api, level = "symbol", img = i, border = "grey")

 # Add the bounding boxes for the misclassified characters.
bbox = ocr(img, "symbol", boundingBox = TRUE, opts = c("tessedit_char_whitelist" = "0123456789."))
m = do.call(rbind, bbox[ miss$symbolIndex ])[, -1]  # get rid of confidence.
rect(m[,1], nrow(i) - m[,2], m[,3], nrow(i) - m[,4], border = "red")


# Note that we get a lot of the digits after the decimal places wrong.

# Let's look at the sub-image corresponding to the first misclassified digit.

pos = bbox[[ miss$symbolIndex[1] ]][-1]
k = i[ pos[2]:pos[4],  pos[1]:pos[3], ]
plot(0, type = "n", xlim = c(0, ncol(k)), ylim = c(0, nrow(k)))
rasterImage(k, 0, 0, ncol(k), nrow(k))
# Code now in plotSubImage


# Now show all the misclassified digits.

par(mfrow = c(12, 12), mar = c(0, 0, 0, 0))
invisible(lapply(bbox[ miss$symbolIndex ],
                  function(box)
                      plotSubImage(box[-1], i, axes = FALSE)))



# What about the alternatives.
alt = ocr(img, "symbol", alternatives = TRUE, opts = c("tessedit_char_whitelist" = "0123456789.", save_alt_choices = "T"))

table(sapply(alt, length)) # no actual alternatives