Recognizing a topological graph in a noisy image - image-processing

Question

I am not at all experienced with machine learning or image processing, so I'm hoping someone can give some pointers of first thoughts on this problem:
The image below is an example of a photograph of tomato plant leaf. We have thousands of these. We need to trace the veins and output a graph. We have already had undergraduates trace the veins by hand for a few hundreds, so I presume that this can be a training set for a machine learning approach.
So my question: what types of filters/classifiers immediately come to mind? Is there anything you recommend I read or take a look at?
Our first thought was, look at directional derivatives. Each pixel can be classified as being in an edge or not in an edge at a given angle, and if a pixel is in an for a lot of different angles, then it's probably a blotch and not a vein. Then the parameters of gradient threshold and angle variation allowed can be adjusted by the learning but probably this is not the best way...
Thank you for any help!

Answer 1

Two methods immediately come to mind
a sliding window neural network classifier
identifying a threshold that sets apart dark/light pixels in the image (this could be done using machine learning or perhaps a simple computation) and then doing a flood fill to identify regions in the image.
The second method should be simpler and quicker, so I'd perhaps prototype it first to see if it gives good enough answers.
In any case, my intuition is that it's going to be easier to solve the dual problem - not trying to find edges and nodes of the graph, but finding its faces. From that, you get the graph itself easily.

Answer 2

I did this very simple program to filter the vein regions using opencv. I've added comments to explain the operations. Resulting images for the intermediate steps are saved. Hope it helps.
#include "stdafx.h"
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <iostream>
using namespace cv;
using namespace std;
#define INPUT_FILE "wMTjH3L.png"
#define OUTPUT_FOLDER_PATH string("")
#define CONTOUR_AREA_THRESHOLD 30.0
int _tmain(int argc, _TCHAR* argv[])
{
// read image as grayscale
Mat im = imread(INPUT_FILE, CV_LOAD_IMAGE_GRAYSCALE);
imwrite(OUTPUT_FOLDER_PATH + string("gray.jpg"), im);
// smooth the image with a gaussian filter
Mat blurred;
GaussianBlur(im, blurred, Size(3, 3), 1.5);
imwrite(OUTPUT_FOLDER_PATH + string("blurred.jpg"), blurred);
// flatten lighter regions while retaining the darker vein regions using morphological opening
Mat morph;
Mat morphKernel = getStructuringElement(MORPH_ELLIPSE, Size(5, 5));
morphologyEx(blurred, morph, MORPH_OPEN, morphKernel);
imwrite(OUTPUT_FOLDER_PATH + string("morph.jpg"), morph);
// apply adaptive thresholding
Mat adaptTh;
adaptiveThreshold(morph, adaptTh, 255.0, ADAPTIVE_THRESH_GAUSSIAN_C, CV_THRESH_BINARY_INV, 7, 2.0);
imwrite(OUTPUT_FOLDER_PATH + string("adaptth.jpg"), adaptTh);
// morphological closing to merge disjoint regions
Mat morphBin;
Mat morphKernelBin = getStructuringElement(MORPH_ELLIPSE, Size(3, 3));
morphologyEx(adaptTh, morphBin, MORPH_CLOSE, morphKernelBin);
imwrite(OUTPUT_FOLDER_PATH + string("adptmorph.jpg"), morphBin);
// find contours
vector<vector<Point>> contours;
vector<Vec4i> hierarchy;
findContours(morphBin, contours, hierarchy, CV_RETR_CCOMP, CV_CHAIN_APPROX_SIMPLE, Point(0, 0));
// filter contours by region areas and draw
RNG rng(12345);
Mat drawing = Mat::zeros(morphBin.size(), CV_8UC3);
for(int idx = 0; idx >= 0; idx = hierarchy[idx][0])
{
if (contourArea(contours[idx]) > CONTOUR_AREA_THRESHOLD)
{
Scalar color( rand()&255, rand()&255, rand()&255 );
drawContours( drawing, contours, idx, color, CV_FILLED, 8, hierarchy );
}
}
imwrite(OUTPUT_FOLDER_PATH + string("cont.jpg"), drawing);
return 0;
}
The output looks like this for the provided sample image: