Feature extraction: weak edges enhancement

Hi all,

I have to train a svm classificator in order to distinguish among 8 types of classes, which represent 8 types of ionizing radiations. Here I show you some samples for each class:

1) unexpected https://s4.postimg.org/5hiyvrpot/Job_5694.jpg

2) scattered https://s3.postimg.org/ib2ojhwf7/Job_3508.jpg

3) direct https://s4.postimg.org/4xuwleh59/Job_4927.jpg | https://s4.postimg.org/5bc28dif1/Job_7730.jpg

4) background https://s4.postimg.org/gp3ngoz99/Job_6160.jpg

5) label https://s3.postimg.org/kahl3vz03/Job_3944.jpg

6) saturated https://s3.postimg.org/5b5baarab/Job_5185.jpg

7) sum of scattered https://s3.postimg.org/qhxkhsk4z/Job_4179.jpg

8) sum of direct https://s4.postimg.org/f3vf1t6cd/Job_4144.jpg

To extract the texture information, I have subdivided each image in 9 windows, and for each of these I have calculated the glcm. From these I calculated different parameters such as Contrast, Homogeneity, Correlation, Energy, etc. But this was not sufficient in order to discriminate between sum direct/direct and sum scattered/scattered, so i thought to extract the edge and then calculate the Histogram of Gradients, because as you can see from the pictures, this 2 classes differ only by the sharpness of the edges. My problem is that in some image the radiation footprint is very weak and the common algorithms (Canny, Sobel, etc) aren't able to extract any kind of edges. I tried to adjust the contrast of the images, but these images become too much noisy and again I can't extract any edge.

Is there any trick in order to enhance and extract these weak edges or i have to extract another feature to discriminate these two classes? If yes, which features I could extract?

I thank everyone for any hint.

Answer 1

[Swathik Kurella Janardhan]

Refer to extract features functionality from Computer Vision System Toolbox that can help you as starting point to extract edges from images.

Also refer to feature detection and extraction to get start with.