OpenCV - Draw only objects that keypoints refer to using python

Do have a look at this approach. It is not so tidy but you can possibly refine it to suit your needs.

I obtained keypoints on the given image using cv2.ORB_create(), for which I obtained this:

img = cv2.imread(filename,0)
orb = cv2.ORB_create()
kp = orb.detect(img,None)
kp, des = orb.compute(img, kp)

img2 = cv2.drawKeypoints(img,kp,None,color=(0,255,0), flags=0)
cv2.imshow('keypoint_on_text.jpg', img2)

You can use surf for better detection of keypoints.

Then I obtained an image having a solid color(black in this case) having the same shape as that of the image. I drew these obtained keypoints on this black image.

mask = np.zeros((img.shape[0], img.shape[1], 3), np.uint8)
mask[:] = (0, 0, 0) 

fmask = cv2.drawKeypoints(mask,kp,None,color=(0,255,0), flags=0)
cv2.imshow('fmask.jpg', fmask)

Now I converted this to grayscale image and applied threshold to binarize it. I then found contours on this image and drew them with a bigger radius.

graymask = cv2.cvtColor(fmask,cv2.COLOR_BGR2GRAY)
ret, th = cv2.threshold(graymask, 50, 255, 0)
_, contours , _= cv2.findContours(th,2,1)
rep = cv2.drawContours(fmask, contours, -1, (0,255,0), 5)
cv2.imshow('contours.jpg',rep)

I converted this to greyscale, binarized it and masked it with the original image to finally obtain this:

repmask = cv2.cvtColor(rep,cv2.COLOR_BGR2GRAY)
ret, th1 = cv2.threshold(repmask, 50, 255, 0)
res = cv2.bitwise_and(img,img,mask = th1)
cv2.imshow('Only_Text.jpg',res)

As you can see, certain portions of the desired text is visible. If you use surf detection, you will be able to obtain more portion of text.