Text Detection and Recognition in Natural Images

img = cv2.imread('test2.jpg',0) #0 indicates read in gray scale

Use MSER to extract regions that may contain text. For more information refer to, Wikipedia: (MSER) MAXIMALLY STABLE EXTREMAL REGIONS

mser = cv2.MSER()
regions = mser.detect(img, None)    #detect and extract MSER lasso-contours

Convex-hulls are formed from the lasso contours

hulls = [cv2.convexHull(p.reshape(-1, 1, 2)) for p in regions]  #take convex-hull of each contour

The convex hulls are converted to bounded rectangles and redundant rectangles that are contained in other rectangles are removed.

x,y,w,h = cv2.boundingRect(hull)    #convert the hull to bounding rectangle

OTSU Binarization of region under rectangle is performed to make the text stand out from the background For more information refer to, Wikipedia: OTSU Binarization

thresh,letter = cv2.threshold(letter,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)

For regions getting binarized into black backgrounds and white text, inversion is performed using simple heuristic

Xnorm = [mynormalization(image,(64,64)) for image in X]