X-ray auto_crop v2

import numpy as np
import cv2
import math

def subimage(image, center, theta, width, height):
    if 45 < theta <= 90:
        theta = theta - 90
        width, height = height, width

    theta *= math.pi / 180 # convert to rad
    v_x = (math.cos(theta), math.sin(theta))
    v_y = (-math.sin(theta), math.cos(theta))
    s_x = center[0] - v_x[0] * (width / 2) - v_y[0] * (height / 2)
    s_y = center[1] - v_x[1] * (width / 2) - v_y[1] * (height / 2)
    mapping = np.array([[v_x[0],v_y[0], s_x], [v_x[1],v_y[1], s_y]])
    return cv2.warpAffine(image, mapping, (width, height), flags=cv2.WARP_INVERSE_MAP, borderMode=cv2.BORDER_REPLICATE)

def get_threshold_clr(image_source):
    h, w = image_source.shape[:2]
    inset = math.ceil(.1*(h + w)/2)
    mismatch = 0
    corners = ((inset, inset), (w-inset, h-inset), (inset, h-inset), (w-inset, inset))
    clr = None
    for crnr1 in corners:
        for crnr2 in corners:
            if (np.linalg.norm(image_src[crnr1] - image_src[crnr2]) > 3):
                mismatch += 1
                break
            else:
                clr = image_src[crnr1]

    if (mismatch > 3 + 3 or clr == None):
        return None
    else:
        # Get one threshold value
        clr = np.mean(clr)
        if (clr > 250):
            return 240
        elif(clr < 10):
            return 20
        else:
            None

def auto_crop(image_source):
    # First slightly crop edge - some images had a rogue 2 pixel black edge on one side
    init_crop = 5
    h, w = image_source.shape[:2]
    image_source = image_source[init_crop:init_crop+(h-init_crop*2), init_crop:init_crop+(w-init_crop*2)]

    threshold = get_threshold_clr(image_source)
    if (threshold == None):
        return image_source

    edge_clr = 255
    if (threshold < 128):
        edge_clr = 0

    # Add back a white border
    image_source = cv2.copyMakeBorder(image_source, 5,5,5,5, cv2.BORDER_CONSTANT, value=(edge_clr, edge_clr, edge_clr))

    image_gray = cv2.cvtColor(image_source, cv2.COLOR_BGR2GRAY)


    _, image_thresh = cv2.threshold(image_gray, threshold, 255, cv2.THRESH_BINARY)

    image_thresh2 = image_thresh.copy()
    image_thresh2 = cv2.Canny(image_thresh2, 100, 100, apertureSize=3)
    points = cv2.findNonZero(image_thresh2)

    centre, dimensions, theta = cv2.minAreaRect(points)
    rect = cv2.minAreaRect(points)

    width = int(dimensions[0])
    height = int(dimensions[1])

    box = cv2.boxPoints(rect)
    box = np.int0(box)

    temp = image_source.copy()
    cv2.drawContours(temp, [box], 0, (255,0,0), 2)

    M = cv2.moments(box)
    cx = int(M['m10']/M['m00'])
    cy = int(M['m01']/M['m00'])

    image_patch = subimage(image_source, (cx, cy), theta+90, height, width)

    # add back a small border
    image_patch = cv2.copyMakeBorder(image_patch, 1,1,1,1, cv2.BORDER_CONSTANT, value=(edge_clr,edge_clr,edge_clr))

    # Convert image to binary, edge is black. Do edge detection and convert edges to a list of points.
    # Then calculate a minimum set of points that can enclose the points.
    _, image_thresh = cv2.threshold(image_patch, threshold, 255, 1)
    image_thresh = cv2.Canny(image_thresh, 100, 100, 3)
    points = cv2.findNonZero(image_thresh)
    hull = cv2.convexHull(points)

    # Find min epsilon resulting in exactly 4 points, typically between 7 and 21
    # This is the smallest set of 4 points to enclose the image.
    for epsilon in range(3, 50):
        hull_simple = cv2.approxPolyDP(hull, epsilon, 1)

        if len(hull_simple) == 4:
            break

    hull = hull_simple

    # Find closest fitting image size and warp/crop to fit
    # (ie reduce scaling to a minimum)

    x,y,w,h = cv2.boundingRect(hull)
    target_corners = np.array([[0,0],[w,0],[w,h],[0,h]], np.float32)

    # Sort hull into tl,tr,br,bl order.
    # n.b. hull is already sorted in clockwise order, we just need to know where top left is.

    source_corners = hull.reshape(-1,2).astype('float32')
    min_dist = 100000
    index = 0

    for n in xrange(len(source_corners)):
        x,y = source_corners[n]
        dist = math.hypot(x,y)

        if dist < min_dist:
            index = n
            min_dist = dist

    # Rotate the array so tl is first
    source_corners = np.roll(source_corners , -(2*index))

    try:
        transform = cv2.getPerspectiveTransform(source_corners, target_corners)
        return cv2.warpPerspective(image_patch, transform, (w,h))

    except:
        print "Warp failure"
        return image_patch


if __name__ == "__main__":
    import os
    import re
    import sys

    cv2.namedWindow("Original")
    cv2.namedWindow("Cropped")

    if (len(sys.argv) < 2):
        sys.exit()
    else:
        source_dir = sys.argv[1]
        if (not re.search("^/", source_dir)):
            source_dir = "{0}/{1}".format(os.path.dirname(os.path.realpath(__file__)),
                                          source_dir)

    if not os.path.isdir(source_dir):
        print "The source directory '{0}' does not exist".format(source_dir)
        sys.exit()

    for (directory, _, files) in os.walk(source_dir):
        for f in files:
            if re.search("\\.png$", f) and\
                not re.search("^cropped_", f):
                path = os.path.join(directory, f)
                image_src = cv2.imread(path)
                image_cropped = auto_crop(image_src)
                cv2.imwrite(os.path.join(directory, "cropped_{0}".format(f)),
                            image_cropped)
                print "Processed file {0}".format(f)
                cv2.imshow("Original", image_src)
                cv2.imshow("Cropped", image_cropped)
                cv2.waitKey(0)