Detect image orientation using OCR

detect_orientation.py

import pathlib
import pytesseract
import cv2
import sys
import exif
import multiprocessing
import enchant

def ocr(img):
	return pytesseract.image_to_string(img, lang='fra', config="--psm 6").replace('\n', ' ') #psm: disable orientation detection
		
def getscore(text):
	dictionary = enchant.Dict("fr_FR")
	words = [word for word in text.split() if len(word) >= 4]
	return sum(1 for word in words if dictionary.check(word))

def rotate_exif(filename, rot):
	with open(filename, 'rb') as sf :
		img = exif.Image(sf)
		if not(img.has_exif): img.orientation = 1		
		neworientation = [1,6,3,8][({1:0, 6:1, 3:2, 8:3}[img.orientation] + rot)%4]
		if img.orientation != neworientation:
			img.orientation = neworientation
			output = img.get_file()
			with open(filename, 'wb') as df:
				df.write(output)		

def detect(path):
	filename = str(path)
	img = cv2.imread(filename)
	img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) #grayscale
	img = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1] #high contrast

	bestscore = bestangle = -1
	for angle in range(4):
		if angle == 0:
			score = getscore(ocr(img))
		#elif angle == 1:
		#	score = getscore(ocr(cv2.rotate(img, cv2.ROTATE_90_CLOCKWISE)))
		elif angle == 2:
			score = getscore(ocr(cv2.rotate(img, cv2.ROTATE_180)))			
		#elif angle == 3:
		#	score = getscore(ocr(cv2.rotate(img, cv2.ROTATE_90_COUNTERCLOCKWISE)))			
		if score>bestscore and score>5: #at least five correct words
			bestscore = score
			bestangle = angle
	
	if bestangle>-1:	
		rotate_exif(filename, bestangle)	
	print(f'{filename} {bestangle}')	
	
if __name__ == '__main__':
	if len(sys.argv) == 1:
		raise SystemExit(f"Usage: {sys.argv[0]} <directory_with_jpg_files>")	
	pool = multiprocessing.Pool(multiprocessing.cpu_count()) 
	pool.map(detect, pathlib.Path(sys.argv[1]).glob('*.jpg'))