ocr mser

User MSER and NMS for OCR

import cv2

import os

import numpy as np

import matplotlib.pyplot as plt

# NMS (Non Maximum Suppression, Non Maximum Suppression)

def nms(boxes, overlapThresh):

    if len(boxes) == 0:

        return []

    if boxes.dtype.kind == "i":

        boxes = boxes.astype("float")

    pick = []

    # Take four coordinate arrays

    x1 = boxes[:, 0]

    y1 = boxes[:, 1]

    x2 = boxes[:, 2]

    y2 = boxes[:, 3]

    # Computational Area Array

    area = (x2 - x1 + 1) * (y2 - y1 + 1)

    # Sort by score (if there is no confidence score, it can be sorted by coordinates from small to large, such as lower right corner coordinates)

    idxs = np.argsort(y2)

    # Start traversing and delete duplicate boxes

    while len(idxs) > 0:

        # Put the bottom right box into the pick array

        last = len(idxs) - 1

        i = idxs[last]

        pick.append(i)

        # Find the maximum and minimum coordinates in the remaining frames

        xx1 = np.maximum(x1[i], x1[idxs[:last]])

        yy1 = np.maximum(y1[i], y1[idxs[:last]])

        xx2 = np.minimum(x2[i], x2[idxs[:last]])

        yy2 = np.minimum(y2[i], y2[idxs[:last]])

        # Calculate the proportion of overlapping area to the corresponding box, i.e. IoU

        w = np.maximum(0, xx2 - xx1 + 1)

        h = np.maximum(0, yy2 - yy1 + 1)

        overlap = (w * h) / area[idxs[:last]]

        # If IoU is greater than the specified threshold, delete

        idxs = np.delete(idxs, np.concatenate(([last], np.where(overlap > overlapThresh)[0])))

    return boxes[pick].astype("int")

folder = r'C:\Work\OCR_Orientation\data\test data'

files = [os.path.join(folder, f) for f in os.listdir(folder)]

# Read pictures

imagePath = files[0]

img = cv2.imread(imagePath)

# Grayscale

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

vis = img.copy()

orig = img.copy()

# Calling MSER algorithm

mser = cv2.MSER_create()

regions, _ = mser.detectRegions(gray)  # Get the text area

hulls = [cv2.convexHull(p.reshape(-1, 1, 2)) for p in regions]  # Drawing text areas

cv2.polylines(img, hulls, 1, (0, 255, 0))

#cv2.imshow('img', img)

# Processing irregular detection boxes into rectangular boxes

keep = []

for c in hulls:

    x, y, w, h = cv2.boundingRect(c)

    keep.append([x, y, x + w, y + h])

    #cv2.rectangle(vis, (x, y), (x + w, y + h), (255, 255, 0), 1)

boxes = np.array(keep)    

merged = nms(boxes, 0.5)

for box in merged:

    x1, y1, x2, y2 = box

    cv2.rectangle(vis, (x1, y1), (x2, y2), (255, 0, 0), 1)

#cv2.imshow("hulls", vis)

plt.figure(figsize=(30,20))

plt.imshow(vis)

plt.show()