Vertical margins in the detected boxes from easyocr.detect don't have a consistent behavior

[fxmarty]

When setting add_margin=0 (not sure about other cases), some boxes detected with easyocr.detect still have a border while some don't, as illustrated below.

How to reproduce

Download

Execute

import torch
from easyocr import Reader
import numpy as np
import cv2
import matplotlib.pyplot as plt

reader = Reader(['ja'])

filename = '/path/to/file.png' # change accordingly
img = cv2.imread(filename, 1)

# Here we add some white border to the image
h,w=img.shape[0:2]
border = 50
base_size=h+border,w+border,3
base=np.zeros(base_size,dtype=np.uint8) + 255
base[border//2:border//2 + h, border//2:border//2 + w]=img

# Detect boxes, without margins
result = reader.detect(base, add_margin=0)

Then execute

k = result[0][0]
cv2.rectangle(base, (k[1], k[3]),
            (k[0], k[2]),
            (0,0,255),
            1)

plt.imshow(base)
plt.show()

and

k = result[0][1]
cv2.rectangle(base, (k[1], k[3]),
            (k[0], k[2]),
            (0,0,255),
            1)

plt.imshow(base)
plt.show()

This returns

and

As you can see, the first one has borders, while the second do not.

I guess it is not too big of an issue, and it can be solved manually, but still, it is not expected behavior here.

Thank you!

[rkcosmos]

add_margin is an extra margin in addition to the output from text detection. Text detection is deep learning-based algorithm and may introduce its own small margin.

[fxmarty]

I see, seems very reasonable. Thank you for your help.