ai-agents-ocr-accuracy-text-recognition-accuracy-drops-when-detected-text-564fd979

THEN Apply OpenCV contour detection on the detected text region to find the minimal bounding rectangle of the text itself, then crop to that rectangle with a small fixed padding (e.g., 1-2 pixels) to match training conditions. This forces the input to resemble the tight-crop training samples, improving recognition accuracy.

import cv2
# img: input cropped region (detected text line)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
_, thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if contours:
    x, y, w, h = cv2.boundingRect(np.vstack(contours))
    padding = 2
    x = max(x - padding, 0)
    y = max(y - padding, 0)
    w = min(w + 2*padding, img.shape[1]-x)
    h = min(h + 2*padding, img.shape[0]-y)
    cropped = img[y:y+h, x:x+w]