flask-ocr-api / app /ocr_engine.py
vamsi1103's picture
Upload 26 files
0e157b6 verified
Raw History Blame Contribute Delete
4.2 kB
import os
import numpy as np
from PIL import Image
from rapidocr_onnxruntime import RapidOCR
from .preprocess import preprocess_image
# 1. Initialize Default RapidOCR (for general text, ASCII, numbers, and rates)
default_engine = RapidOCR()
# 2. Initialize Telugu RapidOCR (if Telugu model is present)
telugu_model_path = os.getenv("REC_MODEL_PATH") or os.path.join("models", "telugu", "rec.onnx")
telugu_keys_path = os.getenv("KEYS_PATH") or os.path.join("models", "telugu", "dict.txt")
telugu_engine = None
if os.path.exists(telugu_model_path) and os.path.exists(telugu_keys_path):
try:
telugu_engine = RapidOCR(rec_model_path=telugu_model_path, rec_keys_path=telugu_keys_path)
print("Telugu RapidOCR model initialized successfully.")
except Exception as e:
print(f"Warning: Failed to initialize Telugu RapidOCR model: {e}")
def run_ocr(image):
"""
Runs text detection (DBNet) and recognition (Default + Telugu SVTR/PP-OCR if available).
Returns list of detected bounding boxes with both default and Telugu text recognitions.
"""
img_np = np.array(image)
processed = preprocess_image(img_np)
res_default, _ = default_engine(processed)
res_telugu = None
if telugu_engine is not None:
try:
res_telugu, _ = telugu_engine(processed)
except Exception as e:
print(f"Warning: Telugu OCR execution failed: {e}")
extracted = []
if not res_default and not res_telugu:
return extracted
# Helper to calculate box metrics
def get_center(box):
xs = [p[0] for p in box]
ys = [p[1] for p in box]
return sum(xs)/len(xs), sum(ys)/len(ys)
# If Telugu engine is not available, return default
if not res_telugu:
for item in res_default or []:
if not item or len(item) < 3:
continue
box = item[0]
extracted.append({
"text": str(item[1]).strip(),
"default_text": str(item[1]).strip(),
"telugu_text": "",
"confidence": float(item[2]),
"box": box.tolist() if hasattr(box, "tolist") else box
})
return extracted
# When both are available, pair fragments by bounding box centers
tel_items = []
for item in res_telugu:
if not item or len(item) < 3:
continue
cx, cy = get_center(item[0])
tel_items.append({
"text": str(item[1]).strip(),
"conf": float(item[2]),
"box": item[0].tolist() if hasattr(item[0], "tolist") else item[0],
"cx": cx,
"cy": cy
})
def_items = []
for item in (res_default or []):
if not item or len(item) < 3:
continue
cx, cy = get_center(item[0])
def_items.append({
"text": str(item[1]).strip(),
"conf": float(item[2]),
"box": item[0].tolist() if hasattr(item[0], "tolist") else item[0],
"cx": cx,
"cy": cy
})
# Pair each Telugu box with nearest Default box
for t in tel_items:
best_d = None
min_dist = 999999
for d in def_items:
dist = (t["cx"] - d["cx"])**2 + (t["cy"] - d["cy"])**2
if dist < min_dist:
min_dist = dist
best_d = d
def_txt = best_d["text"] if (best_d and min_dist < 150) else ""
tel_txt = t["text"]
extracted.append({
"text": tel_txt if tel_txt else def_txt,
"telugu_text": tel_txt,
"default_text": def_txt,
"confidence": round(t["conf"], 2),
"box": t["box"]
})
# Also include any default items that had no nearby Telugu box
for d in def_items:
is_paired = any(((d["cx"] - t["cx"])**2 + (d["cy"] - t["cy"])**2) < 150 for t in tel_items)
if not is_paired:
extracted.append({
"text": d["text"],
"telugu_text": "",
"default_text": d["text"],
"confidence": round(d["conf"], 2),
"box": d["box"]
})
return extracted