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