from fastapi import FastAPI import gradio as gr from PIL import Image from ultralytics import YOLO import easyocr import cv2 from warnings import filterwarnings import numpy as np import re import json import glob import os import shutil import paddleocr from paddleocr import PaddleOCR from rich import print from langchain.chains import ConversationChain from langchain.memory import ConversationBufferMemory from langchain_core.messages import HumanMessage from langchain_groq import ChatGroq # Import ChatGroq from langchain_openai import ChatOpenAI import base64 from fastapi.middleware.cors import CORSMiddleware from warnings import filterwarnings filterwarnings('ignore') # Initialize the memory memory = ConversationBufferMemory() model_key = os.getenv("API_KEY") model_id = "gpt-4o-mini" # Initialize the ChatGroq model # chat_model = ChatGroq(model="llama-3.2-11b-vision-preview", temperature=0, # api_key="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", max_tokens=1000) chat_model = ChatOpenAI(model=model_id, max_tokens=1000, openai_api_key=model_key) filterwarnings('ignore') # Create a reader to do OCR. reader = easyocr.Reader(['en']) # Load a pretrained YOLOv8n model model = YOLO('yolo_models/yolov8l_imgsz800_map85_best.pt', verbose=False) print('Model Loaded...') ocr_model = PaddleOCR(use_angle_cls=True, lang='en', show_log=False, rec_model_dir='rec_r45_abinet', det_model_dir='det_r50_vd_east_v2.0') def perform_pad_ocr(image, conf=0.20, iou=0.91, imgsz=800): if os.path.exists('cropped'): shutil.rmtree('cropped/') # make detection results = model.predict(image, conf=conf, iou=iou, imgsz=imgsz) ocr_result_path = "OCR_RESULTS" image_fname = "uploaded_image" chat1 = [] timestamp_formats = ['%I.%M %p', '%I %p', '%I.%M%p', '%I%M %p', '%I %p', '%I %M %p', '%I:%M %p', '%I:%M:%S %p', '%H:%M', '%H.%M', '%H %M', '%H:%M:%S', '%H.%M.%S', '%H %M %S', '%H:%M:%S.%f', '%H.%M.%S.%f', '%H %M %S %f'] for j in results: os.makedirs(ocr_result_path, exist_ok=True) boxes = j.boxes.cpu().numpy() j.save(filename=f'{ocr_result_path}/{image_fname}_result.jpg') # save to disk j.save_crop('cropped') message_boxes = [box for box in boxes if j.names[int(box.cls[0])] in ['sender_message', 'receiver_message']] replied_boxes = [box for box in boxes if j.names[int(box.cls[0])] in ['replied_to_message']] message_boxes.sort(key=lambda box: box.xyxy[0][1]) replied_boxes.sort(key=lambda box: box.xyxy[0][1]) emoji_boxes = [box for box in boxes if j.names[int(box.cls[0])] == 'emoji'] timestamp_boxes = [box for box in boxes if j.names[int(box.cls[0])] == 'timestamp'] for msg_box, t_box in zip(message_boxes, timestamp_boxes): message_type = j.names[int(msg_box.cls[0])] msg_r = msg_box.xyxy[0].astype(int) roi = (cv2.imread(image, 0))[msg_r[1]:msg_r[3], msg_r[0]:msg_r[2]] ocr_res = ocr_model.ocr(roi, cls=True)[0] ocr_text3 = " " if ocr_res is not None: for i in ocr_res: txt = i[-1][0] txt = txt.replace('\n', ' ') txt = txt.strip() ocr_text3 += f"{txt} " t_i = cv2.imread(image) fin_oc = [] ocr_text1 = " " for i,emoji_box in enumerate(emoji_boxes): emoji_r = emoji_box.xyxy[0].astype(int) kernel = np.ones((8,8),np.float32)/15 t_i[emoji_r[1]:emoji_r[3], emoji_r[0]:emoji_r[2]] = cv2.filter2D(t_i[emoji_r[1]:emoji_r[3], emoji_r[0]:emoji_r[2]], -1, kernel) center_x = int((emoji_r[0] + emoji_r[2]) / 2) center_y = int((emoji_r[1] + emoji_r[3]) / 2) text_size = cv2.getTextSize("%", cv2.FONT_HERSHEY_SIMPLEX, 1, 2)[0] text_x = center_x - text_size[0] // 2 text_y = center_y + text_size[1] // 2 cv2.putText(t_i, "% ", (text_x, text_y), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 0, 0) , 2) cv2.imwrite('t_i.jpg', t_i) roi1 = (t_i[msg_r[1]:msg_r[3], msg_r[0]:msg_r[2]]) ocr_res1 = ocr_model.ocr(roi1, cls=True)[0] new_oc = [] if ocr_res1 is not None: for i in ocr_res1: txt1 = i[-1][0] txt1 = txt1.replace('\n', ' ') txt1 = txt1.strip() ocr_text1 += f"{txt1} " new_oc.append(txt1) fin_oc.append(new_oc) ocr_text2 = " ".join(fin_oc[-1]) ocr_text3 = ocr_text2 ocr_text3 = re.sub(r'(EA|EF|EL|EM|EI|EN|EV|El|FM|FN|FI|FA|FE|FV|M E|E M|EE|MM|NM|MN|MI|NN|IM|M1|1M|MW|AN|SN|Q|%)+', '$$ ', ocr_text3) rep_tt = "" for replied_box in replied_boxes: rep_r = replied_box.xyxy[0].astype(int) if (msg_r[0] <= rep_r[0] and rep_r[2] <= msg_r[2]) and (msg_r[1] <= rep_r[1] and msg_r[3] <= rep_r[3]): rep_roi = cv2.imread(image, 0)[rep_r[1]:rep_r[3], rep_r[0]:rep_r[2]] rep_ocr_res = ocr_model.ocr(rep_roi, cls=True)[0] if rep_ocr_res is not None: for i in rep_ocr_res: txt = i[-1][0] rep_tt += f"{txt} " try: for time_box in timestamp_boxes: time_r = time_box.xyxy[0].astype(int) if (msg_r[0] <= time_r[0] and time_r[2] <= msg_r[2]) and (msg_r[1] <= time_r[1] and time_r[3] <= msg_r[3]): ts_r = time_r ts_roi = cv2.imread(image)[ts_r[1]:ts_r[3], ts_r[0]:ts_r[2]] ts_ocr_res = ocr_model.ocr(ts_roi, cls=True)[0] ts_text = "" if ts_ocr_res is not None: for i in ts_ocr_res: txt = i[-1][0] ts_text += txt timestamp = txt if timestamp is None: match = re.search(r'\b\d{1,2}:\d{2}\s?(?:AM|PM)\b', ocr_text3) timestamp = match.group() if match else None ocr_text3 = ocr_text3.replace(timestamp, ' ') if timestamp else ocr_text3 ocr_text3 = re.sub(r'\b\d{1,2}:\d{2}\s?(?:AM|PM)\b', '', ocr_text3) ocr_text3 = ocr_text1.replace(rep_tt, '') if rep_tt else ocr_text3 dictt1 = {"type": message_type, "text": ocr_text3.strip(), 'timestamp': timestamp} chat1.append(dictt1) except: pass json.dump(chat1, open(f"{ocr_result_path}/{image_fname}_result1.json", 'w')) return chat1 def get_deep_message_extractor_ans(image_path, q): # Construct the prompt with memory fin_q = json.dumps(q, indent=2) with open(image_path, "rb") as image_file: image_base64 = base64.b64encode(image_file.read()).decode('utf-8') prompt = ( f"From the original image" f"You are a deep message and emoji extractor assistant.\n" f"Your task is to process and analyze messages extracted text \n{fin_q} \n: from that image." f"Your main focus is just removing duplicate dictionary with the following condition: \n" f"If more than two dicts shares same text: values, remove the second dict and keep the first one\n" f"Ensure to remove duplicates in dictionary, before outputing final JSON\n" f"Do not modify the extracted texts; more importantly do not remove the emoji placeholders $$\n" f"Ensure your final output is a JSON just like you receive it.\n" f"If there are no duplicates, return original JSON with no modification\n" f"If there are duplicates, JUST REMOVE the later dictionary and return final JSON, *** DO NOT MODIFY THE TEXTS, EXPECIALLY THE emoji placeholders $$ ***\n" f"If you find emoji in the image, but not in the original extracted text, also have it replace with the placeholder $$\n" f"Finally; make sure all the $$ emoji placeholders are replaced with unicode emoji in the final text or JSON\n" f"And if no placeholder found, check from the image and make sure unicode emoji are correctly placed in the texts if available" ) # Add the conversation history from memory history = memory.load_memory_variables({}) if history['history']: prompt = f"{history['history']}\n\n{prompt}" # Invoke the ChatGroq model # response = chat_model.invoke([HumanMessage(content=prompt)]).content.strip() response = chat_model.invoke([ HumanMessage( content=[ {"type": "text", "text": prompt}, {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}}, ] ) ]) # Save the current interaction to memory memory.save_context({"input": q}, {"output": response.content.strip()}) return response.content.strip() def extract_json_from_response(response_text): """ Extracts the JSON portion from the LLM's response. """ # Use regex to find the JSON object or array json_match = re.search(r'\[\s*\{.*\}\s*\]|\{\s*.*\s*\}', response_text, re.DOTALL) if json_match: try: # Parse the extracted JSON json_data = json.loads(json_match.group()) return json_data except json.JSONDecodeError: print("Failed to parse extracted JSON.") return {"error": "Invalid JSON format in the response."} else: return {"error": "No JSON found in the response."} def gradio_interface(image, conf, iou, imgsz): res = perform_pad_ocr(image, conf, iou, imgsz) fin_res = get_deep_message_extractor_ans(image, res) ex_res = extract_json_from_response(fin_res) # print(res) return ex_res # Create the Gradio interface iface = gr.Interface( fn=gradio_interface, inputs=[ gr.Image(type="filepath", label="Upload Image"), gr.Slider(minimum=0.25, maximum=1.0, step=0.01, value=0.25, label="Confidence Threshold"), gr.Slider(minimum=0.45, maximum=1.0, step=0.01, value=0.91, label="IOU Threshold"), gr.Slider(minimum=800, maximum=1000, step=32, value=800, label="Image Size") ], outputs=gr.JSON(label="OCR Results"), title="deep message extractor", description="Upload an image and adjust the parameters to extract text." ) # # Create the FastAPI app # app = FastAPI() # # Add CORS middleware to allow all origins # app.add_middleware( # CORSMiddleware, # allow_origins=["*"], # allow_credentials=True, # allow_methods=["*"], # allow_headers=["*"], # ) # # Mount the Gradio app # app = gr.mount_gradio_app(app, iface, path="/") # # Run the FastAPI app # if __name__ == "__main__": # import uvicorn # uvicorn.run(app, host="0.0.0.0", port=80) iface.launch(favicon_path='smile.png')