ronn12/truth

1import os2import cv23import numpy as np4import tensorflow as tf5import librosa6import pandas as pd7from joblib import load8from moviepy.editor import VideoFileClip9from fastapi import FastAPI, UploadFile, File, HTTPException10from starlette.responses import JSONResponse11 12app = FastAPI()13 14UPLOAD_FOLDER = "uploads"15os.makedirs(UPLOAD_FOLDER, exist_ok=True)16 17saved_model = tf.keras.models.load_model("./lstm_without_masking(94).h5")18audio_model = tf.keras.models.load_model("./audio_model.h5")19scaler = load("./scaler.joblib")20 21latest_result = None22 23# Function to validate input video (one person, clear face)24def input_validation(input_video):25    import mediapipe as mp26    mp_face_detection = mp.solutions.face_detection27    face_detection = mp_face_detection.FaceDetection()28    cap = cv2.VideoCapture(input_video)29 30    valid = True31    while cap.isOpened():32        ret, frame = cap.read()33        if not ret:34            break35        rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)36        results = face_detection.process(rgb_frame)37        if not results.detections or len(results.detections) > 1:38            valid = False39            break40    cap.release()41    return valid42 43# Function to check if the video has audio44def check_audio_in_video(video_file):45    video = VideoFileClip(video_file)46    return 1 if video.audio else 047 48# Extract audio from video49def extract_audio_from_video(video_file, audio_file):50    video = VideoFileClip(video_file)51    audio = video.audio52    audio.write_audiofile(audio_file, codec='pcm_s16le')53 54# Split audio into segments55def split_and_pad_audio(audio_file, segment_duration=10, sample_rate=16000):56    segment_samples = segment_duration * sample_rate57    audio, _ = librosa.load(audio_file, sr=sample_rate)58    num_segments = len(audio) // segment_samples59    remainder = len(audio) % segment_samples60    segments = [audio[i * segment_samples : (i + 1) * segment_samples] for i in range(num_segments)]61    if remainder > 0:62        last_segment = np.pad(audio[num_segments * segment_samples:], (0, segment_samples - remainder), mode='constant')63        segments.append(last_segment)64    return segments65 66# Extract MFCC features67def extract_mfcc_from_segments(segments, sample_rate=16000, n_mfcc=13):68    features = []69    for segment in segments:70        mfcc = librosa.feature.mfcc(y=segment, sr=sample_rate, n_mfcc=n_mfcc)71        mfcc_mean = np.mean(mfcc, axis=1)72        mfcc_std = np.std(mfcc, axis=1)73        features.append(np.concatenate([mfcc_mean, mfcc_std]))74    return np.array(features)75 76# Predict audio lie/truth77def make_prediction_audio(input_video):78    if not check_audio_in_video(input_video):79        return [1]  # Default to truth if no audio80    audio_file = "extracted_audio.wav"81    extract_audio_from_video(input_video, audio_file)82    segments = split_and_pad_audio(audio_file)83    mfcc_features = extract_mfcc_from_segments(segments)84    mfcc_features = scaler.transform(mfcc_features)85    mfcc_features = np.expand_dims(mfcc_features, axis=1)86    predictions = audio_model.predict(mfcc_features)87    return 1 - np.argmax(predictions, axis=1)88 89# Predict facial lie/truth90def make_prediction(input_video):91    video_capture = cv2.VideoCapture(input_video)92    if not video_capture.isOpened():93        raise HTTPException(status_code=400, detail="Could not open video")94    if not input_validation(input_video):95        raise HTTPException(status_code=400, detail="Video should contain one person with a clear face")96    predictions = []97    frame_count = 098    cur_video32 = []99    while True:100        ret, frame = video_capture.read()101        if not ret:102            break103        frame_count += 1104        cur_video32.append(np.zeros(478 * 3))  # Placeholder since face mesh isn't implemented here105        if frame_count == 32:106            predictions.append(np.argmax(saved_model.predict(np.array([cur_video32]))))107            cur_video32 = []108            frame_count = 0109    video_capture.release()110    return predictions111 112# Final decision function113def final_decision(input_video):114    global latest_result115    voice_preds = make_prediction_audio(input_video)116    face_preds = make_prediction(input_video)117    voice_preds = np.repeat(voice_preds, 10)[:len(face_preds)]118    face_preds = np.pad(face_preds, (0, max(0, len(voice_preds) - len(face_preds))), constant_values=1)119    hard_vote = (voice_preds + face_preds) >= 1120    final_result = "Lie" if np.sum(hard_vote) > len(hard_vote) / 2 else "Truth"121    confidence = (max(np.sum(hard_vote), len(hard_vote) - np.sum(hard_vote)) / len(hard_vote)) * 100122    latest_result = {"final_decision": final_result, "confidence": confidence}123    return latest_result124 125@app.post("/process_video")126async def process_video(video: UploadFile = File(...)):127    file_path = os.path.join(UPLOAD_FOLDER, video.filename)128    with open(file_path, "wb") as f:129        f.write(await video.read())130    result = final_decision(file_path)131    return JSONResponse(content=result)132 133@app.get("/result")134async def get_result():135    if latest_result is None:136        raise HTTPException(status_code=404, detail="No analysis performed yet")137    return JSONResponse(content=latest_result)138