Spaces:
Sleeping
Sleeping
thugCodeNinja
commited on
Commit
•
119338f
1
Parent(s):
db21c2e
Update app.py
Browse files
app.py
CHANGED
@@ -2,8 +2,6 @@ import gradio as gr
|
|
2 |
import cv2
|
3 |
import numpy as np
|
4 |
import tensorflow as tf
|
5 |
-
import tensorflow_addons
|
6 |
-
|
7 |
from facenet_pytorch import MTCNN
|
8 |
from PIL import Image
|
9 |
import moviepy.editor as mp
|
@@ -13,61 +11,43 @@ import zipfile
|
|
13 |
# Load face detector
|
14 |
mtcnn = MTCNN(margin=14, keep_all=True, factor=0.7, device='cpu')
|
15 |
|
16 |
-
#Face Detection function
|
17 |
class DetectionPipeline:
|
18 |
-
"""Pipeline class for detecting faces in the frames of a video file."""
|
19 |
-
|
20 |
def __init__(self, detector, n_frames=None, batch_size=60, resize=None):
|
21 |
-
"""Constructor for DetectionPipeline class.
|
22 |
-
"""
|
23 |
self.detector = detector
|
24 |
self.n_frames = n_frames
|
25 |
self.batch_size = batch_size
|
26 |
self.resize = resize
|
27 |
|
28 |
def __call__(self, filename):
|
29 |
-
"""Load frames from an MP4 video and detect faces.
|
30 |
-
|
31 |
-
Arguments:
|
32 |
-
filename {str} -- Path to video.
|
33 |
-
"""
|
34 |
-
# Create video reader and find length
|
35 |
v_cap = cv2.VideoCapture(filename)
|
36 |
v_len = int(v_cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
37 |
|
38 |
-
# Pick 'n_frames' evenly spaced frames to sample
|
39 |
if self.n_frames is None:
|
40 |
sample = np.arange(0, v_len)
|
41 |
else:
|
42 |
sample = np.linspace(0, v_len - 1, self.n_frames).astype(int)
|
43 |
|
44 |
-
# Loop through frames
|
45 |
faces = []
|
46 |
frames = []
|
47 |
for j in range(v_len):
|
48 |
success = v_cap.grab()
|
49 |
if j in sample:
|
50 |
-
# Load frame
|
51 |
success, frame = v_cap.retrieve()
|
52 |
if not success:
|
53 |
continue
|
54 |
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
55 |
-
# frame = Image.fromarray(frame)
|
56 |
|
57 |
-
# Resize frame to desired size
|
58 |
if self.resize is not None:
|
59 |
-
frame =
|
|
|
60 |
frames.append(frame)
|
61 |
|
62 |
-
# When batch is full, detect faces and reset frame list
|
63 |
if len(frames) % self.batch_size == 0 or j == sample[-1]:
|
64 |
-
|
65 |
-
boxes, probs = self.detector.detect(frames)
|
66 |
-
|
67 |
for i in range(len(frames)):
|
68 |
-
|
69 |
if boxes[i] is None:
|
70 |
-
faces.append(face2)
|
71 |
continue
|
72 |
|
73 |
box = boxes[i][0].astype(int)
|
@@ -75,11 +55,10 @@ class DetectionPipeline:
|
|
75 |
face = frame[box[1]:box[3], box[0]:box[2]]
|
76 |
|
77 |
if not face.any():
|
78 |
-
faces.append(face2)
|
79 |
continue
|
80 |
|
81 |
face2 = cv2.resize(face, (224, 224))
|
82 |
-
|
83 |
faces.append(face2)
|
84 |
|
85 |
frames = []
|
@@ -88,14 +67,11 @@ class DetectionPipeline:
|
|
88 |
|
89 |
return faces
|
90 |
|
91 |
-
|
92 |
-
detection_pipeline = DetectionPipeline(detector=mtcnn,n_frames=20, batch_size=60)
|
93 |
|
94 |
model = tf.keras.models.load_model("p1")
|
95 |
|
96 |
-
|
97 |
def deepfakespredict(input_video):
|
98 |
-
|
99 |
faces = detection_pipeline(input_video)
|
100 |
|
101 |
total = 0
|
@@ -103,22 +79,21 @@ def deepfakespredict(input_video):
|
|
103 |
fake = 0
|
104 |
|
105 |
for face in faces:
|
106 |
-
|
107 |
-
face2 = face/255
|
108 |
pred = model.predict(np.expand_dims(face2, axis=0))[0]
|
109 |
-
total+=1
|
110 |
|
111 |
pred2 = pred[1]
|
112 |
|
113 |
if pred2 > 0.5:
|
114 |
-
|
115 |
else:
|
116 |
-
|
117 |
|
118 |
-
fake_ratio = fake/total
|
119 |
|
120 |
-
text =""
|
121 |
-
text2 = "Deepfakes Confidence: " + str(fake_ratio*100) + "%"
|
122 |
|
123 |
if fake_ratio >= 0.5:
|
124 |
text = "The video is FAKE."
|
@@ -126,27 +101,25 @@ def deepfakespredict(input_video):
|
|
126 |
text = "The video is REAL."
|
127 |
|
128 |
face_frames = []
|
129 |
-
|
130 |
for face in faces:
|
131 |
face_frame = Image.fromarray(face.astype('uint8'), 'RGB')
|
132 |
face_frames.append(face_frame)
|
133 |
-
|
134 |
-
face_frames[0].save('results.gif', save_all=True, append_images=face_frames[1:], duration
|
135 |
clip = mp.VideoFileClip("results.gif")
|
136 |
clip.write_videofile("video.mp4")
|
137 |
|
138 |
return text, text2, "video.mp4"
|
139 |
|
140 |
-
|
141 |
-
|
142 |
-
title="Group 2- EfficientNetV2 based Deepfake Video Detector"
|
143 |
-
description='''Please upload videos responsibly and await the results in a gif. The approach in place includes breaking down the video into several frames followed by collecting
|
144 |
the frames that contain a face. Once these frames are collected the trained model attempts to predict if the face is fake or real and contribute to a deepfake confidence. This confidence level eventually
|
145 |
determines if the video can be considered a fake or not.'''
|
146 |
-
|
147 |
gr.Interface(deepfakespredict,
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
|
|
2 |
import cv2
|
3 |
import numpy as np
|
4 |
import tensorflow as tf
|
|
|
|
|
5 |
from facenet_pytorch import MTCNN
|
6 |
from PIL import Image
|
7 |
import moviepy.editor as mp
|
|
|
11 |
# Load face detector
|
12 |
mtcnn = MTCNN(margin=14, keep_all=True, factor=0.7, device='cpu')
|
13 |
|
14 |
+
# Face Detection function
|
15 |
class DetectionPipeline:
|
|
|
|
|
16 |
def __init__(self, detector, n_frames=None, batch_size=60, resize=None):
|
|
|
|
|
17 |
self.detector = detector
|
18 |
self.n_frames = n_frames
|
19 |
self.batch_size = batch_size
|
20 |
self.resize = resize
|
21 |
|
22 |
def __call__(self, filename):
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
v_cap = cv2.VideoCapture(filename)
|
24 |
v_len = int(v_cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
25 |
|
|
|
26 |
if self.n_frames is None:
|
27 |
sample = np.arange(0, v_len)
|
28 |
else:
|
29 |
sample = np.linspace(0, v_len - 1, self.n_frames).astype(int)
|
30 |
|
|
|
31 |
faces = []
|
32 |
frames = []
|
33 |
for j in range(v_len):
|
34 |
success = v_cap.grab()
|
35 |
if j in sample:
|
|
|
36 |
success, frame = v_cap.retrieve()
|
37 |
if not success:
|
38 |
continue
|
39 |
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
|
|
40 |
|
|
|
41 |
if self.resize is not None:
|
42 |
+
frame = cv2.resize(frame, (int(frame.shape[1] * self.resize), int(frame.shape[0] * self.resize)))
|
43 |
+
|
44 |
frames.append(frame)
|
45 |
|
|
|
46 |
if len(frames) % self.batch_size == 0 or j == sample[-1]:
|
47 |
+
boxes, _ = self.detector.detect(frames)
|
|
|
|
|
48 |
for i in range(len(frames)):
|
|
|
49 |
if boxes[i] is None:
|
50 |
+
faces.append(face2)
|
51 |
continue
|
52 |
|
53 |
box = boxes[i][0].astype(int)
|
|
|
55 |
face = frame[box[1]:box[3], box[0]:box[2]]
|
56 |
|
57 |
if not face.any():
|
58 |
+
faces.append(face2)
|
59 |
continue
|
60 |
|
61 |
face2 = cv2.resize(face, (224, 224))
|
|
|
62 |
faces.append(face2)
|
63 |
|
64 |
frames = []
|
|
|
67 |
|
68 |
return faces
|
69 |
|
70 |
+
detection_pipeline = DetectionPipeline(detector=mtcnn, n_frames=20, batch_size=60)
|
|
|
71 |
|
72 |
model = tf.keras.models.load_model("p1")
|
73 |
|
|
|
74 |
def deepfakespredict(input_video):
|
|
|
75 |
faces = detection_pipeline(input_video)
|
76 |
|
77 |
total = 0
|
|
|
79 |
fake = 0
|
80 |
|
81 |
for face in faces:
|
82 |
+
face2 = face / 255
|
|
|
83 |
pred = model.predict(np.expand_dims(face2, axis=0))[0]
|
84 |
+
total += 1
|
85 |
|
86 |
pred2 = pred[1]
|
87 |
|
88 |
if pred2 > 0.5:
|
89 |
+
fake += 1
|
90 |
else:
|
91 |
+
real += 1
|
92 |
|
93 |
+
fake_ratio = fake / total
|
94 |
|
95 |
+
text = ""
|
96 |
+
text2 = "Deepfakes Confidence: " + str(fake_ratio * 100) + "%"
|
97 |
|
98 |
if fake_ratio >= 0.5:
|
99 |
text = "The video is FAKE."
|
|
|
101 |
text = "The video is REAL."
|
102 |
|
103 |
face_frames = []
|
104 |
+
|
105 |
for face in faces:
|
106 |
face_frame = Image.fromarray(face.astype('uint8'), 'RGB')
|
107 |
face_frames.append(face_frame)
|
108 |
+
|
109 |
+
face_frames[0].save('results.gif', save_all=True, append_images=face_frames[1:], duration=250, loop=100)
|
110 |
clip = mp.VideoFileClip("results.gif")
|
111 |
clip.write_videofile("video.mp4")
|
112 |
|
113 |
return text, text2, "video.mp4"
|
114 |
|
115 |
+
title = "Group 2- EfficientNetV2 based Deepfake Video Detector"
|
116 |
+
description = '''Please upload videos responsibly and await the results in a gif. The approach in place includes breaking down the video into several frames followed by collecting
|
|
|
|
|
117 |
the frames that contain a face. Once these frames are collected the trained model attempts to predict if the face is fake or real and contribute to a deepfake confidence. This confidence level eventually
|
118 |
determines if the video can be considered a fake or not.'''
|
119 |
+
|
120 |
gr.Interface(deepfakespredict,
|
121 |
+
inputs=["video"],
|
122 |
+
outputs=["text", "text", gr.Video(label="Detected face sequence")],
|
123 |
+
title=title,
|
124 |
+
description=description
|
125 |
+
).launch()
|