def detect_image(filename):
# Load data
img = np.array(Image.open(filename).convert(params["channel"]))
coords, _, _ = image.slide2d(sz=img.shape[:2], K=64, S=32)
patches = image.crop_patches(img=img, coords=coords, patch_sz=64)
loader = NumpyImageLoader(ndarray_data=patches, batch_size=16, n_workers=cpu_count(), pin_memory=True,
shuffle=False).loader
# Predict
softmaxs = []
model.eval()
for X in loader:
X = X[0].to(DEVICE)
logits = model(X)
softmaxs.append(F.softmax(logits, dim=1).detach().cpu().numpy())
softmaxs = np.concatenate(softmaxs, axis=0)
# Post-processing
labels = image.post_process(softmaxs[:, 1], coords, 8, params["threshold"], 32, pools=pools)
decision = image.fusion(labels)
return decision #1 fake, 0 real
# Test on authentic images
for i, file in tqdm(enumerate(au_files), total=n_au_files):
# Load softmaxs and coords from disk
data = np.load(file).item()
softmaxs, coords = data["softmaxs"], data["coords"]
softmaxs = softmaxs[:, 1]
# Postprocess
labels = image.post_process(softmaxs,
coords,
8,
params["threshold"],
32,
pools=pools)
mark = image.fusion(labels)
scores_au.append(mark)
# In[8]:
# Test on tampered images
for i, file in tqdm(enumerate(tp_files), total=n_tp_files):
# Load softmaxs and coords from disk
data = np.load(file).item()
softmaxs, coords = data["softmaxs"], data["coords"]
softmaxs = softmaxs[:, 1]
# Postprocess
labels = image.post_process(softmaxs,
coords,
coords, _, _ = image.slide2d(sz=img.shape[:2], K=64, S=32)
patches = image.crop_patches( img=img, coords=coords, patch_sz=64)
loader = NumpyImageLoader(ndarray_data=patches, batch_size=16, n_workers=cpu_count(), pin_memory=True, shuffle=False).loader
# Predict
softmaxs = []
model.eval()
for X in loader:
X = X[0].to(DEVICE)
logits = model(X)
softmaxs.append(F.softmax(logits, dim=1).detach().cpu().numpy())
softmaxs = np.concatenate(softmaxs, axis=0)
# Post-processing
labels = image.post_process(softmaxs[:,1], coords, 8, params["threshold"], 32, pools=pools)
decision = image.fusion(labels)
if decision==1:
if GT=="Positive":
print("Prediction: Positive ==> True")
else:
print("Prediction: Positive ==> False")
else:
if GT=="Positive":
print("Prediction: Negative ==> False")
else:
print("Prediction: Negative ==> True")
# Draw score_map
score_map = image.reconstruct_heatmap(softmaxs[:,1], coords, img.shape, 64)
plt.figure(figsize=(16,8))