|
import gradio as gr |
|
import numpy as np |
|
import torch |
|
import torch.nn as nn |
|
import torch.optim as optim |
|
import torch.nn.functional as F |
|
import cv2 |
|
import PIL.Image |
|
from scipy.interpolate import griddata |
|
import matplotlib.pyplot as plt |
|
|
|
def RGB2gray(rgb): |
|
r, g, b = rgb[:,:,0], rgb[:,:,1], rgb[:,:,2] |
|
gray = 0.2989 * r + 0.5870 * g + 0.1140 * b |
|
return gray |
|
|
|
|
|
def img_to_patches(img: PIL.Image.Image) -> tuple: |
|
patch_size = 16 |
|
img = img.convert('RGB') |
|
|
|
grayscale_imgs = [] |
|
imgs = [] |
|
coordinates = [] |
|
|
|
for i in range(0, img.height, patch_size): |
|
for j in range(0, img.width, patch_size): |
|
box = (j, i, j + patch_size, i + patch_size) |
|
img_color = np.asarray(img.crop(box)) |
|
grayscale_image = cv2.cvtColor(src=img_color, code=cv2.COLOR_RGB2GRAY) |
|
grayscale_imgs.append(grayscale_image.astype(dtype=np.int32)) |
|
imgs.append(img_color) |
|
normalized_coord = (i + patch_size // 2, j + patch_size // 2) |
|
coordinates.append(normalized_coord) |
|
|
|
return grayscale_imgs, imgs, coordinates, (img.height, img.width) |
|
|
|
def get_l1(v): |
|
return np.sum(np.abs(v[:, :-1] - v[:, 1:])) |
|
|
|
def get_l2(v): |
|
return np.sum(np.abs(v[:-1, :] - v[1:, :])) |
|
|
|
def get_l3l4(v): |
|
l3 = np.sum(np.abs(v[:-1, :-1] - v[1:, 1:])) |
|
l4 = np.sum(np.abs(v[1:, :-1] - v[:-1, 1:])) |
|
return l3 + l4 |
|
|
|
def get_pixel_var_degree_for_patch(patch: np.array) -> int: |
|
l1 = get_l1(patch) |
|
l2 = get_l2(patch) |
|
l3l4 = get_l3l4(patch) |
|
return l1 + l2 + l3l4 |
|
|
|
def get_rich_poor_patches(img: PIL.Image.Image, coloured=True): |
|
gray_scale_patches, color_patches, coordinates, img_size = img_to_patches(img) |
|
var_with_patch = [] |
|
for i, patch in enumerate(gray_scale_patches): |
|
if coloured: |
|
var_with_patch.append((get_pixel_var_degree_for_patch(patch), color_patches[i], coordinates[i])) |
|
else: |
|
var_with_patch.append((get_pixel_var_degree_for_patch(patch), patch, coordinates[i])) |
|
|
|
var_with_patch.sort(reverse=True, key=lambda x: x[0]) |
|
mid_point = len(var_with_patch) // 2 |
|
r_patch = [(patch, coor) for var, patch, coor in var_with_patch[:mid_point]] |
|
p_patch = [(patch, coor) for var, patch, coor in var_with_patch[mid_point:]] |
|
p_patch.reverse() |
|
return r_patch, p_patch, img_size |
|
|
|
def azimuthalAverage(image, center=None): |
|
y, x = np.indices(image.shape) |
|
if not center: |
|
center = np.array([(x.max() - x.min()) / 2.0, (y.max() - y.min()) / 2.0]) |
|
r = np.hypot(x - center[0], y - center[1]) |
|
ind = np.argsort(r.flat) |
|
r_sorted = r.flat[ind] |
|
i_sorted = image.flat[ind] |
|
r_int = r_sorted.astype(int) |
|
deltar = r_int[1:] - r_int[:-1] |
|
rind = np.where(deltar)[0] |
|
nr = rind[1:] - rind[:-1] |
|
csim = np.cumsum(i_sorted, dtype=float) |
|
tbin = csim[rind[1:]] - csim[rind[:-1]] |
|
radial_prof = tbin / nr |
|
return radial_prof |
|
|
|
def azimuthal_integral(img, epsilon=1e-8, N=50): |
|
if len(img.shape) == 3 and img.shape[2] == 3: |
|
img = RGB2gray(img) |
|
f = np.fft.fft2(img) |
|
fshift = np.fft.fftshift(f) |
|
fshift += epsilon |
|
magnitude_spectrum = 20 * np.log(np.abs(fshift)) |
|
psd1D = azimuthalAverage(magnitude_spectrum) |
|
points = np.linspace(0, N, num=psd1D.size) |
|
xi = np.linspace(0, N, num=N) |
|
interpolated = griddata(points, psd1D, xi, method='cubic') |
|
interpolated = (interpolated - np.min(interpolated)) / (np.max(interpolated) - np.min(interpolated)) |
|
return interpolated.astype(np.float32) |
|
|
|
def positional_emb(coor, im_size, N): |
|
img_height, img_width = im_size |
|
center_y, center_x = coor |
|
normalized_y = center_y / img_height |
|
normalized_x = center_x / img_width |
|
pos_emb = np.zeros(N) |
|
indices = np.arange(N) |
|
div_term = 10000 ** (2 * (indices // 2) / N) |
|
pos_emb[0::2] = np.sin(normalized_y / div_term[0::2]) + np.sin(normalized_x / div_term[0::2]) |
|
pos_emb[1::2] = np.cos(normalized_y / div_term[1::2]) + np.cos(normalized_x / div_term[1::2]) |
|
return pos_emb |
|
|
|
def azi_diff(img: PIL.Image.Image, patch_num, N): |
|
r, p, im_size = get_rich_poor_patches(img) |
|
r_len = len(r) |
|
p_len = len(p) |
|
patch_emb_r = np.zeros((patch_num, N)) |
|
patch_emb_p = np.zeros((patch_num, N)) |
|
positional_emb_r = np.zeros((patch_num, N)) |
|
positional_emb_p = np.zeros((patch_num, N)) |
|
coor_r = [] |
|
coor_p = [] |
|
if r_len != 0: |
|
for idx in range(patch_num): |
|
tmp_patch1 = r[idx % r_len][0] |
|
tmp_coor1 = r[idx % r_len][1] |
|
patch_emb_r[idx] = azimuthal_integral(tmp_patch1, N=N) |
|
positional_emb_r[idx] = positional_emb(tmp_coor1, im_size, N) |
|
coor_r.append(tmp_coor1) |
|
if p_len != 0: |
|
for idx in range(patch_num): |
|
tmp_patch2 = p[idx % p_len][0] |
|
tmp_coor2 = p[idx % p_len][1] |
|
patch_emb_p[idx] = azimuthal_integral(tmp_patch2, N=N) |
|
positional_emb_p[idx] = positional_emb(tmp_coor2, im_size, N) |
|
coor_p.append(tmp_coor2) |
|
output = {"total_emb": [patch_emb_r + positional_emb_r / 5, patch_emb_p + positional_emb_p / 5], |
|
"positional_emb": [positional_emb_r / 5, positional_emb_p / 5], "coor": [coor_r, coor_p], |
|
"image_size": im_size} |
|
return output |
|
|
|
class AttentionBlock(nn.Module): |
|
def __init__(self, input_dim, num_heads, ff_dim, rate=0.1): |
|
super(AttentionBlock, self).__init__() |
|
self.attention = nn.MultiheadAttention(embed_dim=input_dim, num_heads=num_heads) |
|
self.dropout1 = nn.Dropout(rate) |
|
self.layer_norm1 = nn.LayerNorm(input_dim) |
|
self.ffn = nn.Sequential( |
|
nn.Linear(input_dim, ff_dim), |
|
nn.ReLU(), |
|
nn.Dropout(rate), |
|
nn.Linear(ff_dim, input_dim), |
|
nn.Dropout(rate) |
|
) |
|
self.layer_norm2 = nn.LayerNorm(input_dim) |
|
|
|
def forward(self, x): |
|
attn_output, _ = self.attention(x, x, x) |
|
attn_output = self.dropout1(attn_output) |
|
out1 = self.layer_norm1(attn_output + x) |
|
ffn_output = self.ffn(out1) |
|
out2 = self.layer_norm2(ffn_output + out1) |
|
return out2 |
|
|
|
class TextureContrastClassifier(nn.Module): |
|
def __init__(self, input_shape, num_heads=4, key_dim=64, ff_dim=256, rate=0.1): |
|
super(TextureContrastClassifier, self).__init__() |
|
input_dim = input_shape[1] |
|
self.rich_attention_block = AttentionBlock(input_dim, num_heads, ff_dim, rate) |
|
self.rich_dense = nn.Sequential( |
|
nn.Linear(input_dim, 128), |
|
nn.ReLU(), |
|
nn.Dropout(0.5) |
|
) |
|
self.poor_attention_block = AttentionBlock(input_dim, num_heads, ff_dim, rate) |
|
self.poor_dense = nn.Sequential( |
|
nn.Linear(input_dim, 128), |
|
nn.ReLU(), |
|
nn.Dropout(0.5) |
|
) |
|
self.fc = nn.Sequential( |
|
nn.Linear(128 * input_shape[0], 256), |
|
nn.ReLU(), |
|
nn.Dropout(0.5), |
|
nn.Linear(256, 128), |
|
nn.ReLU(), |
|
nn.Dropout(0.5), |
|
nn.Linear(128, 64), |
|
nn.ReLU(), |
|
nn.Dropout(0.5), |
|
nn.Linear(64, 32), |
|
nn.ReLU(), |
|
nn.Dropout(0.5), |
|
nn.Linear(32, 16), |
|
nn.ReLU(), |
|
nn.Dropout(0.5), |
|
nn.Linear(16, 1), |
|
nn.Sigmoid() |
|
) |
|
|
|
def forward(self, rich_texture, poor_texture): |
|
rich_texture = rich_texture.permute(1, 0, 2) |
|
poor_texture = poor_texture.permute(1, 0, 2) |
|
rich_attention = self.rich_attention_block(rich_texture) |
|
rich_attention = rich_attention.permute(1, 0, 2) |
|
rich_features = self.rich_dense(rich_attention) |
|
poor_attention = self.poor_attention_block(poor_texture) |
|
poor_attention = poor_attention.permute(1, 0, 2) |
|
poor_features = self.poor_dense(poor_attention) |
|
difference = rich_features - poor_features |
|
difference = difference.view(difference.size(0), -1) |
|
output = self.fc(difference) |
|
return output |
|
|
|
input_shape = (128, 256) |
|
model = TextureContrastClassifier(input_shape) |
|
model.load_state_dict(torch.load('./model_epoch_36.pth', map_location=torch.device('cpu'))) |
|
|
|
def inference(image, model): |
|
predictions = [] |
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
model.to(device) |
|
model.eval() |
|
tmp = azi_diff(image, patch_num=128, N=256) |
|
rich = tmp["total_emb"][0] |
|
poor = tmp["total_emb"][1] |
|
rich_texture_tensor = torch.tensor(rich, dtype=torch.float32).unsqueeze(0).to(device) |
|
poor_texture_tensor = torch.tensor(poor, dtype=torch.float32).unsqueeze(0).to(device) |
|
with torch.no_grad(): |
|
output = model(rich_texture_tensor, poor_texture_tensor) |
|
prediction = output.cpu().numpy().flatten()[0] |
|
return prediction |
|
|
|
|
|
def predict(image): |
|
prediction = inference(image, model) |
|
return f"{prediction * 100:.2f}% chance AI-generated" |
|
|
|
gr.Interface(fn=predict, inputs=gr.Image(type="pil"), outputs="text").launch() |
|
|