app.py

import cv2
import torch
import gradio as gr
import numpy as np
from PIL import Image
import torch.nn.functional as F

from model import DINOv2DPT, compute_CD_heatmap, cd_to_tensor

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

IMG_SIZE = 518
CHECKPOINT = "best.pt"


# =========================
# LOAD MODEL
# =========================

model = DINOv2DPT(
    img_size=IMG_SIZE,
    features=256,
    hook_indices=(2, 5, 8, 11),
    unfreeze_blocks=3
)

state_dict = torch.load(CHECKPOINT, map_location=DEVICE)
model.load_state_dict(state_dict)

model.to(DEVICE)
model.eval()


# =========================
# INFERENCE
# =========================

def predict(image):

    image_np = np.array(image).astype(np.uint8)

    original_h, original_w = image_np.shape[:2]

    resized = cv2.resize(image_np, (IMG_SIZE, IMG_SIZE))

    input_tensor = (
        torch.tensor(resized / 255.0, dtype=torch.float32)
        .permute(2, 0, 1)
        .unsqueeze(0)
        .to(DEVICE)
    )

    # CD MAP
    image_bgr = cv2.cvtColor(resized, cv2.COLOR_RGB2BGR)

    cd_map = compute_CD_heatmap(image_bgr)

    cd_tensor = cd_to_tensor(cd_map, DEVICE)

    # MODEL
    with torch.no_grad():

        logits = model(input_tensor, cd_map=cd_tensor)

        probs = torch.sigmoid(logits)[0, 0]

    mask = probs.cpu().numpy()

    mask = cv2.resize(mask, (original_w, original_h))

    binary_mask = (mask > 0.5).astype(np.uint8)

    # VISUALIZATION
    mask_vis = (binary_mask * 255).astype(np.uint8)

    overlay = image_np.copy()

    overlay[binary_mask == 1] = [255, 0, 0]

    overlay = cv2.addWeighted(image_np, 0.7, overlay, 0.3, 0)

    return (
        # image_np,
        mask_vis,
        overlay
    )


# =========================
# EXAMPLES
# =========================

examples = [
    "examples/image.png",
    "examples/0008.jpg",
    "examples/apricot.jpg",
]


# =========================
# UI
# =========================
with gr.Blocks(theme=gr.themes.Soft()) as demo:

    gr.Markdown(
        """
        # Semantic Transformer based Adversarial Patch Segmentation

        Upload an image to detect adversarial patches using DINOv2 + DPT.
        """
    )

    with gr.Row():

        # =====================================
        # LEFT PANEL
        # =====================================

        with gr.Column(scale=1):

            input_image = gr.Image(
                type="pil",
                label="Input Image",
                height=400
            )

            run_btn = gr.Button(
                "Run Detection",
                variant="primary"
            )

            gr.Examples(
                examples=examples,
                inputs=input_image,
            )

        # =====================================
        # RIGHT PANEL
        # =====================================

        with gr.Column(scale=2):

            with gr.Tab("Overlay"):

                overlay_output = gr.Image(
                    label="Patch Detection Overlay",
                    height=700
                )

            with gr.Tab("Binary Mask"):

                mask_output = gr.Image(
                    label="Predicted Mask",
                    height=700
                )

    run_btn.click(
        fn=predict,
        inputs=input_image,
        outputs=[
            mask_output,
            overlay_output
        ]
    )

demo.queue().launch()