app.py

from hashlib import sha1
from pathlib import Path

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


def estimate_foreground_ml(image, alpha, return_background=False):
    """
    Estimates the foreground and background of an image based on an alpha mask.

    Parameters:
    - image: numpy array of shape (H, W, 3), the input RGB image.
    - alpha: numpy array of shape (H, W), the alpha mask with values ranging from 0 to 1.
    - return_background: boolean, if True, both foreground and background are returned.

    Returns:
    - If return_background is False, returns only the foreground.
    - If return_background is True, returns a tuple (foreground, background).
    """

    # Estimating foreground
    # Expand alpha dimensions from (H, W) to (H, W, 1) to make it compatible for element-wise multiplication with the RGB image
    foreground = image * alpha

    if return_background:
        # Estimating background
        # Inverse alpha mask to isolate background
        background_alpha = 1 - alpha
        # Assuming a white background. This can be changed based on the application or estimated from the image.
        background = (image * background_alpha) + (1 - background_alpha) * 255

        return foreground, background

    return foreground


def load_entire_model(taskname):
    model_ls = []
    if (taskname == "mask"):
        model = torch.jit.load(Path("./models/sod.pt"))
        model.eval()
        model_ls.append(model)
    elif (taskname == "matting"):
        model = torch.jit.load(Path("./models/trimap.pt"))
        model.eval()
        model_ls.append(model)

        model = torch.jit.load(Path("./models/matting.pt"))
        model.eval()
        model_ls.append(model)
    else:
        model_ls = []

    return model_ls


model_names = [
    "matting",
    "mask"
]
model_dict = {
    name: None
    for name in model_names
}

last_result = {
    "cache_key": None,
    "algorithm": None,
}


def image_matting(
        image: PIL.Image.Image,
        result_type: str,
        bg_color: str,
        algorithm: str,
        morph_op: str,
        morph_op_factor: float,
) -> np.ndarray:
    image_np = np.ascontiguousarray(image)
    width, height = image_np.shape[1], image_np.shape[0]
    cache_key = sha1(image_np).hexdigest()
    if cache_key == last_result["cache_key"] and algorithm == last_result["algorithm"]:
        alpha = last_result["alpha"]
    else:
        model = load_entire_model(algorithm)
        transform = transforms.Compose([
            # transforms.ToPILImage(),
            transforms.Resize((798, 798)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])
        if (algorithm == "mask"):
            input_tensor = transform(image).unsqueeze(0)
            with torch.no_grad():
                alpha = model[0](input_tensor).float()
                alpha = F.interpolate(alpha, [height, width], mode="bilinear")
            alpha = np.array(alpha* 255.).astype(np.uint8)[0][0]
            alpha = np.stack((alpha,alpha,alpha),axis=2)
        else:
            transform2 = transforms.Compose([
                transforms.Resize((800, 800)),
                transforms.ToTensor(),
                # transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
            ])

            input_tensor = transform(image).unsqueeze(0)
            with torch.no_grad():
                output = model[0](input_tensor).float()
                output = F.interpolate(output, [height, width], mode="bilinear")

            trimap = np.array(output[0][0])

            ratio = 0.05
            site = np.where(trimap > 0)
            try:
                bbox = [np.min(site[1]), np.min(site[0]), np.max(site[1]), np.max(site[0])]
            except:
                bbox = [0, 0, width, height]

            x0, y0, x1, y1 = bbox
            H = y1 - y0
            W = x1 - x0
            x0 = int(max(0, x0 - ratio * W))
            x1 = int(min(width, x1 + ratio * W) )
            y0 = int(max(0, y0 - ratio * H) )
            y1 = int(min(height, y1 + ratio * H) )

            Image_input = image.crop((x0, y0, x1, y1))
            # Image_input.save('image.png')
            input_tensor = transform2(Image_input).unsqueeze(0)

            trimap = trimap[y0:y1, x0:x1]
            trimap = np.where(trimap < 1, 0, trimap)
            trimap = np.where(trimap > 1, 255, trimap)
            trimap = np.where(trimap == 1, 128, trimap)
            # cv2.imwrite("trimap.png", trimap)

            trimap = Image.fromarray(np.uint8(trimap)).convert('L')
            input_tensor2 = transform2(trimap).unsqueeze(0)
            with torch.no_grad():
                output = model[1]({'image': input_tensor, 'trimap': input_tensor2})['phas']
                output = F.interpolate(output, [Image_input.size[1],Image_input.size[0]], mode="bilinear")[0].numpy()

            numpy_image = (output * 255.).astype(np.uint8)  # Scale to [0, 255] and convert to uint8

            # Step 4: Remove the channel dimension since it's a grayscale image
            numpy_image = numpy_image.squeeze(0)  # Convert from (1, H, W) to (H, W)
            pil_image = Image.fromarray(numpy_image, mode='L')
            alpha = Image.new(mode='RGB', size=image.size)
            alpha.paste(pil_image, (x0, y0))
            # alpha.save('tmp.png')

        alpha = np.array(alpha).astype(np.uint8)
        last_result["cache_key"] = cache_key
        last_result["algorithm"] = algorithm
        last_result["alpha"] = alpha

    # alpha = (alpha * 255).astype(np.uint8)
    image = np.array(image)
    kernel = np.ones((morph_op_factor, morph_op_factor), np.uint8)
    if morph_op == "Dilate":
        alpha = cv2.dilate(alpha, kernel, iterations=int(morph_op_factor))
    elif morph_op == "Erode":
        alpha = cv2.erode(alpha, kernel, iterations=int(morph_op_factor))
    else:
        alpha = alpha
    alpha = (alpha / 255).astype("float32")

    image = (image / 255.0).astype("float32")
    fg = estimate_foreground_ml(image, alpha)

    if result_type == "Remove BG":
        result = fg
    elif result_type == "Replace BG":
        bg_r = int(bg_color[1:3], base=16)
        bg_g = int(bg_color[3:5], base=16)
        bg_b = int(bg_color[5:7], base=16)

        bg = np.zeros_like(fg)
        bg[:, :, 0] = bg_r / 255.
        bg[:, :, 1] = bg_g / 255.
        bg[:, :, 2] = bg_b / 255.

        result = alpha * image + (1 - alpha) * bg
        result = np.clip(result, 0, 1)
    else:
        result = alpha

    return result


def main():
    with gr.Blocks() as app:
        gr.Markdown("Salient Object Matting")

        with gr.Row(variant="panel"):
            image_input = gr.Image(type='pil')
            image_output = gr.Image()

        with gr.Row(variant="panel"):
            result_type = gr.Radio(
                label="Mode",
                show_label=True,
                choices=[
                    "Remove BG",
                    "Replace BG",
                    "Generate Mask",
                ],
                value="Remove BG",
            )
            bg_color = gr.ColorPicker(
                label="BG Color",
                show_label=True,
                value="#000000",
            )
            algorithm = gr.Dropdown(
                label="Algorithm",
                show_label=True,
                choices=model_names,
                value="matting"
            )

        with gr.Row(variant="panel"):
            morph_op = gr.Radio(
                label="Post-process",
                show_label=True,
                choices=[
                    "None",
                    "Erode",
                    "Dilate",
                ],
                value="None",
            )

            morph_op_factor = gr.Slider(
                label="Factor",
                show_label=True,
                minimum=3,
                maximum=20,
                value=3,
                step=2,
            )

        run_button = gr.Button("Run")

        run_button.click(
            image_matting,
            inputs=[
                image_input,
                result_type,
                bg_color,
                algorithm,
                morph_op,
                morph_op_factor,
            ],
            outputs=image_output,
        )

    app.launch()


if __name__ == "__main__":
    main()