model/modnet.py

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


# Backbone: U2NET-like architecture (simplified for inference only)
class BasicConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(BasicConvBlock, self).__init__()
        self.block = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, 3, 1, 1),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(out_channels, out_channels, 3, 1, 1),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),
        )

    def forward(self, x):
        return self.block(x)


class SimpleMODNetBackbone(nn.Module):
    def __init__(self):
        super(SimpleMODNetBackbone, self).__init__()
        self.stage1 = BasicConvBlock(3, 64)
        self.pool1 = nn.MaxPool2d(2, 2)
        self.stage2 = BasicConvBlock(64, 128)
        self.pool2 = nn.MaxPool2d(2, 2)
        self.stage3 = BasicConvBlock(128, 256)

    def forward(self, x):
        x = self.stage1(x)
        x = self.pool1(x)
        x = self.stage2(x)
        x = self.pool2(x)
        x = self.stage3(x)
        return x


class MODNet(nn.Module):
    def __init__(self):
        super(MODNet, self).__init__()
        self.backbone = SimpleMODNetBackbone()
        self.seg_head = nn.Sequential(
            nn.Conv2d(256, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(64, 1, kernel_size=1),
            nn.Sigmoid()
        )

    def forward(self, x):
        features = self.backbone(x)
        pred_matte = self.seg_head(features)
        return pred_matte


def preprocess_image(image: Image.Image, device: torch.device) -> torch.Tensor:
    transform = transforms.Compose([
        transforms.Resize((512, 512)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ])
    img_tensor = transform(image.convert("RGB")).unsqueeze(0).to(device)
    return img_tensor


def remove_background_modnet(image: Image.Image) -> Image.Image:
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    modnet = MODNet()
    modnet.to(device)

    # Skip loading weights (simple version)
    modnet.eval()

    img_tensor = preprocess_image(image, device)

    with torch.no_grad():
        pred_matte = modnet(img_tensor)

    matte = pred_matte[0][0].cpu().numpy()
    matte = cv2.resize(matte, image.size)
    matte = np.uint8(matte * 255)

    image = image.convert("RGBA")
    image_np = np.array(image)
    image_np[:, :, 3] = matte
    return Image.fromarray(image_np)