app.py

import cv2
from skimage.restoration import denoise_nl_means, estimate_sigma
import gradio as gr
import os
from PIL import Image
import numpy as np
import torch
from torch.autograd import Variable
from torchvision import transforms
import torch.nn.functional as F
import gdown
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings("ignore")

os.system("git clone https://github.com/xuebinqin/DIS")
os.system("mv DIS/IS-Net/* .")

# project imports
from data_loader_cache import normalize, im_reader, im_preprocess 
from models import *

#Helpers
device = 'cuda' if torch.cuda.is_available() else 'cpu'

# Download official weights
if not os.path.exists("saved_models"):
    os.mkdir("saved_models")
    os.system("mv isnet.pth saved_models/")
    
class GOSNormalize(object):
    '''
    Normalize the Image using torch.transforms
    '''
    def __init__(self, mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]):
        self.mean = mean
        self.std = std

    def __call__(self,image):
        image = normalize(image,self.mean,self.std)
        return image


transform =  transforms.Compose([GOSNormalize([0.5,0.5,0.5],[1.0,1.0,1.0])])

def load_image(image, hypar):
    """
    Load and preprocess an image.
    :param image: The image to load. This can be either a file path or a PIL.Image object.
    :param hypar: Hyperparameters for preprocessing.
    :return: A tuple of the preprocessed image tensor and its original shape.
    """
    # Check if the image is a file path or a PIL.Image object
    if isinstance(image, str):
        # If it's a file path, read the image from disk
        im = im_reader(image)
    elif isinstance(image, Image.Image):
        # If it's a PIL.Image object, convert it to a NumPy array
        im = np.array(image)
    else:
        raise TypeError("Unsupported image type")

    # Preprocess the image
    im, im_shp = im_preprocess(im, hypar["cache_size"])
    im = torch.divide(im, 255.0)
    shape = torch.from_numpy(np.array(im_shp))

    # Normalize and add batch dimension
    im = transform(im).unsqueeze(0)
    shape = shape.unsqueeze(0)  # Add batch dimension to shape

    return im, shape


def build_model(hypar,device):
    net = hypar["model"]#GOSNETINC(3,1)

    # convert to half precision
    if(hypar["model_digit"]=="half"):
        net.half()
        for layer in net.modules():
            if isinstance(layer, nn.BatchNorm2d):
                layer.float()

    net.to(device)

    if(hypar["restore_model"]!=""):
        net.load_state_dict(torch.load(hypar["model_path"]+"/"+hypar["restore_model"], map_location=device))
        net.to(device)
    net.eval()  
    return net


def crop_signature(original_image_path, mask, padding=32):
    # Convert the mask to a binary image
    _, binary_mask = cv2.threshold(mask, 127, 255, cv2.THRESH_BINARY)
    
    # Find contours from the binary mask
    contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    # Open the original image
    original_image = Image.open(original_image_path).convert("RGB")
    
    # If contours are found, proceed to crop
    if contours:
        # Find the combined bounding box of all contours
        min_x, min_y = original_image.width, original_image.height
        max_x = max_y = 0
        for contour in contours:
            x, y, w, h = cv2.boundingRect(contour)
            min_x, min_y = min(min_x, x), min(min_y, y)
            max_x, max_y = max(max_x, x + w), max(max_y, y + h)

        # Add padding to the bounding box
        x_padded = max(min_x - padding, 0)
        y_padded = max(min_y - padding, 0)
        w_padded = min(max_x + padding, original_image.width) - x_padded
        h_padded = min(max_y + padding, original_image.height) - y_padded

        # Crop the mask using the combined bounding box with padding
        cropped_mask = binary_mask[y_padded:y_padded+h_padded, x_padded:x_padded+w_padded]

        # Apply smoothing and denoising to the cropped mask
        smooth_denoised_mask = smooth_edges(cropped_mask)

        # Create an RGBA image with a black background and the denoised mask as the alpha channel
        mask_image = Image.new('RGBA', (w_padded, h_padded), (0, 0, 0))
        pil_mask = Image.fromarray(cropped_mask).convert('L')
        mask_image.putalpha(pil_mask)

        return mask_image

    # If no contours are found, return the original image
    return original_image

'''
def crop_signature(original_image_path, mask, padding=32):
    """
    Crop the signature from the original image using the provided mask.

    :param original_image_path: The file path of the original image.
    :param mask: The binary mask of the signature.
    :param padding: Padding to add around the bounding box of the signature.
    :return: Cropped image containing the signature.
    """
    # Convert the mask to a binary image
    _, binary_mask = cv2.threshold(mask, 127, 255, cv2.THRESH_BINARY)

    # Find contours from the binary mask
    contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    # Open the original image
    original_image = Image.open(original_image_path).convert("RGB")

    # If contours are found, proceed to crop
    if contours:
        # Find the combined bounding box of all contours
        min_x, min_y = original_image.width, original_image.height
        max_x = max_y = 0

        for contour in contours:
            x, y, w, h = cv2.boundingRect(contour)
            min_x, min_y = min(min_x, x), min(min_y, y)
            max_x, max_y = max(max_x, x + w), max(max_y, y + h)

        # Add padding to the bounding box
        x_padded = max(min_x - padding, 0)
        y_padded = max(min_y - padding, 0)
        w_padded = min(max_x + padding, original_image.width) - x_padded
        h_padded = min(max_y + padding, original_image.height) - y_padded

        # Crop the original image using the combined bounding box with padding
        cropped_image = original_image.crop((x_padded, y_padded, x_padded + w_padded, y_padded + h_padded))

        return cropped_image

    # If no contours are found, return the original image
    return original_image
'''

def smooth_and_denoise(mask):
    """
    Apply smoothing and denoising to the mask.
    :param mask: The binary mask of the signature.
    :return: Processed mask.
    """
    # Ensure the mask is a 2D array
    if mask.ndim > 2:
        mask = mask[..., 0]

    # Apply Gaussian Blurring for smoothing
    smoothed_mask = cv2.GaussianBlur(mask, (5, 5), 0)

    # Estimate noise standard deviation from the image
    sigma_est = np.mean(estimate_sigma(smoothed_mask, channel_axis=None))

    # Apply Non-Local Means Denoising
    denoised_mask = denoise_nl_means(smoothed_mask, h=1.15 * sigma_est, fast_mode=True,
                                     patch_size=5, patch_distance=3, channel_axis=None)
    return denoised_mask


def smooth_edges(mask):
    """
    Smooth edges of a binary mask using morphological operations and anti-aliasing resizing.
    :param mask: The binary mask of the signature.
    :return: Mask with smoothed edges.
    """
    # Convert mask to uint8 type if it isn't already
    mask = mask.astype(np.uint8)
    
    # Define a kernel for morphological operations
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
    
    # Use morphological close operation to close small holes in the mask
    closing = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel, iterations=2)
    
    # Use morphological open operation to remove noise
    opening = cv2.morphologyEx(closing, cv2.MORPH_OPEN, kernel, iterations=2)
    
    # Dilate the mask to make the signature slightly thicker
    dilated = cv2.dilate(opening, kernel, iterations=1)
    
    # Convert dilated mask to a PIL image for anti-aliasing resizing
    pil_mask = Image.fromarray(dilated)
    
    # Resize the image to a smaller size, then back to the original size
    small_size = (pil_mask.width // 2, pil_mask.height // 2)
    pil_mask_small = pil_mask.resize(small_size, Image.Resampling.LANCZOS)
    pil_mask_smooth = pil_mask_small.resize(pil_mask.size, Image.Resampling.LANCZOS)
    
    # Convert back to a numpy array
    smoothed_mask = np.array(pil_mask_smooth)
    final_mask = cv2.bilateralFilter(smoothed_mask, d=9, sigmaColor=75, sigmaSpace=75)

    return final_mask

    
def predict(net,  inputs_val, shapes_val, hypar, device):
    '''
    Given an Image, predict the mask
    '''
    net.eval()

    if(hypar["model_digit"]=="full"):
        inputs_val = inputs_val.type(torch.FloatTensor)
    else:
        inputs_val = inputs_val.type(torch.HalfTensor)

  
    inputs_val_v = Variable(inputs_val, requires_grad=False).to(device) # wrap inputs in Variable
   
    ds_val = net(inputs_val_v)[0] # list of 6 results

    pred_val = ds_val[0][0,:,:,:] # B x 1 x H x W    # we want the first one which is the most accurate prediction

    ## recover the prediction spatial size to the orignal image size
    pred_val = torch.squeeze(F.upsample(torch.unsqueeze(pred_val,0),(shapes_val[0][0],shapes_val[0][1]),mode='bilinear'))

    ma = torch.max(pred_val)
    mi = torch.min(pred_val)
    pred_val = (pred_val-mi)/(ma-mi) # max = 1

    if device == 'cuda': torch.cuda.empty_cache()
    return (pred_val.detach().cpu().numpy()*255).astype(np.uint8) # it is the mask we need
    
# Set Parameters
hypar = {} # paramters for inferencing


hypar["model_path"] ="./saved_models" ## load trained weights from this path
hypar["restore_model"] = "isnet.pth" ## name of the to-be-loaded weights
hypar["interm_sup"] = False ## indicate if activate intermediate feature supervision

##  choose floating point accuracy --
hypar["model_digit"] = "full" ## indicates "half" or "full" accuracy of float number
hypar["seed"] = 0

hypar["cache_size"] = [1024, 1024] ## cached input spatial resolution, can be configured into different size

## data augmentation parameters ---
hypar["input_size"] = [1024, 1024] ## mdoel input spatial size, usually use the same value hypar["cache_size"], which means we don't further resize the images
hypar["crop_size"] = [1024, 1024] ## random crop size from the input, it is usually set as smaller than hypar["cache_size"], e.g., [920,920] for data augmentation

hypar["model"] = ISNetDIS()

 # Build Model
net = build_model(hypar, device)


def inference(image):
    image_path = image
    
    image_tensor, orig_size = load_image(image_path, hypar) 
    original_mask = predict(net, image_tensor, orig_size, hypar, device)
     
    # Process the original mask with smoothing and denoising
    processed_mask = smooth_and_denoise(original_mask)

    # Convert processed mask to PIL image
    pil_processed_mask = Image.fromarray((processed_mask * 255).astype(np.uint8)).convert('L')
    pil_original_mask = Image.fromarray(original_mask).convert('L')
    
    im_rgb = Image.open(image).convert("RGB")
    im_dark = Image.new('RGB', im_rgb.size, (0, 0, 0))
    cropped_signature_image = crop_signature(image_path, original_mask, 64)
    
    # Apply processed mask to images
    im_rgba = im_rgb.copy()
    im_rgba.putalpha(pil_original_mask)
    im_dark.putalpha(pil_original_mask)
    
    return [cropped_signature_image, im_rgba, im_dark]

title = "Mysign.id - Signature Background removal based on DIS"
description = "ML Model based on ECCV2022/dis-background-removal specifically made for removing background from signatures."

interface = gr.Interface(
    fn=inference,
    inputs=gr.Image(type='filepath'),
    outputs=["image", "image", "image"],
    examples=[['example-1.jpg'], ['example-2.jpg']],
    title=title,
    description=description,
    allow_flagging='never',
    cache_examples=False,
    ).queue(api_open=True).launch(show_api=True, show_error=True)