Fix for cropped image

app.py

@@ -1,4 +1,7 @@
+from data_loader_cache import normalize, im_reader, im_preprocess
+from models import *
 import cv2
+from skimage.restoration import denoise_nl_means, estimate_sigma
 import gradio as gr
 import os
 from PIL import Image
@@ -16,45 +19,48 @@ 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
+# 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]):
+
+    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)
+    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])])
+transform = transforms.Compose(
+    [GOSNormalize([0.5, 0.5, 0.5], [1.0, 1.0, 1.0])])
+
 
 def load_image(im_path, hypar):
     im = im_reader(im_path)
     im, im_shp = im_preprocess(im, hypar["cache_size"])
-    im = torch.divide(im,255.0)
+    im = torch.divide(im, 255.0)
     shape = torch.from_numpy(np.array(im_shp))
-    return transform(im).unsqueeze(0), shape.unsqueeze(0) # make a batch of image, shape
+    # make a batch of image, shape
+    return transform(im).unsqueeze(0), shape.unsqueeze(0)
 
 
-def build_model(hypar,device):
-    net = hypar["model"]#GOSNETINC(3,1)
+def build_model(hypar, device):
+    net = hypar["model"]  # GOSNETINC(3,1)
 
     # convert to half precision
-    if(hypar["model_digit"]=="half"):
+    if (hypar["model_digit"] == "half"):
         net.half()
         for layer in net.modules():
             if isinstance(layer, nn.BatchNorm2d):
@@ -62,96 +68,145 @@ def build_model(hypar,device):
 
     net.to(device)
 
-    if(hypar["restore_model"]!=""):
-        net.load_state_dict(torch.load(hypar["model_path"]+"/"+hypar["restore_model"], map_location=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()  
+    net.eval()
     return net
 
 
-def crop_to_signature(mask):
-    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+def crop_to_signature(mask, padding=32):
+    """
+    Crop the signature area based on the mask and add padding.
+    :param mask: The binary mask of the signature.
+    :param padding: Padding around the cropped signature.
+    :return: Cropped mask of the signature with padding.
+    """
+    contours, _ = cv2.findContours(
+        mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
     if contours:
         # Assume the largest contour is the signature
         x, y, w, h = cv2.boundingRect(max(contours, key=cv2.contourArea))
-        # Add some padding to the bounding box
-        padding = 32
-        x, y, w, h = x-padding, y-padding, w+2*padding, h+2*padding
+
+        # Add padding to the bounding box
+        x = max(x - padding, 0)
+        y = max(y - padding, 0)
+        w = min(w + 2 * padding, mask.shape[1] - x)
+        h = min(h + 2 * padding, mask.shape[0] - y)
+
         # Crop the mask
         cropped_mask = mask[y:y+h, x:x+w]
-        return cropped_mask
+        return cropped_mask, (x, y, w, h)
     else:
-        return mask  # Return the original mask if no contours are found
+        # Return the original mask if no contours are found
+        return mask, (0, 0, mask.shape[1], mask.shape[0])
+
+
+def smooth_and_denoise(mask):
+    """
+    Apply smoothing and denoising to the mask.
+    :param mask: The binary mask of the signature.
+    :return: Processed mask.
+    """
+    # 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, multichannel=True))
+
+    # 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, multichannel=True)
+    return denoised_mask
+
 
-    
 def predict(net,  inputs_val, shapes_val, hypar, device):
     '''
     Given an Image, predict the mask
     '''
     net.eval()
 
-    if(hypar["model_digit"]=="full"):
+    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
+    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
+    # B x 1 x H x W    # we want the first one which is the most accurate prediction
+    pred_val = ds_val[0][0, :, :, :]
 
-    ## 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'))
+    # 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
+    pred_val = (pred_val-mi)/(ma-mi)  # max = 1
+
+    if device == 'cuda':
+        torch.cuda.empty_cache()
+    # it is the mask we need
+    return (pred_val.detach().cpu().numpy()*255).astype(np.uint8)
+
 
-    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 = {}  # 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
+hypar["model_path"] = "./saved_models"  # load trained weights from this path
+hypar["restore_model"] = "isnet.pth"  # name of the to-be-loaded weights
+# indicate if activate intermediate feature supervision
+hypar["interm_sup"] = False
 
-##  choose floating point accuracy --
-hypar["model_digit"] = "full" ## indicates "half" or "full" accuracy of float number
+# choose floating point accuracy --
+# indicates "half" or "full" accuracy of float number
+hypar["model_digit"] = "full"
 hypar["seed"] = 0
 
-hypar["cache_size"] = [1024, 1024] ## cached input spatial resolution, can be configured into different size
+# cached input spatial resolution, can be configured into different size
+hypar["cache_size"] = [1024, 1024]
 
-## 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
+# data augmentation parameters ---
+# mdoel input spatial size, usually use the same value hypar["cache_size"], which means we don't further resize the images
+hypar["input_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["crop_size"] = [1024, 1024]
 
 hypar["model"] = ISNetDIS()
 
- # Build Model
+# Build Model
 net = build_model(hypar, device)
 
 
 def inference(image):
-  image_path = image
-  
-  image_tensor, orig_size = load_image(image_path, hypar) 
-  mask = predict(net, image_tensor, orig_size, hypar, device)
-  
-  pil_mask = Image.fromarray(mask).convert('L')
-  im_rgb = Image.open(image).convert("RGB")
-  im_dark = Image.new('RGB', im_rgb.size, (0, 0, 0))
-  
-  
-  im_rgba = im_rgb.copy()
-  im_rgba.putalpha(pil_mask)
-  im_dark.putalpha(pil_mask)
-
-  return [im_rgba, pil_mask, im_dark]
+    image_path = image
+
+    image_tensor, orig_size = load_image(image_path, hypar)
+    mask = predict(net, image_tensor, orig_size, hypar, device)
+
+    cropped_mask, (x, y, w, h) = crop_to_signature(mask)
+    processed_mask = smooth_and_denoise(cropped_mask)
+
+    pil_mask = Image.fromarray(processed_mask).convert('L')
+    im_rgb = Image.open(image).convert("RGB")
+
+    # Crop the original image to match the mask
+    cropped_image = im_rgb.crop((x, y, x + w, y + h))
+
+    # Create a dark background image of the same size as the cropped image
+    im_dark = Image.new('RGB', (w, h), (0, 0, 0))
+
+    # Apply the mask to the cropped image
+    im_rgba = cropped_image.copy()
+    im_rgba.putalpha(pil_mask)
+    im_dark.putalpha(pil_mask)
+
+    return [im_rgba, pil_mask, im_dark]
 
 
 title = "Mysign.id - Signature Background removal based on DIS"
@@ -166,4 +221,4 @@ interface = gr.Interface(
     description=description,
     allow_flagging='never',
     cache_examples=False,
-    ).queue(api_open=True).launch(show_api=True, show_error=True)
+).queue(api_open=True).launch(show_api=True, show_error=True)