Segmented input area

vision
1

Hello all,
I have my input image segmented into water region and land region.
I have my masked image:
image = Image.open(image_path).convert(“RGB”)
mask = Image.open(mask_path) .convert(“L”)

Convert to numpy arrays

image_np = np.array(image)
mask_np = np.array(mask)
binary_mask = (mask_np > 0).astype(np.uint8)

Apply mask to the image

segmented_part = image_np * np.expand_dims(mask_np, axis=2)
segmented_part.astype(“float32”)
segmented_part_image = Image.fromarray(segmented_part)

I have a U-Net network and my input is the masked image the tensor size is (batch size, channels (I have 3), width, height) like below
plot

How can I work with the features that are only in the water region, just predict the water region and have the loss for that region?

Thank you for your help

2

Write your own Dataset class that works with three things: image, object_mask, water_mask (where water pixel =1, land_pixel=0)

Do augmentation (I use albumentations lib) like this:
sample = self.augmentation(image=image, masks=[object_mask, water_mask ])

Write your own loss (likely IoU + BCE) and metrics that takes three objects (pred, object_mask, water_mask), then do
pred = pred * water_mask
object_mask= object_mask * water_mask
and then do standard IoU_loss(pred, object_mask) + BCE(pred, object_mask) on these new entities.

ChatGPT writes decent code for this task.