app.py

import gradio as gr
import tensorflow as tf
import numpy as np
import cv2
from keras.utils import normalize
from PIL import Image

def dice_coef(y_true, y_pred):
    smooth = 1e-5
    intersection = K.sum(y_true * y_pred, axis=[1, 2, 3])
    union = K.sum(y_true, axis=[1, 2, 3]) + K.sum(y_pred, axis=[1, 2, 3])
    return K.mean((2.0 * intersection + smooth) / (union + smooth), axis=0)

def predict_segmentation(image):
    original_size = (image.shape[1], image.shape[0])  # (width, height)
    
    # Resize to the model's input size
    SIZE_X = 128
    SIZE_Y = 128
    img = cv2.resize(image, (SIZE_Y, SIZE_X))

    if len(img.shape) == 3 and img.shape[2] == 3:  # If the image is RGB
        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)  # Convert to grayscale

    img = np.expand_dims(img, axis=2)  # Add the channel dimension
    img = normalize(img, axis=1)
    X_test = np.expand_dims(img, axis=0)  # Add the batch dimension

    custom_objects = {'dice_coef': dice_coef}
    with tf.keras.utils.custom_object_scope(custom_objects):
        model = tf.keras.models.load_model("model_best.h5")

    # Get the prediction
    prediction = model.predict(X_test)
    predicted_img = np.argmax(prediction, axis=3)[0, :, :]

    # Resize prediction back to original image size
    predicted_img_resized = cv2.resize(predicted_img, original_size, interpolation=cv2.INTER_NEAREST)

    # Create an RGBA image with a transparent background
    rgba_img = np.zeros((predicted_img_resized.shape[0], predicted_img_resized.shape[1], 4), dtype=np.uint8)

    # Define the color for the segmented area (e.g., red)
    segmented_color = [255, 0, 0]  # Red color in RGB

    # Set the segmented area to the desired color
    for i in range(3):
        rgba_img[:, :, i] = np.where(predicted_img_resized > 0, segmented_color[i], 0)

    # Create an alpha channel: 255 where there is segmentation, 0 otherwise
    rgba_img[:, :, 3] = np.where(predicted_img_resized > 0, 255, 0)

    # Convert the numpy array to an image
    output_image = Image.fromarray(rgba_img)

    # Save the image as PNG to return it
    output_image_path = "/tmp/segmented_output.png"
    output_image.save(output_image_path)

    return output_image_path

# Gradio Interface
iface = gr.Interface(
    fn=predict_segmentation,
    inputs="image",
    outputs="file",  # Return the file path to download the PNG
    live=False
)

iface.launch()