app.py

import streamlit as st
import cv2
import numpy as np
import torch
from torchvision import transforms, models
from PIL import Image
from TranSalNet_Res import TranSalNet
import torch.nn as nn
from utils.data_process import preprocess_img, postprocess_img

device = torch.device('cpu')
model = TranSalNet()
model.load_state_dict(torch.load('pretrained_models/TranSalNet_Res.pth', map_location=torch.device('cpu')))
model.to(device)
model.eval()

def count_and_label_red_patches(heatmap, threshold=200):
    red_mask = heatmap[:, :, 2] > threshold
    contours, _ = cv2.findContours(red_mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    
    # Sort the contours based on their areas in descending order
    contours = sorted(contours, key=cv2.contourArea, reverse=True)
    
    original_image = np.array(image)
    
    # Find the centroid of the red spot with the highest area
    M_largest = cv2.moments(contours[0])
    if M_largest["m00"] != 0:
        cX_largest = int(M_largest["m10"] / M_largest["m00"])
        cY_largest = int(M_largest["m01"] / M_largest["m00"])
    else:
        cX_largest, cY_largest = 0, 0
    
    for i, contour in enumerate(contours, start=1):
        # Compute the centroid of the current contour
        M = cv2.moments(contour)
        if M["m00"] != 0:
            cX = int(M["m10"] / M["m00"])
            cY = int(M["m01"] / M["m00"])
        else:
            cX, cY = 0, 0
        
        radius = 20  # Adjust the circle radius to fit the numbers
        circle_color = (0, 0, 0)  # Blue color
        cv2.circle(original_image, (cX, cY), radius, circle_color, -1)  # Draw blue circle

        # Connect the current red spot to the red spot with the highest area
        line_color = (0, 0, 0)  # Red color
        cv2.line(original_image, (cX, cY), (cX_largest, cY_largest), line_color, 2)

        font = cv2.FONT_HERSHEY_SIMPLEX
        font_scale = 1
        font_color = (255, 255, 255)
        line_type = cv2.LINE_AA
        cv2.putText(original_image, str(i), (cX - 10, cY + 10), font, font_scale, font_color, 2, line_type)

    return original_image, len(contours)

st.title('Saliency Detection App')
st.write('Upload an image for saliency detection:')
uploaded_image = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])

if uploaded_image:
    image = Image.open(uploaded_image)
    st.image(image, caption='Uploaded Image', use_column_width=True)

    if st.button('Detect Saliency'):
        img = image.resize((384, 288))
        img = np.array(img)
        img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)  # Convert to BGR color space
        img = np.array(img) / 255.
        img = np.expand_dims(np.transpose(img, (2, 0, 1)), axis=0)
        img = torch.from_numpy(img)
        img = img.type(torch.FloatTensor).to(device)

        pred_saliency = model(img).squeeze().detach().numpy()

        heatmap = (pred_saliency * 255).astype(np.uint8)
        heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)  # Use a blue colormap (JET)

        heatmap = cv2.resize(heatmap, (image.width, image.height))

        enhanced_image = np.array(image)
        b, g, r = cv2.split(enhanced_image)
        clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
        b_enhanced = clahe.apply(b)
        enhanced_image = cv2.merge((b_enhanced, g, r))

        alpha = 0.7
        blended_img = cv2.addWeighted(enhanced_image, 1 - alpha, heatmap, alpha, 0)

        original_image, num_red_patches = count_and_label_red_patches(heatmap)

        st.image(original_image, caption=f'Image with {num_red_patches} Red Patches', use_column_width=True, channels='RGB')

        st.image(blended_img, caption='Blended Image', use_column_width=True, channels='BGR')

        # Create a dir with name example to save
        cv2.imwrite('example/result15.png', blended_img, [int(cv2.IMWRITE_JPEG_QUALITY), 200])
        st.success('Saliency detection complete. Result saved as "example/result15.png".')