# -*- coding: utf-8 -*-
import gradio as gr
import cv2
import numpy as np
import torch
from PIL import Image
from ultralytics import YOLO
from transformers import BlipProcessor, BlipForConditionalGeneration
from huggingface_hub import hf_hub_download
import warnings
warnings.filterwarnings('ignore')
# ============================================================
# 1. LOAD MODELS (CHANGE USERNAME BELOW)
# ============================================================
print("Loading YOLOv11 model from Hugging Face Hub...")
# ⚠️ IMPORTANT: Replace "YOUR_USERNAME" with your actual Hugging Face username
model_path = hf_hub_download(
repo_id="SkyGuardAI/drone-detection-yolov11", # <--- CHANGE THIS
filename="best.pt"
)
model = YOLO(model_path)
print("YOLO model loaded successfully.")
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
print("Loading BLIP model...")
blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(device)
blip_model.eval()
print("BLIP model loaded successfully.")
# ============================================================
# 2. HEATMAP FUNCTIONS (unchanged)
# ============================================================
layer_outputs = {}
def hook_fn(module, input, output):
layer_outputs['feature_map'] = output.detach()
def get_best_layer(model):
best_layer = None
best_depth = 0
pytorch_model = model.model if hasattr(model, 'model') else model
for name, layer in pytorch_model.named_modules():
if isinstance(layer, torch.nn.Conv2d):
depth = name.count('.')
if depth > best_depth:
best_depth = depth
best_layer = layer
return best_layer
def generate_heatmap(model, image):
try:
layer_outputs.clear()
if isinstance(image, Image.Image):
image = np.array(image)
img_resized = cv2.resize(image, (640, 640))
pytorch_model = model.model if hasattr(model, 'model') else model
target_layer = get_best_layer(pytorch_model)
if target_layer is None:
return None
hook = target_layer.register_forward_hook(hook_fn)
results = model(img_resized)
hook.remove()
if 'feature_map' not in layer_outputs:
return None
feature_map = layer_outputs['feature_map']
if feature_map.dim() == 4:
heatmap = feature_map[0].mean(dim=0).cpu().numpy()
else:
heatmap = feature_map.cpu().numpy()
heatmap = cv2.GaussianBlur(heatmap, (5, 5), 0)
min_val = heatmap.min()
max_val = heatmap.max()
if max_val - min_val > 1e-8:
heatmap = (heatmap - min_val) / (max_val - min_val)
else:
heatmap = np.zeros_like(heatmap)
heatmap = cv2.resize(heatmap, (640, 640))
threshold = np.percentile(heatmap, 70)
heatmap = np.where(heatmap > threshold, heatmap, 0)
if heatmap.max() > 0:
heatmap = heatmap / heatmap.max()
heatmap_colored = cv2.applyColorMap((heatmap * 255).astype(np.uint8), cv2.COLORMAP_JET)
heatmap_colored = cv2.cvtColor(heatmap_colored, cv2.COLOR_BGR2RGB)
overlay = cv2.addWeighted(img_resized, 0.6, heatmap_colored, 0.4, 0)
return overlay
except Exception as e:
print(f"Heatmap error: {e}")
return None
# ============================================================
# 3. DYNAMIC CAPTION (BLIP)
# ============================================================
def generate_dynamic_caption(image):
try:
if isinstance(image, np.ndarray):
image = Image.fromarray(image).convert('RGB')
elif isinstance(image, Image.Image):
image = image.convert('RGB')
inputs = blip_processor(image, return_tensors="pt").to(device)
with torch.no_grad():
out = blip_model.generate(
**inputs,
max_length=60,
num_beams=5,
temperature=0.7,
repetition_penalty=1.2
)
caption = blip_processor.decode(out[0], skip_special_tokens=True)
return caption
except Exception as e:
print(f"Caption error: {e}")
return "AI model is analyzing the scene."
# ============================================================
# 4. XAI REPORT
# ============================================================
def build_xai_report(is_drone, confidence, drone_count, processing_time, image_caption):
if is_drone:
drone_text = "a drone" if drone_count == 1 else f"{drone_count} drones"
if confidence >= 0.8:
confidence_level = "VERY HIGH"
confidence_assessment = "excellent"
elif confidence >= 0.6:
confidence_level = "HIGH"
confidence_assessment = "good"
elif confidence >= 0.5:
confidence_level = "MODERATE"
confidence_assessment = "acceptable"
else:
confidence_level = "LOW"
confidence_assessment = "limited"
report = f"""
================================================================================
XAI DRONE DETECTION REPORT
================================================================================
[DYNAMIC IMAGE ANALYSIS]
{image_caption}
[DETECTION RESULTS]
• Status: CONFIRMED
• Confidence: {confidence:.1%} (Level: {confidence_level})
• Drones Detected: {drone_count}
• Processing Time: {processing_time:.0f} milliseconds
• Model: YOLOv11
• XAI Method: Convolutional Feature Map Extraction
[XAI HEATMAP INTERPRETATION]
The heatmap shows red regions where the neural network focused its attention.
Strong red activation on {drone_text} confirms the model successfully learned
discriminative features for drone detection.
The model demonstrates {confidence_assessment} confidence, as evidenced by
the concentrated activation pattern in the heatmap.
[TECHNICAL DETAILS]
• Heatmap: Extracted from deepest convolutional layer of YOLOv11
• Color Code: Red = High activation (model focus) | Blue = Low activation
• Processing: Gaussian blur (5x5 kernel)
• Threshold: Top 30% activation retained
[XAI CONCLUSION]
The YOLOv11 model has successfully detected {drone_text} in this image with
{confidence_assessment} confidence. The heatmap confirms correct feature
learning as the neural network focused on the drone's location.
"""
else:
report = f"""
================================================================================
XAI DRONE DETECTION REPORT
================================================================================
[DYNAMIC IMAGE ANALYSIS]
{image_caption}
[DETECTION RESULTS]
• Status: NOT CONFIRMED
• Highest Confidence: {confidence:.1%}
• Processing Time: {processing_time:.0f} milliseconds
• Model: YOLOv11
• XAI Method: Convolutional Feature Map Extraction
[XAI HEATMAP INTERPRETATION]
The heatmap shows scattered or unfocused activation patterns without strong
concentration on any specific region. This indicates the model did not identify
strong drone-like features in this image.
[POSSIBLE REASONS]
• No drone is present in the image
• Drone is too small or too far from the camera
• Poor lighting conditions reduce feature visibility
• Image blur or motion blur affects detection quality
• Drone is partially occluded by other objects
[RECOMMENDATIONS]
• Ensure adequate lighting in the scene
• Position the drone closer to the camera
• Use higher resolution images without motion blur
• Avoid cluttered backgrounds that may confuse the model
[TECHNICAL NOTE]
The heatmap was extracted from the deepest convolutional layer of YOLOv11.
Scattered activation pattern confirms absence of strong drone-like features.
"""
return report
# ============================================================
# 5. MAIN PIPELINE FUNCTION
# ============================================================
def drone_detection_pipeline(input_image):
try:
if isinstance(input_image, Image.Image):
img = np.array(input_image)
else:
img = input_image.copy()
original_h, original_w = img.shape[:2]
results = model(img)
heatmap_overlay = generate_heatmap(model, img)
is_drone = False
confidence = 0.0
drone_count = 0
if results and len(results) > 0 and hasattr(results[0], 'boxes'):
boxes_data = results[0].boxes
if boxes_data and boxes_data.data is not None:
data = boxes_data.data.cpu().numpy()
for det in data:
if len(det) >= 6:
conf = float(det[4])
cls = int(det[5])
class_name = model.names[cls]
if class_name.lower() == 'drone' and conf >= 0.3:
drone_count += 1
confidence = max(confidence, conf)
is_drone = drone_count > 0
result_img = results[0].plot() if len(results) > 0 else img
result_img_rgb = cv2.cvtColor(result_img, cv2.COLOR_BGR2RGB)
caption = generate_dynamic_caption(img)
processing_time_ms = 0
report = build_xai_report(is_drone, confidence, drone_count, processing_time_ms, caption)
if heatmap_overlay is not None:
heatmap_resized = cv2.resize(heatmap_overlay, (original_w, original_h))
else:
heatmap_resized = np.zeros_like(result_img_rgb)
return result_img_rgb, heatmap_resized, report
except Exception as e:
error_msg = f"An error occurred during processing: {str(e)}"
print(error_msg)
blank = np.zeros((480, 640, 3), dtype=np.uint8)
return blank, blank, error_msg
# ============================================================
# 6. INTERFACE - DARK BLUE ONLY, NO PURPLE/VIOLET
# ============================================================
custom_css = """
/* Force white background */
.gradio-container, body, .gradio-container .main, .gradio-container .gradio-container {
background: #ffffff !important;
font-family: 'Segoe UI', 'Roboto', sans-serif;
}
/* Override any purple/violet colors from Gradio default theme */
:root {
--primary-50: #0a2b4e !important;
--primary-100: #1e4a76 !important;
--primary-200: #2c5282 !important;
--primary-300: #1e3a8a !important;
--primary-400: #0a2b4e !important;
--primary-500: #0a2b4e !important;
--primary-600: #0a2b4e !important;
--primary-700: #0a2b4e !important;
--primary-800: #0a2b4e !important;
--primary-900: #0a2b4e !important;
}
/* Header - dark blue */
.skyguard-header {
text-align: center;
padding: 1.5rem 0 0.5rem 0;
margin-bottom: 1.5rem;
}
.skyguard-header h1 {
font-size: 3rem;
font-weight: 800;
letter-spacing: 3px;
color: #0a2b4e !important;
margin-bottom: 0.2rem;
}
.skyguard-header .tagline {
font-size: 1.2rem;
font-weight: 600;
color: #1e4a76 !important;
letter-spacing: 1px;
margin-top: 0.2rem;
}
.skyguard-header .sub {
font-size: 0.9rem;
color: #2c5282 !important;
margin-top: 0.5rem;
}
/* Primary button - solid dark blue, no gradient, no purple */
.gr-button-primary, button.gr-button-primary, .gr-button.primary {
background: #0a2b4e !important;
border: none !important;
color: white !important;
font-weight: 700 !important;
border-radius: 30px !important;
padding: 10px 28px !important;
transition: all 0.2s ease !important;
text-transform: uppercase !important;
letter-spacing: 1px;
box-shadow: none !important;
}
.gr-button-primary:hover, button.gr-button-primary:hover, .gr-button.primary:hover {
background: #1e4a76 !important;
transform: scale(1.02);
box-shadow: 0 2px 8px rgba(10,43,78,0.2) !important;
}
/* Tabs - remove purple */
.gr-tabs .tab-nav button {
background-color: #eef2f5 !important;
color: #0a2b4e !important;
font-weight: 600 !important;
border-radius: 8px 8px 0 0 !important;
padding: 8px 20px !important;
margin-right: 4px;
border: none !important;
}
.gr-tabs .tab-nav button.selected {
background-color: #0a2b4e !important;
color: white !important;
}
/* Input and output boxes */
.gr-box, .gr-form, .gr-input, .gr-panel {
background-color: #f9fafb !important;
border: 1px solid #d1d5db !important;
border-radius: 12px !important;
}
/* Labels */
label, .gr-form label {
color: #1e293b !important;
font-weight: 500 !important;
}
/* Markdown text */
.gr-markdown p, .gr-markdown {
color: #1e293b !important;
}
/* Footer */
.skyguard-footer {
text-align: center;
margin-top: 30px;
font-size: 12px;
color: #6c757d;
border-top: 1px solid #e2e8f0;
padding-top: 15px;
}
/* Additional overrides for any purple elements */
.gr-button, .gr-button-lg, .gr-button-sm, button {
--tw-ring-color: #0a2b4e !important;
}
.prose a, a {
color: #1e4a76 !important;
}
input:focus, textarea:focus, select:focus {
border-color: #0a2b4e !important;
ring-color: #0a2b4e !important;
}
"""
with gr.Blocks(title="SkyGuard - Drone Detection System", theme=gr.themes.Soft(), css=custom_css) as demo:
gr.HTML("""
""")
with gr.Row(equal_height=False):
with gr.Column(scale=1, min_width=300):
input_image = gr.Image(label="Upload Aerial Image", type="pil")
analyze_btn = gr.Button("Analyze Threat", variant="primary")
with gr.Column(scale=2):
with gr.Tabs():
with gr.TabItem("Detection"):
output_image = gr.Image(label="Bounding Boxes")
with gr.TabItem("Heatmap (XAI)"):
heatmap_image = gr.Image(label="Neural Activation Heatmap")
with gr.TabItem("Intelligence Report"):
report_text = gr.Markdown(label="XAI Analysis Report")
analyze_btn.click(
fn=drone_detection_pipeline,
inputs=[input_image],
outputs=[output_image, heatmap_image, report_text]
)
gr.HTML("""
""")
# ============================================================
# 7. RUN APP
# ============================================================
if __name__ == "__main__":
demo.launch()