app.py

import gradio as gr
from transformers import AutoImageProcessor, SiglipForImageClassification
from PIL import Image
import torch
import json
import base64
import io
import numpy as np

model_name = "prithivMLmods/Rice-Leaf-Disease"
model = SiglipForImageClassification.from_pretrained(model_name)
processor = AutoImageProcessor.from_pretrained(model_name)
model.eval()

labels = {0: "Bacterial Leaf Blight", 1: "Rice Blast", 2: "Brown Spot", 3: "Healthy", 4: "Rice Tungro Virus"}

def is_likely_rice_leaf(image):
    """Check if image is obviously NOT a plant (laptop, face, etc).
    PERMISSIVE: only rejects clearly non-plant images.
    Diseased leaves can be brown/yellow/dry — we must accept those."""
    img = image.resize((64, 64))
    arr = np.array(img).astype(float)
    r, g, b = arr[:,:,0], arr[:,:,1], arr[:,:,2]
    total_pixels = r.size

    # Green pixels: green channel dominates (healthy leaves)
    green_mask = (g > r + 5) & (g > b + 5) & (g > 30)
    green_ratio = np.sum(green_mask) / total_pixels

    # Brown/tan pixels: diseased or dry leaves
    brown_mask = (r > b + 10) & (r > 40) & (r < 220) & (g > 30)
    brown_ratio = np.sum(brown_mask) / total_pixels

    # Yellow/orange: tungro, nutrient deficiency
    yellow_mask = (r > 80) & (g > 60) & (b < g) & (np.abs(r.astype(int) - g.astype(int)) < 60)
    yellow_ratio = np.sum(yellow_mask) / total_pixels

    # Any natural/organic color
    natural_ratio = green_ratio + brown_ratio * 0.8 + yellow_ratio * 0.6

    # Pure gray/metallic: electronics, concrete (all channels nearly equal AND not bright white)
    gray_mask = (np.abs(r.astype(int) - g.astype(int)) < 12) & (np.abs(g.astype(int) - b.astype(int)) < 12) & (r < 200)
    gray_ratio = np.sum(gray_mask) / total_pixels

    # Pure black: screens, dark objects
    black_mask = (r < 30) & (g < 30) & (b < 30)
    black_ratio = np.sum(black_mask) / total_pixels

    # Only reject if image is CLEARLY not a plant:
    # - More than 60% gray/metallic AND less than 8% natural colors
    # - Or more than 50% pure black (screens, dark electronics)
    is_not_plant = (gray_ratio > 0.6 and natural_ratio < 0.08) or (black_ratio > 0.5)

    return not is_not_plant, round(natural_ratio * 100, 1), round(gray_ratio * 100, 1)

def classify(image):
    image = Image.fromarray(image).convert("RGB")

    # Step 1: Color validation (only rejects obvious non-plants)
    is_leaf, plant_pct, gray_pct = is_likely_rice_leaf(image)

    # Step 2: Run disease model
    inputs = processor(images=image, return_tensors="pt")
    with torch.no_grad():
        outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=1).squeeze().tolist()

    results = []
    for i in range(len(probs)):
        results.append({"label": labels[i], "score": round(probs[i], 4)})
    results.sort(key=lambda x: x["score"], reverse=True)
    top = results[0]

    return json.dumps({
        "disease": top["label"],
        "confidence": round(top["score"] * 100, 1),
        "probabilities": results,
        "is_valid_rice_leaf": is_leaf,
        "validation_message": None if is_leaf else f"This does not appear to be a rice leaf (natural colors: {plant_pct}%, gray: {gray_pct}%). Please scan a rice leaf.",
        "plant_pixel_ratio": plant_pct,
        "gray_pixel_ratio": gray_pct
    })

def detect_base64(img_b64):
    try:
        if ',' in img_b64:
            img_b64 = img_b64.split(',')[1]
        img_bytes = base64.b64decode(img_b64)
        image = Image.open(io.BytesIO(img_bytes)).convert("RGB")
        return classify(np.array(image))
    except Exception as e:
        return json.dumps({"error": str(e)})

demo = gr.TabbedInterface(
    [
        gr.Interface(fn=classify, inputs=gr.Image(), outputs=gr.Textbox(label="Result"), title="RiceGuard Disease Detection"),
        gr.Interface(fn=detect_base64, inputs=gr.Textbox(label="Base64 Image"), outputs=gr.Textbox(label="Result"), title="API")
    ],
    ["Upload Image", "Base64 API"]
)

demo.launch()