""" Streamlit Web UI for the Diatom Classifier Pipeline. """ import os import streamlit as st import pandas as pd import numpy as np from pathlib import Path from PIL import Image, ImageDraw, ImageFont from dotenv import load_dotenv # Load environment variables from .env file (if it exists) load_dotenv() # -- Setup Environment & Paths -- APP_ENV = os.environ.get("APP_ENV", "production") REPO_ID = "kemalbsoylu/diatom-models" BASE_DIR = Path(__file__).resolve().parent.parent st.set_page_config(page_title="Diatom AI", page_icon="🔬", layout="wide") @st.cache_resource def load_models(): from ultralytics import YOLO from fastai.vision.all import load_learner if APP_ENV == "development": # LOCAL: Use models stored on your hard drive yolo_path = BASE_DIR / "models" / "yolo_diatom_detector.pt" resnet_path = BASE_DIR / "models" / "v2_resnet18_weighted.pkl" else: # PRODUCTION: Download models from the Hugging Face Hub from huggingface_hub import hf_hub_download yolo_path = hf_hub_download(repo_id=REPO_ID, filename="yolo_diatom_detector.pt") resnet_path = hf_hub_download(repo_id=REPO_ID, filename="v2_resnet18_weighted.pkl") # Load them into memory yolo = YOLO(yolo_path) resnet = load_learner(resnet_path) return yolo, resnet st.title("🔬 Diatom Detection & Classification AI") st.markdown(""" Upload a microscope image. Use **Full Slide Analysis** to automatically detect and classify multiple diatoms, or use **Single Diatom Crop** if you already have a cropped image of a single diatom. """) with st.spinner("Loading AI Models into memory..."): yolo_model, resnet_model = load_models() if APP_ENV == "development": st.sidebar.success("🔧 Running in Development Mode (Local Models)") # -- Sidebar Controls -- st.sidebar.header("Configuration") app_mode = st.sidebar.radio("Select Analysis Mode:", ["Full Slide Analysis", "Single Diatom Crop"]) conf_threshold = 0.25 if app_mode == "Full Slide Analysis": conf_threshold = st.sidebar.slider("Detection Confidence", 0.1, 1.0, 0.25, 0.05) st.sidebar.markdown("*Lowering the threshold finds more diatoms but increases false positives. (Default: 0.25)*") # -- Sidebar Footer (Portfolio & License) -- st.sidebar.markdown("---") st.sidebar.markdown("### About") st.sidebar.markdown("Developed by **Kemal Soylu**") st.sidebar.markdown("[View Source Code on GitHub](https://github.com/kemalbsoylu/diatom-classifier)") st.sidebar.markdown(""" Licenses: Code (MIT), Detector (AGPL-3.0).
Data: Trained on dataset by Gündüz et al. (CC BY-NC-SA 4.0). """, unsafe_allow_html=True) # -- Main File Uploader -- uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"]) if uploaded_file is not None: # Silence PyCharm type warning assert not isinstance(uploaded_file, list) original_image = Image.open(uploaded_file).convert("RGB") st.markdown("---") st.subheader("Analysis Results") # ----------------------------------------------------- # MODE 1: SINGLE CROP CLASSIFICATION # ----------------------------------------------------- if app_mode == "Single Diatom Crop": with st.spinner("Classifying diatom..."): # Convert standard PIL Image directly to a numpy array img_array = np.array(original_image) # Predict directly on the array pred_class, pred_idx, probs = resnet_model.predict(img_array) conf = probs[pred_idx].item() * 100 st.success("Classification Complete!") st.metric(label="Predicted Genus", value=f"**{pred_class}**", delta=f"{conf:.2f}% Confidence", border=True) st.info("Note: This mode bypassed the automatic detector and evaluated the entire image as a single diatom. For best results, ensure your image is cropped tightly around the diatom with a maximum of 15% background margin.") st.markdown("---") st.subheader("Image Viewer") st.image(original_image, caption="Original Upload", use_container_width=False) # ----------------------------------------------------- # MODE 2: FULL SLIDE YOLO + RESNET # ----------------------------------------------------- else: display_image = original_image.copy() report_data = [] diatom_count = 0 with st.spinner("Scanning slide & Classifying..."): results = yolo_model(original_image, conf=conf_threshold, verbose=False)[0] draw = ImageDraw.Draw(display_image) try: font = ImageFont.truetype("arial.ttf", 24) except IOError: font = ImageFont.load_default() for box in results.boxes.xyxy: diatom_count += 1 x1, y1, x2, y2 = map(int, box.tolist()) # Apply 15% margin for cropping box_w, box_h = x2 - x1, y2 - y1 margin_x, margin_y = int(box_w * 0.15), int(box_h * 0.15) crop_x1 = max(0, x1 - margin_x) crop_y1 = max(0, y1 - margin_y) crop_x2 = min(original_image.width, x2 + margin_x) crop_y2 = min(original_image.height, y2 + margin_y) cropped_img = original_image.crop((crop_x1, crop_y1, crop_x2, crop_y2)) # Convert cropped PIL Image to a numpy array img_array = np.array(cropped_img) # Classify directly with ResNet pred_class, pred_idx, probs = resnet_model.predict(img_array) conf = probs[pred_idx].item() * 100 # Draw bounding box on display image draw.rectangle([x1, y1, x2, y2], outline="red", width=3) # Format text with the ID matching the CSV report label_text = f"#{diatom_count} {pred_class} ({conf:.1f}%)" # Calculate text background size for readability left, top, right, bottom = font.getbbox(label_text) text_width = right - left text_height = bottom - top # Draw solid red background for text (positioned inside top-left of the box) draw.rectangle([x1, y1, x1 + text_width + 6, y1 + text_height + 6], fill="red") # Draw white text over the red background draw.text((x1 + 3, y1 + 3), label_text, fill="white", font=font, stroke_width=0.5, stroke_fill="white") # Save to report report_data.append({ "ID": diatom_count, "Genus": pred_class, "Confidence": f"{conf:.2f}%" }) # Render Full Slide Results if report_data: word = "diatom" if diatom_count == 1 else "diatoms" st.success(f"Successfully found {diatom_count} {word}!") df = pd.DataFrame(report_data) st.dataframe(df, use_container_width=True) csv = df.to_csv(index=False).encode('utf-8') st.download_button( label="📥 Download CSV Report", data=csv, file_name=f"analysis_{uploaded_file.name}.csv", mime="text/csv", ) else: st.warning("No diatoms detected. Try lowering the detection confidence threshold in the sidebar.") st.markdown("---") st.subheader("Image Viewer") col1, col2 = st.columns(2) with col1: st.image(original_image, caption="Original Upload", use_container_width=True) with col2: st.image(display_image, caption="Analyzed Image", use_container_width=True) # -- Footer AI Warning -- st.markdown("---") st.caption("**Disclaimer:** This application utilizes artificial intelligence and may produce inaccurate results. Always verify critical findings with a qualified domain expert.")