pepperumo/MVTec_Website
3
1import streamlit as st2import pandas as pd3import torch4 5import plotly.express as px6import plotly.graph_objects as go7import numpy as np8from PIL import Image9import os10import pickle11import joblib12import cv213from data_processing import (14 dataset_statistics, dataset_distribution_chart, load_dataset, 15 plot_bgr_pixel_densities, plot_pair_plots16)17from metrics_calculation import (18 load_evaluation_metrics,19 plot_roc_curve,20 plot_confusion_matrix21 )22 23from prediction import (24 run_inference_autoencoder, load_model_autoencoder, run_inference_knn, load_model_knn25)26 27 28# Overview Page29def overview_page():30 col1, col2, col3 = st.columns([1, 6, 1])31 32 st.title("🚀 Beyond Normal: Unveiling Image Anomalies with AI")33 st.markdown("---")34 35 try:36 st.image("images/overview_image.png", use_container_width=True, 37 caption="Normal vs. Anomalous Samples with Segmentation Masks")38 except FileNotFoundError:39 st.error("⚠️ Image file not found. Please check if 'images/overview_image.png' exists.")40 41 # Create three columns for key metrics42 col1, col2, col3 = st.columns(3)43 with col1:44 st.markdown("""45 <div>46 <h2>5,450</h2>47 <p>High-Resolution Images</p>48 </div>49 """, unsafe_allow_html=True)50 with col2:51 st.markdown("""52 <div>53 <h2>15</h2>54 <p>Object Categories</p>55 </div>56 """, unsafe_allow_html=True)57 with col3:58 st.markdown("""59 <div>60 <h2>70+</h2>61 <p>Defect Types</p>62 </div>63 """, unsafe_allow_html=True)64 65 st.markdown("---")66 67 68 st.header("🔍 What is Anomaly Detection?")69 st.write("""70 Imagine a world where machines can **spot defects** in products just like human inspectors—but **faster and with higher accuracy**! 71 This is exactly what **MVTec AD**, a powerful dataset, helps us achieve. It is designed for **automated quality control** in manufacturing by detecting **flaws** such as scratches, dents, and missing parts in different objects and textures.72 73 - 🖼️ **5,450 high-resolution images** across **15 object and texture categories**74 - ✅ **Training set**: Only contains **defect-free** images 75 - 🐞 **Test set**: Includes images with **over 70 different types of defects** 76 - 🎯 **Goal**: Automatically **detect and highlight anomalies** with **pixel-precise segmentation**77 """)78 79 st.subheader("🧐 **How Does Anomaly Detection Work?**")80 st.write("""81 The AI model learns what a **perfect product** looks like by studying thousands of **defect-free images**. 82 When it sees a **new image**, it checks:83 84 1️⃣ **Does this image match what I’ve seen before?** 85 2️⃣ **If not, where is the defect?** 86 87 The result? A **heatmap** showing the suspicious areas, along with a **segmentation mask** to pinpoint the defect.88 """)89 90 st.subheader("✨ **Bringing Anomalies to Light**: Real-World Examples")91 st.write("""92 Below are **three real examples** of AI-powered anomaly detection. 93 """)94 95 # Display anomaly detection images directly without additional data processing96 st.image("images/anomaly_visual_example_1.png", use_container_width=True, caption="Defective Wood - Liquid Stain")97 st.image("images/anomaly_visual_example_2.png", use_container_width=True, caption="Hazelnut - Hole Defect")98 st.image("images/anomaly_visual_example_3.png", use_container_width=True, caption="Leather - Cut Defect")99 100 st.write("""101 **First Column**: 102 - Original object with an anomaly 103 104 **The Heatmap (Second Column)**: 105 - AI scans the object and **highlights unusual areas** in **red/yellow**, indicating anomaly. 106 107 **Segmentation Map (Third Column)**: 108 - Shows the **exact shape** of the detected anomaly, crucial for precise localization.109 110 **Ground Truth (Fourth Column)**: 111 - The manually labeled anomaly **used for AI validation**.112 113 This technology helps manufacturers **automate anomaly detection, reduce waste, and ensure top-tier product quality** at an industrial scale. 🚀 114 """)115 116 117# Dataset Page118def dataset_page():119 st.title("📊 Dataset Analysis & Exploratory Data")120 121 tab1, tab2 = st.tabs(["📈 Dataset Overview", "🎨 Feature Analysis"])122 123 with tab1:124 st.subheader("Dataset Structure")125 complete_df = load_dataset()126 if complete_df is not None:127 st.markdown("""128 <div>129 <ul>130 <li>Total Samples: {}</li>131 <li>Categories: {}</li>132 </ul>133 </div>134 """.format(135 len(complete_df),136 len(complete_df['category'].unique())137 ), unsafe_allow_html=True)138 139 with st.expander("🔍 View Full Dataset"):140 st.dataframe(complete_df, use_container_width=True)141 else:142 st.error("❌ Error: Dataset could not be loaded.")143 144 145 st.subheader("Statistical Analysis")146 df = dataset_statistics()147 if df is not None:148 149 dataset_distribution_chart(df)150 151 with st.expander("📊 Detailed Statistics"):152 st.dataframe(df, use_container_width=True)153 else:154 st.error("❌ Error: Statistics could not be computed.")155 156 with tab2:157 st.subheader("Feature Engineering & Analysis")158 159 st.markdown("""160 <div>161 <h3>🔄 Dimensionality Reduction (PCA)</h3>162 <p>Our feature extraction pipeline includes:</p>163 <ul>164 <li>Image preprocessing and normalization</li>165 <li>Feature extraction using ResNet50</li>166 <li>PCA transformation preserving 95% variance</li>167 </ul>168 </div>169 """, unsafe_allow_html=True)170 171 complete_df = load_dataset()172 if complete_df is not None:173 selected_category = st.selectbox("Select a Category", 174 complete_df['category'].unique(), index=list(complete_df['category'].unique()).index('wood'))175 176 col1, col2 = st.columns(2)177 with col1:178 plot_bgr_pixel_densities(179 complete_df[complete_df['category'] == selected_category],180 pixel_columns=['num_pixels_b', 'num_pixels_g', 'num_pixels_r']181 )182 with col2:183 plot_pair_plots(complete_df[complete_df['category'] == selected_category])184 st.write("""185 **Key Takeaways from Feature Relationships:**186 187 **Strong Correlations:** 188 - The BGR pixel values exhibit **strong linear relationships**, which is expected as they represent color intensity.189 - **Perceived brightness** also shows a linear trend, confirming its dependence on RGB values.190 191 **Separation of Normal vs. Anomalous Data:** 192 - Some categories, like `Hazelnut` and `Tile`, show **clear separation** between normal (blue) and anomalous (red) points, indicating that anomalies have **distinct feature distributions**.193 - Other categories, such as `Screw` and `Transistor`, show **more overlap**, meaning their anomalies are harder to distinguish based only on pixel values.194 195 **Density Distributions:** 196 - Categories like `Carpet` and `Capsule` show **multi-modal distributions**, meaning that anomalies have **different types of defects**.197 - Categories like `Leather` and `Metal Nut` show anomalies with **different brightness levels**, suggesting that brightness-based anomaly detection could be effective.198 199 Overall, this analysis helps us understand which **features are useful for distinguishing anomalies** and which categories might need **additional feature engineering**.200 """)201 202 st.success("✅ Using PCA, BGR pixel distributions, and feature relationships, we ensure that the dataset is **optimized for training accurate anomaly detection models**.")203 204 205 206def synthetic_data_page():207 st.title("🔬 Data Enhancement Techniques")208 209 tab1, tab2 = st.tabs(["🧪 Synthetic Data", "🔄 Data Augmentation"])210 211 with tab1:212 st.header("Synthetic Data Generation")213 214 st.info("""215 🔍 **Synthetic Data** refers to artificially generated images that simulate anomalies 216 by modifying normal samples. Unlike data augmentation, synthetic data aims to create 217 new, realistic defect patterns that weren't present in the original dataset.218 """)219 st.write("""220 #### 🔍 Why Do We Need Synthetic Data?221 The **MVTec Anomaly Detection Dataset** contains **15 object categories**, but for each category:222 223 - The dataset is relatively **small**.224 - There are **far fewer anomaly images** than normal images.225 - Splitting test data for validation would leave even **less data** to train the model.226 227 To solve this problem, **we create additional "fake" anomaly images** to train the model better. 228 Instead of taking images of real defective objects (which are limited), we **manipulate normal images** by adding synthetic defects, such as **twisting, distorting, or overlaying textures**.229 230 """)231 st.image(232 "images/synthetic_example.png",233 caption="Synthetic Anomaly"234 )235 col1, col2 = st.columns(2)236 with col1:237 st.markdown("""238 <div>239 <h4>Synthetic Anomaly Generation</h4>240 <ol>241 <li><strong>Base Selection:</strong> Choose a normal image as base</li>242 <li><strong>Defect Injection:</strong> Apply artificial defects:243 <ul>244 <li>Scratch patterns</li>245 <li>Surface contamination</li>246 <li>Structural deformations</li>247 <li>Missing components</li>248 </ul>249 </li>250 <li><strong>Validation:</strong> Ensure defect realism</li>251 <li><strong>Integration:</strong> Add to validation dataset</li>252 </ol>253 </div>254 """, unsafe_allow_html=True)255 256 st.success("""257 ✨ **Benefits of Synthetic Data**258 - Creates diverse anomaly patterns259 - Controls defect characteristics260 - Balances class distribution261 - Reduces data collection costs262 """)263 264 with col2:265 266 st.warning("""267 ⚠️ **Important Considerations**268 - Synthetic defects must be realistic269 - Validation against real defects is crucial270 - Balance between synthetic and real data needed271 """)272 273 with tab2:274 st.header("Data Augmentation Techniques")275 276 st.info("""277 🔄 **Data Augmentation** applies label-preserving transformations to existing images278 to increase dataset variety and prevent overfitting. Unlike synthetic data, augmentation279 doesn't create new defect types but enhances model robustness through variations.280 """)281 282 st.write("""283 **Augmented data** is different from synthetic data. Instead of creating **new artificial images**, we **modify existing images** by applying **small transformations** like flipping, rotating, and resizing. 284 285 #### 🔍 Why Do We Need Data Augmentation?286 Even with synthetic data, **our dataset is still small** compared to what is needed for deep learning. 287 If we train a model on a **limited number of images**, the model might **memorize** the training data instead of **learning general patterns**. This is called **overfitting**.288 289 **To prevent overfitting, we increase the dataset size by applying transformations to images.** 290 """)291 292 col1, col2, col3 = st.columns(3)293 with col1:294 st.markdown("""295 <div>296 <h5>🔲 Basic Transformations</h5>297 <ul>298 <li><strong>Resizing</strong>: Fixed 224×224 pixels</li>299 <li><strong>Horizontal Flip</strong>: 50% probability</li>300 <li><strong>Impact</strong>: Standardized input size</li>301 </ul>302 </div>303 """, unsafe_allow_html=True)304 with col2:305 st.markdown("""306 <div>307 <h5>🎨 Geometric Operations</h5>308 <ul>309 <li><strong>Rotation</strong>: 75% prob. (-90°, 90°, 180°)</li>310 <li><strong>Scaling & Translation</strong>: 75% probability</li>311 <li><strong>Impact</strong>: Position invariance</li>312 </ul>313 </div>314 """, unsafe_allow_html=True)315 with col3:316 st.markdown("""317 <div>318 <h5>🔄 Advanced Processing</h5>319 <ul>320 <li><strong>Gaussian Blur</strong>: Kernel size 3</li>321 <li><strong>Sigma Range</strong>: 0.01-0.05</li>322 <li><strong>Final Step</strong>: Tensor conversion</li>323 </ul>324 </div>325 """, unsafe_allow_html=True)326 327 st.markdown("### 📈 Augmentation examples")328 col1, col2 = st.columns(2)329 with col1:330 st.image(331 "images/augmented_example.png",332 caption="Augmented Images Examples",333 use_container_width=True334 )335 with col2:336 st.image(337 "images/original_example.png",338 caption="Original Images Examples",339 use_container_width=True340 )341 342 st.success("""343 ✨ **Benefits of Data Augmentation**344 - Prevents overfitting345 - Improves model generalization346 - Increases effective dataset size347 - Maintains label validity348 """)349 350def resnet50_page():351 st.title("🔍 ResNet50 Feature Extraction")352 col1, col2= st.columns([3, 2])353 with col1:354 st.markdown("""355 ### Why Use a Pretrained Model (Transfer Learning)? 🧠356 Instead of starting from scratch, we take advantage of **ResNet50**, 357 a popular neural network that has already been trained on a large image dataset (ImageNet). 358 Because ResNet50 has “seen” many kinds of shapes, objects, and patterns,359 it has learned to recognize important features in images.360 361 By using these **pretrained features**, we:362 1. ⏱️ Save time and resources (no need to train a big model from zero).363 2. 🖼️ Gain access to a representation that already captures key visual patterns.364 3. 🎯 Focus on fine-tuning the model for our specific task (anomaly detection).365 """)366 367 with col2:368 st.image("images/resnet50.png", caption="Resnet50 Latent features extraction", use_container_width=True)369 370 st.markdown("""371 372 ### How We Extract Features from ResNet50373 We focus on two parts (or “blocks”) of ResNet50 and gather their outputs:374 - **Block 1**: Produces 512 features.375 - **Block 2**: Produces 1024 features.376 377 We then **combine** (concatenate) these for a total of **(512 + 1024) = 1536** features. 378 These numbers come from the internal layers of ResNet50.379 380 Essentially, these **1536 features** act like a summary of the image’s most important elements. 📝381 382 """)383 st.image("images/internal_features.png", caption="ResNet50 Block 1 & 2, random internal features", use_container_width=True)384 385def models_page():386 st.title("🤖 Anomaly Detection Models")387 388 tab1, tab2 = st.tabs(["KNN", "Autoencoder"])389 390 with tab1:391 st.header("KNN for Anomaly Detection") 392 col1, col2 = st.columns([2.5,2])393 with col1:394 st.markdown("""395 KNN (K-Nearest Neighbors) is a simple method that checks how "close" a new sample is 396 to existing samples. Here's the idea:397 1. We first collect "normal" images and extract their 1536 features. 398 We call this collection our **memory bank** of normal features. 🏦399 2. When a **test** image comes in:400 - ⚙️ We **extract** its 1536 features with the exact same process (ResNet blocks).401 - 📏 We measure its **distance** to each normal feature vector in the memory bank.402 - 🔎 We pick the **1 closest** neighbor (because k=1) and calculate the **average distance**.403 - If this average distance is **small**, it is likely "normal." ✅404 - If this average distance is **large**, it might be "anomalous" or unusual. 🐞405 """)406 st.image("images/k-nearest-neighbors-algorithm.png", caption="Visualizing the K-Nearest Neighbors approach")407 with col2:408 st.image("images/KNN_Pipeline.png", caption="KNN Anomaly Detection Pipeline")409 410 411 412 with tab2:413 st.header("Autoencoder for Anomaly Detection")414 col1, col2 = st.columns([2.5,2])415 with col1: 416 st.write("""417 Deep learning models have demonstrated remarkable performance in anomaly detection tasks, 418 particularly in complex scenarios where traditional methods often struggle. 419 These models can automatically learn intricate patterns and features from data, 420 making them highly effective at identifying subtle anomalies.421 """)422 423 st.write("""424 ### What is an autoencoder?425 426 An autoencoder is a specialized type of neural network designed to compress and reconstruct input images. 427 When applied to anomaly detection, the model is trained exclusively on normal images to learn the typical characteristics of the dataset.428 """)429 430 st.write("""431 ### Anomaly Detection Workflow432 433 #### 1. Feature Extraction434 - Collect a set of normal images.435 - Extract 1,536 features from each image using a **ResNet50** model.436 437 #### 2. Training Process438 - Train the autoencoder exclusively on normal data.439 - The autoencoder learns to efficiently encode and decode normal patterns.440 - The model optimizes its reconstruction error using normal samples.441 442 #### 3. Anomaly Detection443 - A new test image is passed through the trained autoencoder.444 - The reconstructed output is compared to the original input.445 - A high reconstruction error suggests an anomaly.446 - A low reconstruction error indicates that the sample is likely normal.447 """)448 449 st.success("""450 ## Key Advantages451 - **Fully Unsupervised Learning** – No need for labeled anomaly data.452 - **Ability to Capture Complex Normal Patterns** – The model generalizes well to unseen normal variations.453 - **Effective for High-Dimensional Image Data** – Works well with large and detailed datasets.454 """)455 st.warning("""456 ## Limitations457 - **Computationally Intensive Training** – Training deep autoencoders requires significant computational resources.458 - **Performance Sensitivity to Model Architecture** – The effectiveness depends heavily on model design.459 - **Difficulty Detecting Subtle Anomalies** – If anomalies resemble normal patterns closely, they may be overlooked.460 """)461 462 with col2:463 st.image("images/Autoencoder_Pipeline.png", caption="Autoencoder Anomaly Detection Pipeline", use_container_width=True)464 st.image("images/encoder_decoder.png", caption="Autoencoder structure", use_container_width=True)465 466def analysis_page():467 st.title("📊 Model Performance Analysis")468 469 # Move category selection outside of tabs470 selected_category = st.selectbox(471 "Select Category",472 ["bottle", "cable", "capsule", "carpet", "grid", 473 "hazelnut", "leather", "metal_nut", "pill", "screw",474 "tile", "toothbrush", "transistor", "wood", "zipper"],475 index=1 # Default to 'cable'476 )477 478 # Create two columns479 col1, col2 = st.columns(2)480 481 with col1:482 st.subheader("KNN Model")483 try:484 confusion_matrices, roc_curves, auc_scores, f1_scores = load_evaluation_metrics('models/evaluation_metrics_knn.pkl')485 486 # Display ROC curve with unique key487 roc_fig = plot_roc_curve(selected_category, roc_curves, auc_scores)488 st.plotly_chart(roc_fig, use_container_width=True, key="knn_roc")489 490 # Display confusion matrix with unique key491 cm_fig = plot_confusion_matrix(selected_category, confusion_matrices, f1_scores)492 st.plotly_chart(cm_fig, use_container_width=True, key="knn_cm")493 494 except FileNotFoundError:495 st.error("KNN evaluation metrics file not found. Please run model evaluation first.") 496 except Exception as e:497 st.error(f"Error loading KNN metrics: {str(e)}")498 499 with col2:500 st.subheader("Autoencoder Model")501 try:502 confusion_matrices, roc_curves, auc_scores, f1_scores = load_evaluation_metrics('models/evaluation_metrics_autoencoder.pkl')503 504 # Display ROC curve with unique key505 roc_fig = plot_roc_curve(selected_category, roc_curves, auc_scores)506 st.plotly_chart(roc_fig, use_container_width=True, key="ae_roc")507 508 # Display confusion matrix with unique key509 cm_fig = plot_confusion_matrix(selected_category, confusion_matrices, f1_scores)510 st.plotly_chart(cm_fig, use_container_width=True, key="ae_cm")511 512 except FileNotFoundError:513 st.error("Autoencoder evaluation metrics file not found. Please run model evaluation first.")514 except Exception as e:515 st.error(f"Error loading Autoencoder metrics: {e}")516 517 518 519def prediction_page():520 """521 Streamlit page to view anomaly detection images.522 """523 st.title("🔍 Anomaly Detection - Image Viewer")524 525 st.info("Select a category and image to view the anomaly detection result.")526 527 # Select a category528 selected_category = st.selectbox(529 "Select a Category",530 ["bottle", "cable", "capsule", "carpet", "grid", 531 "hazelnut", "leather", "metal_nut", "pill", "screw",532 "tile", "toothbrush", "transistor", "wood", "zipper"],533 key="shared_category" # Unique key for each selectbox534 )535 536 # List available images in the selected category537 category_dir_knn = os.path.join("images/Dataset_knn", selected_category)538 category_dir_autoencoder = os.path.join("images/Dataset_autoencoder", selected_category)539 540 available_images_knn = []541 if os.path.exists(category_dir_knn):542 available_images_knn = [f for f in os.listdir(category_dir_knn) if os.path.isfile(os.path.join(category_dir_knn, f))]543 544 available_images_autoencoder = []545 if os.path.exists(category_dir_autoencoder):546 available_images_autoencoder = [f for f in os.listdir(category_dir_autoencoder) if os.path.isfile(os.path.join(category_dir_autoencoder, f))]547 548 # Find common images549 available_images = list(set(available_images_knn) & set(available_images_autoencoder))550 551 selected_image = st.selectbox(552 "Select an Image", 553 available_images,554 key="shared_image" # Unique key for each selectbox555 )556 557 run_prediction = st.button("Run Prediction")558 559 tab1, tab2 = st.tabs(["KNN", "Autoencoder"])560 561 with tab1:562 st.subheader("KNN Model")563 if run_prediction:564 display_image("knn", selected_category, selected_image)565 566 with tab2:567 st.subheader("Autoencoder Model")568 if run_prediction:569 display_image("autoencoder", selected_category, selected_image)570 571def display_image(model_type, selected_category, selected_image):572 """573 Helper function to display a single image based on the selected model type.574 """575 if selected_image:576 # Construct file paths577 image_path = os.path.join(f"images/Dataset_{model_type}", selected_category, selected_image)578 579 # Display image580 if os.path.exists(image_path):581 try:582 image = Image.open(image_path)583 st.image(image, use_container_width=True)584 except Exception as e:585 st.error(f"Error displaying {model_type} image: {e}")586 else:587 st.error(f"{model_type} Image not found: {image_path}")588 589 st.markdown("""590 ### Explanation of the displayed image:591 592 1️⃣ **Top Left: Original Image**593 - This is the raw image from the dataset.594 - The object is analyzed to detect potential anomalies.595 596 2️⃣ **Top Right: Heatmap**597 - The heatmap represents the anomaly score distribution.598 - Color Legend:599 - 🔴 Red/Yellow: High anomaly score (defective region).600 - 🔵 Blue: Normal areas with low anomaly probability.601 - The defect region is highlighted based on the anomaly model’s prediction.602 603 3️⃣ **Bottom Left: Segmentation Map**604 - This is a binary mask that highlights the detected defect areas.605 - White pixels represent anomalous regions identified by the model.606 - It is created by thresholding the heatmap to localize defects.607 608 4️⃣ **Bottom Right: Ground Truth**609 - The ground truth mask is a manually labeled reference.610 - It defines the true defective areas for validation.611 - The segmentation map should ideally match this mask for accurate detection.612 """)613 614 615 616def conclusion_and_improvements_page():617 st.title("🏁 Conclusion and Future Improvements")618 619 st.write("""620 ## Conclusion621 622 Throughout the project, several Machine Learning and Deep Learning models have been applied to the available image data. We compared different models and approaches in data preparation to optimize anomaly detection performance.623 624 The evaluation shows that the Convolutional Autoencoder and the KNN approach significantly outperform other models, and both yield better results using the deep feature extraction approach.625 626 ### Key Takeaways627 628 - **Deep Feature Extraction**: Significantly outperforms manual extracted features, as seen in the consistently better performance of ResNet50-based methods.629 - **ResNet50 Block Performance**: Two evaluated deep feature extraction approaches, one using ResNet50 blocks 1 and 2, the other using block 3, both seem similarly suitable on average over all categories. However, within single categories, one often showed significantly better performance than the other.630 - **KNN Approach**: Leverages memory-based similarity comparisons, providing a robust alternative to parametric models.631 - **Autoencoder Approach**: Deep learning using an autoencoder provides strong reconstruction-based anomaly detection, ensuring comprehensive feature learning.632 - **Synthetic Validation Data**: Introduced diversity and improved cross-validation but also posed challenges, as some anomalies may not perfectly mimic real-world defects.633 634 ## Future Improvements635 636 - **Data Augmentation Refinement**: Exploring more advanced augmentation techniques, such as generative models (e.g., GANs), could enhance training diversity.637 - **Model Ensembling**: Combining multiple anomaly detection models could improve robustness and generalization.638 - **Hyperparameter Optimization**: Fine-tuning model hyperparameters further, using techniques like Bayesian optimization, could boost performance.639 - **Alternative Architectures**: Exploring transformer-based architectures or vision encoders like ViTs for anomaly detection could yield even better results.640 - **Better Synthetic Data**: Refining the process of synthetic anomaly generation to better align with real-world defects.641 642 Overall, the introduced approaches provide a solid foundation for industrial anomaly detection tasks. Further optimizations and explorations in feature extraction and model selection could push performance even higher in future studies.643 """)644 645# Bibliography Page646def bibliography_page():647 st.title("📚 Bibliography & References")648 649 st.markdown("""650 ### Core Papers & Methods651 652 1. Liu, J., Xie, G., Wang, J., Li, S., Wang, C., Zheng, F., & Jin, Y. (2023). 653 _Deep Industrial Image Anomaly Detection: A Survey_. Springer Nature. 654 [arXiv:2301.11514](https://arxiv.org/abs/2301.11514)655 656 2. Yang, J., Shi, Y., & Qi, Z. (2020).657 _DFR: Deep Feature Reconstruction for Unsupervised Anomaly Segmentation_. 658 [arXiv:2012.07122](https://arxiv.org/abs/2012.07122)659 660 3. Bühler, J., Fehrenbach, J., Steinmann, L., Nauck, C., & Koulakis, M. (2024).661 _Domain-independent detection of known anomalies_. Karlsruhe Institute of Technology (KIT) & preML GmbH. 662 [arXiv:2407.02910](https://arxiv.org/abs/2407.02910)663 664 4. Heckler, L., & König, R. (2024).665 _Feature Selection for Unsupervised Anomaly Detection and Localization Using Synthetic Defects_.666 MVTec Software GmbH & Technical University of Munich.667 In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024), 154-165. 668 [DOI: 10.5220/0012385500003660](https://doi.org/10.5220/0012385500003660)669 670 5. Rippel, O., Mertens, P., & Merhof, D. (2020).671 _Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection_.672 RWTH Aachen University. 673 [arXiv:2005.14140](https://arxiv.org/abs/2005.14140)674 675 6. Bergmann, P., Fauser, M., Sattlegger, D., & Steger, C. (2019).676 _MVTec AD – A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection_.677 MVTec Software GmbH. 678 [MVTec AD Dataset](https://www.mvtec.com/company/research/datasets/mvtec-ad)679 680 7. Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., & Gehler, P. (2022).681 _Towards Total Recall in Industrial Anomaly Detection_.682 University of Tübingen & Amazon AWS. 683 [arXiv:2106.08265](https://arxiv.org/abs/2106.08265)684 685 8. Zheng, Y., Wang, X., Qi, Y., Li, W., & Wu, L. (2022).686 _Benchmarking Unsupervised Anomaly Detection and Localization_.687 University of Chinese Academy of Sciences, SenseTime Research, & Tsinghua University. 688 [arXiv:2205.14852](https://arxiv.org/abs/2205.14852)689 """)690 691def main():692 with st.sidebar:693 st.image("images/logo.png", width=100) # Add your logo here694 st.title("Navigation")695 696 697 selection = st.radio(698 "Select a Section",699 ["Overview", 700 "Dataset & EDM",701 "Synthetic Data & Augmentation",702 "Transfer Learning - Resnet50",703 "Models",704 "Analysis",705 "Prediction",706 "Conclusion and Improvements",707 "Bibliography"],708 format_func=lambda x: f" {x}"709 )710 st.markdown("---")711 st.markdown("""712 <div style='text-align: center; color: #666;'>713 714 <div style='margin: 10px 0;'>715 This app is maintained by:<br>716 <a href="https://www.linkedin.com/in/giuseppe-rumore-b2599961" target="_blank">Giuseppe Rumore</a> |717 <a href="https://www.linkedin.com/in/micaela-w%C3%BCnsche-9baaa710b/" target="_blank">Micaela Wünsche</a> |718 <a href="https://www.linkedin.com/in/majid-jafari-62909071/" target="_blank">Majid Jafari</a>719 </div>720 </a>721 <img src="https://content.linkedin.com/content/dam/me/business/en-us/amp/brand-site/v2/bg/LI-Logo.svg.original.svg" 722 width="80" 723 alt="LinkedIn"724 style="margin-top: 10px;">725 </a>726 <br>727 <a href="https://github.com/pepperumo/MVTEC-anomaly-detection" target="_blank">728 <img src="https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png"729 width="40"730 alt="GitHub"731 style="margin-top: 10px; border-radius: 50%;">732 </a>733 <div style='text-align: center; color: #666;'>734 <small>Version 1.0.0</small><br>735 <small>© 2025 MVTec Anomaly Detection</small><br>736 </div>737 """, unsafe_allow_html=True)738 739 if selection == "Overview":740 overview_page()741 elif selection == "Dataset & EDM":742 dataset_page()743 elif selection == "Synthetic Data & Augmentation":744 synthetic_data_page()745 elif selection == "Transfer Learning - Resnet50":746 resnet50_page()747 elif selection == "Models":748 models_page()749 elif selection == "Analysis":750 analysis_page()751 elif selection == "Prediction":752 prediction_page()753 elif selection == "Conclusion and Improvements":754 conclusion_and_improvements_page()755 elif selection == "Bibliography":756 bibliography_page()757 758if __name__ == "__main__":759 main()760