---

✨ What is Visiopedia?

Visiopedia is an interactive visual concept explorer that demonstrates how machines can learn and reason about images without explicit labels.

Instead of predicting class names like a traditional classifier, Visiopedia:

• Learns visual concepts from a small set of example images
• Builds self-supervised visual prototypes
• Explains what it “sees” in new images using visual similarity, not predictions

💡 In short: Visiopedia shows how machines understand images by comparing them to previously seen visual examples.

---

🔬 Background: The Betel Leaf SSL Project

Visiopedia is built on top of a larger academic mini-project focused on Self-Supervised Learning (SSL) using the Betel Leaf dataset.

In the original project, I conducted a complete, research-oriented study that included:

• Exploratory Data Analysis (EDA) on the Betel Leaf dataset
• Supervised baselines using pretrained CNN architectures
• Self-Supervised Learning using SimSiam
• Label-free pretraining and downstream evaluation
• Linear probing, k-NN evaluation, and full fine-tuning
• Embedding analysis using PCA, t-SNE, and UMAP
• Label-efficiency studies and ablation experiments

The SimSiam encoder trained during this project is the core visual representation model used in Visiopedia.

📄 For full academic details, experiments, and works, please refer to the Project Description PDF included in this repository. For Training Notebook or Model you can contact anytime.

🧠 From Research to Visiopedia

After completing the Betel Leaf SSL project, I extended the work into Visiopedia to make the learned representations:

• Interactive
• Explainable
• Usable beyond a single dataset

Visiopedia transforms the trained SSL encoder into a concept-learning and reasoning system, allowing users to teach, test, and explore visual understanding directly in a web interface.

---

🚀 What Visiopedia Does

✅ Teaching Phase (No Labels Required)

• Upload at least 3 images of the same object [More Picture More Detailed Learn]
• Assign a human-readable name [The name is not learned, It is only used to describe similarity results]
• Visiopedia builds a visual prototype using self-supervised embeddings

🔍 Testing Phase (Similarity-Based Reasoning)

When a new image is uploaded, Visiopedia:

• Compares it against all learned visual prototypes

• Measures relative similarity in embedding space

• Explains results using natural language, such as:

  • “I can clearly see this object”
  • “I might also see this one, but I’m not confident”
  • “I don’t see any trace of this object”

• Multiple objects can be detected simultaneously if they are visually present.

• Visiopedia itself is dataset-agnostic and can be used with any images

🌐 How to Use Visiopedia

Visiopedia works in two simple stages: Teach and Explore.
You do not train a classifier. Instead, you teach the system visual concepts and observe how it reasons using visual similarity.

🔹 Point 1: Teach Visual Objects

Navigate to Step 1: Teach Visual Objects.

For each visual object:

1. Upload at least 3 images of the same object
2. Enter a name for the object
   • This name is for your reference only
3. Click Save Object

Repeat this process to add more objects.

Key Details

• You can add up to 5 visual objects
• Saved objects appear in the Learned Objects gallery
• The system does not learn the name
• It learns the shared visual pattern across the uploaded images

🔹 Point 2: Lock the Knowledge

Once you finish teaching:

1. Click Lock & Start Testing

This action:

• Freezes all learned objects
• Prevents further edits during testing
• Ensures fair and consistent comparison

You can unlock the system at any time if you need to modify or add objects.

🔹 Point 3: Explore New Images

After Lock It Will Go to Step 2: Test Images.

1. Upload one or more new images

For each image, Visiopedia will:

• Compare it against all learned visual objects
• Display relative similarity bars
• Generate a natural language explanation describing what it perceives, such as:
  • Clearly visible objects
  • Possibly present objects
  • Weak or uncertain traces
  • No recognized objects

Multiple objects can be detected in a single image if sufficient visual similarity exists.

🔹 Point 4: Understanding the Output

• Confidence bars represent relative similarity, not probability
• Higher bars indicate stronger visual resemblance
• The explanation text summarizes what the model “sees”

Important Clarifications

• Visiopedia does not classify
• It does not predict labels
• All reasoning is based purely on visual similarity in embedding space

🔹 Point 5: Editing or Resetting

You can unlock the system at any time to:

• Delete existing objects
• Add new objects
• Teach different visual concepts

After making changes, lock the system again to resume testing.

---

⚠️ Important Notes

• Results depend on visual similarity, not semantic correctness
• The system may detect multiple objects in one image
• If no learned object matches, Visiopedia will explicitly state that nothing was recognized

---

✨ Key Features

• 🧠 Self-supervised visual representation learning (SimSiam)
• 🏷️ No label-based classification
• 📦 Prototype-based concept memory (up to 5 objects)
• 🔐 Lock / Unlock workflow to prevent accidental edits
• 📊 Confidence bars for each learned object
• 🗑️ Object management (add / delete concepts)
• 🌐 Interactive Streamlit web app
• 🎨 Premium, Friendly UI

🛠️ Tech Stack

• Python
• PyTorch
• SimSiam (Self-Supervised Learning)
• NumPy
• Streamlit
• PIL

Training and experimentation were conducted using Kaggle / Colab environments, while Visiopedia is designed as a lightweight interactive deployment.

🎉 Happy exploring visual reasoning with Visiopedia. Hope you like it. 🎉

---