README.md

---
title: SiameseCNN ImgMatching
emoji: 📚
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
short_description: 🔢 Digit matching & recognition  through Siamese CNN
---
# 🔢 Siamese CNN: Number Recognition & Similarities 
**Note:** See model performances and examples [here 📈](https://github.com/Lulloooo/SiameseNet-NumRecognition/blob/main/Performances-Examples.md)
<br/><br/>
## 🎯 PURPOSES
This project aims to develop a Siamese Neural Network 🧠 able to match and identify different kinds of numerical digits 🔢.  
The model learns to measure how similar two images are rather than directly classifying them. This makes it suitable for tasks like:  

- **🤝 Digits matching and recognition**: Compare the uploaded digits with those in the test set. Once a match is found, label the uploaded digit with the same label as the matched one.
- **🔓 Security unlocking mechanisms**: The camera captures an istant image of a face or an object. If the similarity with those in the training pool is high, it unlocks the device/program.
- **🔎 Duplicate detection**: Define a similarity score between the external image and those in the reference set. If similarity with one of these is high, it is likely is a duplicate.
<br/><br/>
## 🛠️ WORKFLOW
The workflow is quite straightforward: the user is prompted to upload a picture of a number or to draw it in a canvas, and the Siamese Network returns:  
- 1️⃣ Which digit is the uploaded/drawn picture (The uploaded img is a ...) along with its similarity score.
- 2️⃣ 3 matching pics coming from the testing dataset.
<br/><br/>
## ⚠️ WARNINGS 
Be fully aware of the model's limitations by reading the [model limitations doc ✋](https://github.com/Lulloooo/SiameseNet-NumRecognition/blob/main/Performances-Examples.md)
<br/><br/>
## 🧠 MODEL OVERVIEW  
The Siamese architecture leverages shared convolutional layers to extract features from two images, then computes their Euclidean distance in embedding space. A contrastive loss function is employed:

&nbsp;&nbsp;**L = (1 - Y) * (1/2) * D^2 + Y * (1/2) * (max(0, m - D))^2**

with:
- D: distance between embeddings
- Y = 0 for similar and Y = 1 for dissimilar
- m: margin parameters (can be changed)
<br/><br/>

__________________________________________________________________________________________________________
**Note**: I got the idea to implement this kind of network for this kind of task + some code chunks from the [Daily Dose of Data Science](https://www.dailydoseofds.com/) newsletter 📨. They are really awesome, and everyone interested in data science should check them too 😉.  
  

**Note2**: As this was a work-related project, it has been approved for posting, and all sensitive information and data are omitted to protect privacy. Code snippet containing sensitive data has not been posted. 
__________________________________________________________________________________________________________