https://github.com/preethi2805/image_processing_cnn_using_pytorch

📸 Image Processing & CNN Classification with PyTorch This project explores feature extraction in image processing, pooling operations, and CNN-based classification using PyTorch and the load_digits dataset.
https://github.com/preethi2805/image_processing_cnn_using_pytorch

convolutional-neural-networks kernel pooling pytorch

Last synced: about 1 month ago
JSON representation

📸 Image Processing & CNN Classification with PyTorch This project explores feature extraction in image processing, pooling operations, and CNN-based classification using PyTorch and the load_digits dataset.

• Host: GitHub
• URL: https://github.com/preethi2805/image_processing_cnn_using_pytorch
• Owner: Preethi2805
• Created: 2025-02-20T03:17:02.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2025-02-20T03:51:41.000Z (over 1 year ago)
• Last Synced: 2025-10-29T06:39:49.316Z (8 months ago)
• Topics: convolutional-neural-networks, kernel, pooling, pytorch
• Language: Jupyter Notebook
• Homepage:
• Size: 312 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
### **README.md**  

# 🖼️ **Feature Extraction & CNN Classification with PyTorch**  

## 📌 **Overview**  

This repository contains five tasks designed to explore **image feature extraction, pooling, CNN training, and model comparison** using PyTorch and scikit-learn.  

### 🔍 **Tasks Overview:**  

✅ **Task 1:** Implement convolutional filters for **edge detection, blurring, and identity mapping**.  

✅ **Task 2:** Apply **max-pooling & average-pooling** to understand spatial resolution impact.  

✅ **Task 3:** Build a **CNN from scratch** using PyTorch for digit classification.  

✅ **Task 4:** Train the CNN with **SGD optimization** and monitor loss trends.  

✅ **Task 5:** Compare CNN vs **Support Vector Classifier (SVC)** on classification performance.  

---

## 🏗️ **Task 1: Implementing Convolutional Filters**  

This task explores **image feature extraction** using **custom convolutional filters** implemented in PyTorch.  

### 🔬 **Implemented Filters:**  

| **Kernel Type**      | **Purpose**                      |

|----------------------|---------------------------------|

| **Horizontal Edge**  | Detects horizontal edges        |

| **Vertical Edge**    | Detects vertical edges          |

| **Diagonal Edge 1**  | Detects main diagonal edges     |

| **Diagonal Edge 2**  | Detects anti-diagonal edges     |

| **Blurring Kernel**  | Smoothens the image            |

| **Identity Kernel**  | Leaves the image unchanged     |

### 🏗 **Implementation Details:**  

- Define **custom convolution kernels** in PyTorch.  

- Implement `corr2d()` function to **perform 2D convolution** manually.  

- Use **padding (2 pixels)** to handle edge detection at boundaries.  

- Apply each filter to **5 sample images** and visualize the results.

### 📷 **Example Visualization:**  

💡 **Before & After Applying Filters**  

| ![original](task1.png) | 

---

## 🔄 **Task 2: Pooling Operations**  

To understand the impact of pooling on feature maps, we:  

- Implement `pool2d()` for **Max Pooling** & **Average Pooling**.  

- Apply it to **feature maps** from Task 1.  

- Observe how pooling **reduces resolution but retains key features**.  

### 🏗 **Implementation Details:**  

✅ Supports **Max-Pooling** & **Average-Pooling**  

✅ Customizable **pool size** parameter  

✅ Visualizes pooled feature maps  

### 📉 **Dimensionality Reduction Example:**  

```

Before Pooling: (8, 8)

After Pooling (2x2): (4, 4)

```

📷 **Example Pooled Image Output:**  

| Original Feature Map | Max-Pooled (2x2) | Avg-Pooled (2x2) |

|----------------------|------------------|------------------|

| ![original](task2.png) |

---

## 🤖 **Task 3: Building a CNN in PyTorch**  

This task involves implementing a **Convolutional Neural Network (CNN)** from scratch.  

### **🛠 Network Architecture:**  

| Layer Type | Filters/Neurons | Kernel Size | Activation |

|-----------|---------------|-------------|------------|

| Conv1 | 8 | 3x3 | ReLU |

| Conv2 | 3 | 3x3 | ReLU |

| Max Pooling | - | 2x2 | - |

| Fully Connected (FC1) | 120 | - | ReLU |

| Fully Connected (FC2) | 84 | - | ReLU |

| Output Layer | 10 | - | Softmax |

---

## 🎯 **Task 4: Training the CNN on load_digits Dataset**  

- **Preprocess the dataset** by normalizing pixel values **(0 to 1)**.  

- **Split the dataset** (40% training, 40% validation, 20% test).  

- **Train the CNN** using **SGD optimizer (lr=0.001, momentum=0.9)**.  

- **Train for 1000 epochs**, monitoring **training & validation loss**.  

- **Checkpointing:** Save the best model when **validation loss improves**.  

- **Use tqdm progress bar** to track training.

📉 **Loss Trends Over Time:**  

```

Epoch 1000: Accuracy = 93.89%

```

📈 **Loss Graphs:**  

![loss_graph](task4.png)  

---

## ⚖️ **Task 5: Comparing CNN vs SVC**  

- **Train an SVM model using sklearn's SVC** with `gamma=0.001`.  

- **Flatten images** (treating pixels as independent features).  

- **Compare CNN and SVC performance on the test set**.  

### 📊 **Model Comparison Results:**  

| Model | Test Accuracy |

|-------|--------------|

| **CNN** | **93.89%** |

| **SVC (gamma=0.001)** | **97.04%** |

### 📝 **Observations & Improvements:**  

**The SVM model is performing better compared to the CNN model.**

---

## 🤝 **Contributing**  

Feel free to **open an issue** or **submit a pull request** if you have improvements or suggestions!

---

### ✅ **Let me know if you need any modifications! 🚀**