---

TL;DR:
This project implements Convolutional Neural Networks (CNNs) for classifying natural and man-made scenes from the Intel Image Dataset.
It compares two models — a VGG19 Transfer Learning CNN and a Custom CNN trained from scratch — to evaluate their performance on real-world scene recognition tasks.

---

• Project Overview
• Features
• Technologies & Tools
• Dataset
• Getting Started
  • Prerequisites
  • Installation
  • Usage
• Model Architectures
• Results & Comparison
  • Key Observation
  • Graphs of Training Loss & Accuracy
• Project Structure
• License

---

This project focuses on developing and evaluating Convolutional Neural Networks (CNNs) for the classification of images from the Intel Image Dataset.

The dataset consists of images categorized into six distinct classes:
🏢 Buildings | 🌲 Forest | 🏔 Glacier | ⛰ Mountain | 🌊 Sea | 🛣 Street

Two different approaches are implemented to assess performance and effectiveness:

1. CNN Model with Transfer Learning – A model leveraging VGG19, a pre-trained deep learning architecture, to enhance feature extraction and improve classification accuracy.
2. CNN Model Trained from Scratch – A custom-built convolutional neural network trained without any pre-existing weights.

 
 🡅 Back to Top
 

---

• Deep Learning Model: CNN-based classifier implemented using TensorFlow/Keras.
• Transfer Learning: Uses VGG19 as the pre-trained model for improved accuracy.
• Performance Metrics: Evaluates accuracy, precision, recall, and confusion matrix.
• Modular Code Structure: Well-organized for easy modification and experimentation.

 
 🡅 Back to Top
 

---

• IDE: Jupyter Lab
• Programming Language: Python
• Deep Learning Framework: TensorFlow/Keras
• Data Processing: OpenCV, NumPy, Pandas
• Visualization: Matplotlib, Seaborn
• Hardware Acceleration: GPU (CUDA-enabled for TensorFlow)

 
 🡅 Back to Top
 

---

The Intel Image Dataset consists of images categorized into six natural and man-made scenery classes. It is a widely used benchmark dataset for scene recognition and classification tasks. The dataset is structured into training, validation, and test sets to facilitate model evaluation.

Dataset Structure

• Train Set: 14,034 images
• Test Set: 3,000 images
• Prediction Set: 7,301 images

Tip

You can download the dataset from here: Intel Image Classification Dataset

 
 🡅 Back to Top
 

---

🔧 Prerequisites

Important

• Ensure Anaconda is installed, if not you can download from Anaconda and also Git (if not available, download from Github).
• Also download the mentioned dataset before running any of the notebooks, and change the paths in the notebooks whereever necessary.

⚙️ Installation

Once Anaconda is installed, open the Anaconda Prompt and run the following commands:

1. Clone the repository:

git clone https://github.com/NSANTRA/Intel-CNN-Image-Classification

2. Navigate to the project directory:

cd Intel-CNN-Image-Classification

3. Create a new Conda environment:

conda env create -f "Tensorflow.yml"

4. Activate the environment:

conda activate Tensorflow

▶️ Usage

After activating the environment:

• Open Jupyter Notebook or JupyterLab within the environment.
• Navigate to the project folder and open the desired notebook.
• Ensure dataset paths are correctly configured in each notebook.
• Run the cells sequentially to execute the project.

 
 🡅 Back to Top
 

---

1️⃣ CNN Model with Transfer Learning (VGG19)

Layer (Type)	Output Shape	Parameters
vgg19 (Functional)	(None, 4, 4, 512)	20,024,384
flatten_2 (Flatten)	(None, 8192)	0
dense_8 (Dense)	(None, 512)	4,194,816
batch_normalization_6 (BatchNormalization)	(None, 512)	2,048
dense_9 (Dense)	(None, 256)	131,328
batch_normalization_7 (BatchNormalization)	(None, 256)	1,024
dense_10 (Dense)	(None, 128)	32,896
batch_normalization_8 (BatchNormalization)	(None, 128)	512
dense_11 (Dense)	(None, 6)	774

• Non-Trainable Parameters: 20,026,176

• Trainable Parameters: 4,361,606

• Total Parameters: 24,387,782

• Optimizer: Adam

• Loss Function: Sparse Categorical Crossentropy

2️⃣ CNN Model (Trained from Scratch)

Layer (Type)	Output Shape	Parameters
conv2d (Conv2D)	(None, 150, 150, 64)	1,792
batch_normalization (BatchNormalization)	(None, 150, 150, 64)	256
conv2d_1 (Conv2D)	(None, 150, 150, 64)	36,928
batch_normalization_1 (BatchNormalization)	(None, 150, 150, 64)	256
max_pooling2d (MaxPooling2D)	(None, 75, 75, 64)	0
conv2d_2 (Conv2D)	(None, 75, 75, 128)	73,856
batch_normalization_2 (BatchNormalization)	(None, 75, 75, 128)	512
conv2d_3 (Conv2D)	(None, 75, 75, 128)	147,584
batch_normalization_3 (BatchNormalization)	(None, 75, 75, 128)	512

• Non-trainable Parameters: 1,216

• Trainable Parameters: 22,701,734

• Total Parameters: 22,702,950

• Optimizer: Adam

• Loss Function: Sparse Categorical Crossentropy

 
 🡅 Back to Top
 

---

Classification Report for CNN With Transfer Learning

	Precision	Recall	F1-Score	Support
Buildings	0.92	0.89	0.91	437
Forest	0.96	0.99	0.98	474
Glacier	0.83	0.80	0.82	553
Mountain	0.82	0.83	0.82	525
Sea	0.93	0.91	0.92	510
Street	0.89	0.93	0.91	501
accuracy			0.89	3000
macro avg	0.89	0.89	0.89	3000
weighted avg	0.89	0.89	0.89	3000

🔹 Key Observations:

• ✅ High Overall Accuracy: 89% – The model performs well across all classes.
• ✅ Forest category has the highest accuracy (Precision: 0.96, Recall: 0.99, F1-score: 0.98) – Very few misclassifications.
• ✅ Buildings, Sea, and Street categories also perform well (F1-score: ~0.91).
• ✅ Glacier and Mountain have the lowest scores (F1-score: ~0.82) – These categories are harder to classify correctly.

Class-Wise Weights

Category	Precision	Recall	F1-Score	Observations
Buildings	0.92	0.89	0.91	Some Buildings misclassified as Streets.
Forest	0.96	0.99	0.98	Best performing class – almost perfect classification.
Glacier	0.83	0.80	0.82	Some Glaciers misclassified as Mountains.
Mountain	0.82	0.83	0.82	Often confused with Glaciers.
Sea	0.93	0.91	0.92	Often confused with Glaciers.
Street	0.89	0.93	0.91	Often confused with Glaciers.

🔹 Key Takeaways:

• 📌 Transfer Learning significantly boosts accuracy, with an overall F1-score of 0.89.
• 📌 Forest classification is near-perfect, while Glacier and Mountain have the most confusion.
• 📌 Further improvements can be made by refining the model’s ability to differentiate Glaciers and Mountains.

---

Classification Report for CNN Without Transfer Learning

	Precision	Recall	F1-Score	Support
Buildings	0.86	0.80	0.83	437
Forest	0.93	0.96	0.94	474
Glacier	0.79	0.66	0.72	553
Mountain	0.78	0.75	0.77	525
Sea	0.75	0.88	0.81	510
Street	0.83	0.88	0.86	501
accuracy			0.82	3000
macro avg	0.82	0.82	0.82	3000
weighted avg	0.82	0.82	0.82	3000

🔹 Key Observations

• ✅ Overall Accuracy: The CNN model achieves an accuracy of 82%, showing strong generalization across multiple scene classes.
• ✅ Best Performing Class: Forest shows the highest performance (Precision: 0.93, Recall: 0.96, F1-score: 0.94), indicating clear, distinctive visual features.
• ✅ Moderate Performance: Buildings, Street, and Sea classes show solid results (F1-scores ≈ 0.83–0.86), though some confusion occurs among visually similar categories.
• ⚠️ Challenging Categories: Glacier and Mountain have the lowest F1-scores (0.72 and 0.77), suggesting overlap in texture and color patterns between these landscapes.

Class-Wise Weights

Category	Precision	Recall	F1-Score	Observations
Buildings	0.86	0.80	0.83	Good accuracy (F1: 0.83); likely confused with Street scenes due to similar man-made structures.
Forest	0.93	0.96	0.94	Excellent classification; high recall (0.96) shows strong detection capability for vegetation.
Glacier	0.79	0.66	0.72	Underperforms (F1: 0.72); frequent confusion with Mountain due to similar snowy textures.
Mountain	0.78	0.75	0.77	Moderate accuracy (F1: 0.77); overlaps visually with Glacier terrain.
Sea	0.75	0.88	0.81	Performs well (F1: 0.81); sometimes confused with Glacier scenes.
Street	0.83	0.88	0.86	Solid performance (F1: 0.86); some misclassifications with Buildings.

🔹 Key Takeaways

• 📌 The CNN without transfer learning achieves a balanced overall performance (F1-score: 0.82).
• 📌 The model performs best on natural, distinct textures (Forest), while struggling with visually overlapping classes (Glacier, Mountain).
• 📌 Feature extraction could be improved with deeper architectures or transfer learning to boost recognition of subtle scene differences.
• 📌 These results form a strong baseline for comparison against models using transfer learning.

---

Confusion Matrices

CNN With Transfer Learning

Key Observations:

• ✅ High overall accuracy, fewer misclassifications compared to the second model.
• ✅ Forest category is nearly perfect – 468 out of 474 correctly classified.
• ✅ Buildings misclassified mainly as Streets (43 cases) – likely due to urban similarities.
• ✅ Glacier vs. Mountain confusion – 69 Glacier images misclassified as Mountains.
• ✅ Minimal errors in the Sea category – 465 out of 510 correctly classified.

Major Misclassifications:

• Glacier mistaken as Mountain (69 cases) – Snow-covered landscapes might be confusing.
• Street misclassified as Buildings (43 cases) – Similar structures in urban settings.
• Sea occasionally confused with Glacier & Mountain – Landscape similarities.

Takeaway:

• 🔥 Transfer Learning improves classification significantly, but Glacier vs. Mountain remains a challenge.

CNN From Sratch

Key Observations:

• ❌ Lower overall accuracy – More misclassifications across most categories.
• ❌ Glacier category struggles the most – 99 Glaciers classified as Mountains (compared to 69 in TL model).
• ❌ Buildings misclassified as Streets (67 cases) – Worse than Transfer Learning model (43 cases).
• ✅ Forest category still performs well – 453 correctly classified out of 474.
• ❌ Sea and Mountain confusion is more frequent than in the Transfer Learning model.

Major Misclassifications:

• Glacier confused with Mountain (99 cases) – Worse than Transfer Learning model.
• Street misclassified as Buildings (41 cases) – A bit better than the TL model but still a concern.
• Sea misclassified as Glacier (16 cases) – More than in the TL model.

Takeaway:

• 🚨 Training from scratch struggles more, particularly with Glacier-Mountain and Sea-Glacier distinctions. Transfer Learning is clearly more effective for this task.

Summary

Key Observations and Comparison

Metric	Transfer Learning CNN	CNN From Scratch
Overall Accuracy	Higher (Fewer misclassifications)	Lower (More misclassifications)
Buildings Accuracy	391 correctly classified, 43 misclassified as Street	351 correctly classified, 67 misclassified as Street
Forest Accuracy	468 correctly classified, almost no errors	453 correctly classified, some errors
Glacier vs. Mountain	69 Glaciers misclassified as Mountains	99 Glaciers misclassified as Mountains (worse)
Sea vs. Mountain	Few misclassifications	More confusion between Sea and Mountain
Street vs. Buildings	Some confusion but better handling	More Streets misclassified as Buildings

Key Takeaways

• ✅ Transfer Learning performs better overall
• ✅ CNN from Scratch struggles more with Glacier & Mountain misclassifications
• ✅ Transfer Learning model has more confident predictions, fewer mixed-up cases
• ✅ Fine-tuning the CNN from Scratch might improve its performance

---

📈 Graphs of Training Loss and Accuracy

CNN With Transfer Learning

CNN Without Transfer Learning

 
 🡅 Back to Top
 

---

Intel-CNN-Image-Classification/
├── Dataset/                                             # Intel Image Dataset
│   └── seg_train/                                       # Training Images
│   └── seg_test/                                        # Testing Images
│   └── seg_pred/                                        # Unlabeled Images                                    
├── Models/                                              # Saved Models (.h5 files)
│   └── Model With Transfer Learning.h5
│   └── Model Without Transfer Learning.h5
├── Notebooks/                                           # Jupyter Notebooks for training & evaluation
│   └── Classification With Transfer Learning.ipynb
│   └── Classification Without Transfer Learning.ipynb
├── requirements.txt                                     # Required Dependencies
├── README.md                                            # Project Documentation
└── .gitignore                                           # Git Ignore File

 
 🡅 Back to Top
 

---

MIT License

Copyright (c) 2025 Neelotpal Santra

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

 
 🡅 Back to Top