Cat vs Dog Image Classification using CNN

Content Page

• 1.0 Problem Statement
• 1.1 Implementation
• 1.2 Highlights (Metrics)

---

1.0 Problem Statement

Build a Convolutional Neural Network capable of classifying images of cats and dogs with accuracy greater than 63 %, using a limited training dataset of 2,000 images. The key challenge is preventing overfitting while maintaining strong generalization performance.

---

1.1 Implementation

Quick Overview (2-Minute Read)

Problem Build a CNN to classify cat and dog images with greater than 63 % accuracy using limited training data.

Solution Implemented an end-to-end deep learning pipeline using TensorFlow and Keras, including aggressive data augmentation to combat overfitting.

Impact Delivered a functional computer vision model that reliably classifies pet images, demonstrating applied CNN knowledge and practical ML workflow execution.

Complete work

https://colab.research.google.com/drive/1prIfgPHEy1ppiIdpXizuVUBUy0RdJdKv?usp=sharing

---

Step-by-Step Documentation

Step 1: Environment Setup and Data Loading

• Imported TensorFlow 2.0 and Keras
• Loaded the Cats vs Dogs dataset

Dataset Structure

• Training set: 2,000 images (1,000 cats, 1,000 dogs)
• Validation set: 1,000 images (500 cats, 500 dogs)
• Test set: 50 unlabeled images

---

Step 2: Data Preprocessing

• Created ImageDataGenerator pipelines for training, validation, and test sets
• Normalized pixel values from 0–255 to 0–1 using rescale
• Used flow_from_directory for batch loading
• Disabled shuffling for test data to preserve prediction order

Verification Output

Found 2000 images belonging to 2 classes
Found 1000 images belonging to 2 classes
Found 50 images belonging to 1 class

---

Step 3: Data Augmentation

Issue Identified High overfitting risk due to limited training data

Solution Applied Enhanced the training generator with multiple augmentation techniques:

• Random rotations
• Width and height shifts
• Horizontal flips
• Zoom variations
• Shear transformations

Result Artificial expansion of the training dataset, improving model generalization and validation performance.

---

Step 4: CNN Architecture Design

Model Components

• Multiple Conv2D layers with ReLU activation
• MaxPooling2D layers for spatial downsampling
• Flatten layer
• Fully connected Dense layers
• Final sigmoid output layer for binary classification

Compilation Settings

• Optimizer: Adam or RMSprop
• Loss function: Binary crossentropy
• Metric: Accuracy

---

Step 5: Model Training

• Trained using model.fit with data generators
• Configured epochs and steps per epoch
• Monitored validation performance during training

Outcome The model learned meaningful feature representations from augmented image data.

---

Step 6: Performance Analysis

• Plotted training versus validation accuracy
• Plotted training versus validation loss

Purpose Validated convergence behavior and monitored overfitting trends.

---

Step 7: Model Evaluation

• Generated predictions on 50 unseen test images
• Converted probability outputs into class labels
• Visualized predictions with confidence %ages

Interpretation Each test image was labeled as cat or dog with an associated confidence score.

---

Step 8: Final Validation

• Verified final accuracy exceeded the 63 % requirement
• Project requirements successfully met

---

1.2 Highlights (Metrics)

Resume-Ready Metrics

Technical Results

• Achieved 72 % classification accuracy
• Processed 3,050 images across training, validation, and test sets
• Applied data augmentation to reduce overfitting and improve validation accuracy
• Generated predictions for 50 test images with confidence scores

Model Summary

Total params: 1,061,313 (4.05 MB)
Trainable params: 1,061,313 (4.05 MB)
Non-trainable params: 0 (0.00 B)

---

Technologies Used

• TensorFlow 2.0
• Keras
• Convolutional Neural Networks
• Computer Vision
• Data Augmentation
• Model Evaluation and Visualization

---

Key Skills Demonstrated

1. Problem solving through overfitting detection and mitigation
2. End-to-end machine learning pipeline development
3. Clean, structured experimentation and documentation
4. Results-driven modeling meeting defined accuracy benchmarks