About Project

• Small input size: CIFAR-10 images are only 32×32, limiting detail and context
• Overfitting: Models often memorize training data rather than generalize
• Poor robustness: Struggle with varied lighting, backgrounds, blur, or webcam noise
• Deployment hurdles: Real-time webcam predictions suffer from latency and performance issues, especially on low-resource hardware

Problem Statement

• Small input size: CIFAR-10 images are only 32×32, limiting detail and context
• Overfitting: Models often memorize training data rather than generalize
• Poor robustness: Struggle with varied lighting, backgrounds, blur, or webcam noise
• Deployment hurdles: Real-time webcam predictions suffer from latency and performance issues, especially on low-resource hardware

Objective

To build a robust and deployable image classification system capable of performing well in real-world scenarios.

• Convolutional Neural Networks (CNNs): Custom or pre-trained models for baseline performance
• Data Augmentation: Improve generalization with transformations (e.g., flip, zoom, shift, blur)
• Transfer Learning: Utilize ResNet50 with fine-tuning for efficient feature extraction
• Real-Time Prediction: Capture and classify images from webcam stream
• Deployment: Integrated into a web application using Django (optionally with FastAPI for backend inference)

Proposed Solution

• Trained a custom CNN on the CIFAR-10 dataset
• Experimented with:
  • Optimizers: SGD, RMSProp, Adam
  • Batch Sizes: 16, 32, 64
  • Dropout Rates: 0.1, 0.2, 0.3, 0.5
  • Model Variants: Deeper vs Wider architectures
• Applied data augmentation (rotation, zoom, contrast, horizontal flipping)
• Implemented Transfer Learning with ResNet50:
  • Adjusted for CIFAR-10's 32×32 input size
  • Froze initial layers to retain pre-trained weights

• Deployed the best model using a Django-based web interface
• Tested real-time webcam input to perform live image classification

Technologies Used

• Python, TensorFlow/Keras: CNN model development and training
• Pandas, NumPy, Matplotlib: Data preprocessing and visualization
• ResNet50 (Transfer Learning): Used as an ImageNet-pretrained backbone
• Django: Web interface and deployment
• OpenCV: Real-time webcam image capture and handling
• Google Colab (T4 GPU): Model training and experimentation environment

Challenges Faced

• Slow Training: Each training run took over 589 seconds, especially with deeper CNN architectures
• Overfitting: Baseline CNN showed signs of overfitting despite using regularization and dropout
• Transfer Learning Issues: ResNet50 underperformed due to input size mismatch (32×32 vs. 224×224)
• Heavy Deployment: Django web app crashed under memory load during real-time webcam predictions
• Real-time Noise: Webcam inputs had noise, blur, and resolution issues, leading to poor classification results
• Compute Requirements: Fine-tuning ResNet50 required image resizing and high compute resources

Methodology

• Batch Sizes: 16, 32, 64
• Optimizers: SGD, RMSProp, Adam
• Dropout Rates: 0.1, 0.2, 0.3, 0.5
• Model Variants:
  • Deeper: 3 convolutional layers
  • Wider: 2 layers × 64 filters
• Best Configuration: RMSProp + Deeper model + Dropout 0.1
• Best Validation Accuracy: 56.86%

• Applied: Horizontal flip, random rotation, zoom, contrast enhancement
• Outcome: Marginal overfitting reduction, but short-term accuracy dropped

• include_top=False, added custom dense layers with GlobalAveragePooling
• Frozen pretrained layers to preserve learned features
• Validation Accuracy: 23.48% (poor due to CIFAR’s small input size)

• Performance order: Baseline CNN > Augmentation > Transfer Learning
• Deeper CNNs outperformed wider ones
• RMSProp was the most stable optimizer across runs
• Clear overfitting observed via divergence in training and validation curves

Result / Outcome

Configuration	Accuracy (Train / Validation)	Remarks
Best CNN (RMSProp, 3-layer)	78.7% / 73.5%	Best overall performance
CNN + Augmentation	65.0% / 54.8%	Slight overfitting, improved robustness
Transfer Learning (ResNet50)	16.6% / 23.4%	Poor performance due to input size mismatch

• Confusion observed between visually similar classes (e.g., cats vs dogs, trucks vs cars)
• Frogs and airplanes were classified with the least error
• Webcam input: Lower accuracy due to blur, lighting, and low resolution
• Deployment: Initial version built on Django; planned upgrade to Django + FastAPI