Google SVHN Prediction using Custom (Tailored) CNN and Pre-trained models

Street View House Numbers (SVHN) is a real-world image dataset obtained from house numbers in Google Street View images.

Used Pre-trained models

1. LeNet-5 (custom made network - but same features)
2. Vgg16

Key Achievement

• Implemented Custom Deep CNN Model with Keras Tuner and LeNet-5 architecture model for comparison.
• Achieved Model Accuracy:
  • CNN = 81.15% SVHN - only CNN
  • LeNet-5 = 80.30% LeNet-5 vs CNN
  • Vgg16 = ~42% Vgg16

Process Overview

1. Data Conversion and Preprocessing:

   • Converted .mat files to numpy format using scipy.
   • Preprocessed data by reducing channels and applying normalization.

2. Custom CNN Model:

   • Developed a Custom CNN model using Keras Tuner.
   • Total params: 15,698
   • Trainable params: 15,650
   • Non-trainable params: 48 (used for Batch Normalization)

3. LeNet-5 Architecture:

   • Implemented the traditional LeNet-5 alongside CNN to compare the base results of the both.

4. Vgg16

   • Achieved accuracy of ~42%

5. Results:

   • Although tuned-custom made CNN achieved higher results when compared to LeNet, LeNet showed stable increase and decrease in accuracy and loss respectively in comparison to CNN.
   • Vgg16 can be applied to more complex datasets.

6. Dataset Link:

   • Download SVHN Dataset