Insight Evaluation on Traditional and CNN Features

This is the implementation of ICMLSC 2022 paper "Insight Evaluation on Traditional and CNN Features". If the paper and code helps you, we would appreciate your kindly citations of our paper.

1. Dataset

We use two main datasets, namely MNIST and CIFAR-10 with the partitions divided as below:

2. Structures

Folder: Traditional Filters

In this folder, we store the code for 14 traditional filters used to extract features from two mentioned datasets:

1. Prewitt (with X-axis and Y-axis)
2. Sobel filter (with X-axis and Y-axis)
3. Laplacian filter
4. Gaussian Blur filter
5. Local Bimnary Pattern filter
6. Center-Symmetric Local Binary Pattern
7. Spatial Colour Binary Pattern
8. Multi-channel Local Binary Pattern
9. Local Ternary Pattern
10. Completed Local Ternary Pattern
11. Scale Invarian Local Ternary Pattern
12. Scale Invariant Center-Symmetric Local Ternary Pattern
13. Extended Center-Symmetric Local Ternary Pattern
14. Generalized Local Ternary Pattern

Folder: CNN Approach

In this folder, we conducted two practical experiments and store the results evaluated on the test set:

• Experiment 1: Apply a basic CNN network (14 Conv2D Filters, MaxPooling, Flatten and two FC layers).
• Experiment 2: Apply a basic CNN network as above with additional layers (BatchNormalization and Dropout).

Files

• CIFAR - Traditional Filters.ipynb: Store the code for applying traditional filters on CIFAR-10 dataset and create a model prepared for test phase.
• MNIST-Traditional Filters.ipynb: Store the code for applying traditional filters on MNIST dataset and create a model prepared for test phase.
• Evaluation - CIFAR - Traditional Filters.ipynb: Store the results evaluated on the test set when apply the combination between traditional filters on CIFAR-10 dataset.
• Evaluation - MNIST - Traditional Filters.ipynb: Store the results evaluated on the test set when apply the combination between traditional filters on MNIST dataset.

3. Architectures for each experiments

Traditional Filters

MNIST Dataset Here is the general process how we build model which is combined betwwen diverse traditional filters and start performing ten-class classification task on MNIST dataset:

CIFAR-10 Dataset

Here is the general process how we build model which is combined betwwen diverse traditional filters and start performing ten-class classification task on CIFAR-10 dataset:

CNN Filters

Here is the general process how we build model which uses CNN Filters to perform ten-class classification task on both datasets:

4. Final results

For further analysis, please refer to the extended version of our paper in the folder extended-paper.