This repository contains code, documentation, and analysis for a project exploring the impact of image permutation on the performance of neural network classifiers. The project investigates how segmenting images into tiles and applying various permutation patterns affects classification accuracy and the learning process of neural networks.
The project delves into the realm of image permutation, dissecting images into tiles and reorganizing them to challenge neural networks in recognizing altered spatial relationships. We explore factors such as tile granularity, permutation complexity, and network architecture depth to understand their effects on classification accuracy. Additionally, the project compares pretrained and custom-designed convolutional neural networks (CNNs), evaluates the efficacy of image augmentation techniques, and examines the interpretability of learned features post-permutation.
- Exploration of tile resolution's impact on classification accuracy.
- Development of a method to rank permutations based on their influence on classification performance.
- Comparison of pretrained and custom-designed CNNs in adapting to permuted images.
- Assessment of image augmentation techniques for enhancing classification of permuted images.
- Examination of the interpretability of learned features post-permutation.
Through rigorous experimentation and analysis, the project aims to provide insights into how image permutation influences neural network performance and identify strategies to enhance network robustness in this context.