This repository contains the code for a simple experiment to see what influence the flattening method in a flatten layer has on image classification, when using 1D convolutions over the flattened image.
Given an image of size (n, n, 3) where n = 2^k for some k in N, k > 1 the HCFlatten layer defines a mapping
H_n: N**2 -> N such that:
A---B C---D
| |
E---F G---H
HCFlatten | | ---> A-B-F-E-I-M-N-J-K-O-P-L-H-G-C-D
I J---K L
| | | |
M---N O---P
A---B---C---D
┌-----------┘
E---F---G---H
Flatten ┌-----------┘ ---> A-B-C-D-E-F-G-H-I-J-K-L-M-N-O-P
I---J---K---L
┌-----------┘
M---N---O---P
Three different models are defined in benchmarks.py. The models are trained on the NWPU-RESISC45 dataset
[1]. Due to memory limitations the training data was limited to the first 60%, validation data to the
last 5%. Image augmentation is performed on the training data. The following three models are compared.
- 2d ConvNet (1,677,133 Parameters)
- 1D ConvNet with simple flatten (1,689,413 Parameters)
- 1D ConvNet with HCFlatten (1,689,413 Parameters)
Images are all 512x512 pixels with three color channels.
The 2D ConvNet unsurprisingly outperforms both 1D variants.
The hilbert curve based flattening layer only marginally outperformed the simple flattening approach.
| Architecture | Val Loss | Val Accuracy |
|---|---|---|
| 2D Convnet | 1.1674 | 0.6711 |
| 1D Simple | 1.3788 | 0.6178 |
| 1D HCFlatten | 1.2528 | 0.6463 |
[1] G. Cheng, J. Han, X. Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE.