Biologically inspired deep learning model for efficient foveal-peripheral vision

Code accompanying the paper "Biologically inspired deep learning model for efficient foveal-peripheral vision" by Hristofor Lukanov, Peter König and Gordon Pipa, Institute of Cognitive science, University of Osnabrück.

Requirements:

Python 3.6
tensorflow-gpu 1.14.0
Keras 2.2.4
numpy 1.15.4
SharedArray 3.2.1
albumentations 0.4.5
opencv-python 4.2.0.32
h5py 2.9.0

Overview

This repository consists of four main folders:

fov - contains tools for creating a sampling matrix, as well as precompiled mappings (forward and reverse) for foveation
preprocessing contains tools for preprocessing the datasets
weights contains weights for models trained on imagenet and core50
utils contains tools for performing foveation and utilities helpful for training and testing

Datasets and data preparation

The dataset needs to be in *.hdf5 format and be named imagenet.hdf5 or core50.hdf5.

The structure of the hdf5 file is the following:

Key train/images includes the train image set with shape (TOTAL_IMAGES,H,W,C) in uint8 datatype, where H=W=256 and C=3. TOTAL_IMAGES is the total number of images in the dataset, H and W stand for respectively height and width, and C stands for color channels.
Key train/labels contains the labels for training data for ILSVRC'10 in onehot format. train/labels10, train/labels50 respectively contain labels with 10 classes or 50 classes for Core50.
Key test/images is identical to train/images but with the test set.
Keys test/labels, test/labels10 and test/labels50 are identical to their training counterparts but for test data.

ILSVRC 2010

The dataset can be downloaded from here. The validation set in ILSVRC 2010 is directly downloaded from the source page.

All images are resized such that the smaller dimension equals 256 pixels and cropped to a 256x256 pixel square images.

Core50

Core50 can be downloaded from here.

The test set of Core50 contains the following folders:

s1: o47, o17, o6, o13, o28, o37, o42, o35
s2: o43, o39, o42, o25, o46, o13, o22, o34
s3: o50, o21, o47, o28, o15, o3, o49, o9
s4: o39, o3, o38, o26, o30, o5, o33, o29
s5: o28, o31, o37, o3, o11, o10, o23
s6: o14, o17, o2, o21, o34, o35, o18, o42
s7: o36, o7, o22, o45, o12, o4, o49
s8: o19, o34, o12, o20, o15, o38
s9: o29, o16, o19, o18, o24, o31
s10: o10, o48, o25, o27, o44, o38, o49
s11: o25, o9, o27, o19, o18, o40, o7

All images are resized to 256x256 pixel size.

How to train and test the model

To train execute:

python main.py --model=[imagenet|core50] --task=train

To test execute:

python main.py --model=[imagenet|core50] --task=test

License

[to be added]

Cite the paper