A unified codebase, programmed in Python 3, for executing a collection of neural networks being developed at Federal University of Rio Grande by one Computer Engineering student and a Ph.D student:

• Mr. Giovanni Gatti De Giacomo
• MSc Matheus Machado dos Santos

In the codebase there are three supported applications that run on neural networks:

• Semantic segmentation of water and stationary and moving objects in satellite images (U-Net);
• Matching between acoustic and satellite images (DizygoticNet);
• General translation of a large satellite image and acoustic image pair to a single equivalent satellite image (W-Net).

Currently, our codebase depends on the following packages:

• abseil >= 0.7.1
• namegenerator >= 1.0.6
• pandas >= 0.24.2
• scikit-learn >= 0.21.3
• tensorflow >= 2.0.0
• tqdm >= 4.36.1

To choose the neural network and dataset combination to use, the file 'main.py' needs to be edited. If you need to change default configuration, such as shape or batch size, then you will also want to edit 'config.py'.

After setting up the appropriate networks, datasets and configurations, execution is as simple as:

$ python general.py

The command for tensorboard will be automatically generated and printed by the program in the terminal.

Configuration can also be specified via command-line arguments. For a list of all possibilities take a look at 'config.py', an example is provided below:

$ python general.py --batch_size=16 --learning_rate=1e-4

Below is a list of papers we published that are related to this codebase:

• Satellite and Underwater Sonar Image Matching using Deep Learning
• Semantic Segmentation of Static and Dynamic Structures of Marina Satellite Images using Deep Learning
• Sonar-to-Satellite Translation using Deep Learning
• Underwater Sonar and Aerial Images Data Fusion for Robot Localization

• Download ARACATI 2017 Matching Dataset
• Download ARACATI 2017 Segmentation Dataset
• Download ARACATI 2017 Translation Dataset