/fisheries-monitoring-deep-learning

Deep learning model to detect and classify fishes

Primary LanguageJupyter Notebook

🐟 Fisheries monitoring using deep convolutional networks

This is the submission to The Nature Conservancy Fisheries Monitoring Kaggle competition, using deep convolutional networks and pre-trained models.

Standard multi output architecture

Abstract

Spanish version

En este trabajo se detallan las soluciones presentadas a la competición The Nature Conservancy Fisheries Monitoring de la plataforma de competiciones predictivas Kaggle. La competición consiste en detectar si aparecen y clasificar diferentes especies de grandes peces en imágenes de barcos pesqueros con el objetivo de analizar el comportamiento de estos.

Las soluciones exploran el uso de arquitecturas de aprendizaje profundo, usando redes neuronales convolucionales preentrenadas para otras competciones de reconocimiento de imágenes mucho más generalistas junto con arquitecturas diseñadas para este problema.

La solución final presentada en este trabajo consigue una medalla de bronce en la competición, clasificándose en el mejor 10 % de los participantes.

English version

In this thesis I outline the submissions presented to The Nature Conservancy Fisheries Monitoring Kaggle competition. The competition consists in detect and classify different kinds of big fishes in fishing boat images, to analyze their behaviour.

The solutions explore the use of deep learning architectures, using pre-trained convolutional neural networks in other competitions, combined with ad-hoc architectures.

The final submission has been awarded with a bronze medal, scoring between the top 10 % submissions.

Structure

All submissions have their own notebook inside the notebooks directory. If you want to reproduce them, install the requirements in the requirements.txt file.

Biography

You can read the entire biography at the end of Master-thesis.pdf file.