| Titre : | Fish Recognition Using Deep Learning | | Type de document : | projet fin études | | Auteurs : | Bouzaouit Adnane, Auteur | | Langues : | Français (fre) | | Catégories : | Ingénierie Finance et la Gestion des Risques (IFGR)
| | Index. décimale : | mast 114/18 | | Résumé : | A great deal of researches has been conducted around the computer vision
eld yet the undersea studies are still far away from being satisfying, nowadays,
underwater object recognition is in great demand, while there are not much solid
systems for this task. For this matter, we present in this report a sh recognition
framework based on a foreground extraction to remove the background in
uence
, followed by a data augmentation, in an attempt to balance the number of
observations for each class , then a convolutional neural network architecture for
the recognition of 23 di erent sh species, in a real-world sh recognition dataset
lmed in the wild unrestricted natural environment we achieved an accuracy of
98.54%. |
Fish Recognition Using Deep Learning [projet fin études] / Bouzaouit Adnane, Auteur . - [s.d.]. Langues : Français ( fre) | Catégories : | Ingénierie Finance et la Gestion des Risques (IFGR)
| | Index. décimale : | mast 114/18 | | Résumé : | A great deal of researches has been conducted around the computer vision
eld yet the undersea studies are still far away from being satisfying, nowadays,
underwater object recognition is in great demand, while there are not much solid
systems for this task. For this matter, we present in this report a sh recognition
framework based on a foreground extraction to remove the background in
uence
, followed by a data augmentation, in an attempt to balance the number of
observations for each class , then a convolutional neural network architecture for
the recognition of 23 di erent sh species, in a real-world sh recognition dataset
lmed in the wild unrestricted natural environment we achieved an accuracy of
98.54%. |
|
Faire une suggestion Affiner la recherche
Fish Recognition Using Deep Learning / Bouzaouit Adnane
