One of the biggest challenges for Convolutional Neural Networks (CNNs), now that they are vastly used in many contexts, is the lack of adequate training sets for robust training sessions and no overfitting. In some cases, it can be difficult to gather enough data for hard classification tasks in the medical field and similar, due to the lack of cases to document or the vast knowledge needed in order to correctly classify the images in a dataset. In the last few years for these reasons data augmentation, rather than other methods, has been under the spotlight mostly thanks to the simplicity and effectiveness that simple methods, such as rotation and flipping for images or random noise added to other sources, have. Here we want to propose various classic and newer methods compared using ResNet50 and MobileNetV2 on thirteen different datasets. Results show that ensembles built combining a variety of the analyzed methods can achieve state of the art performance or even surpass the best-known approaches on various tasks.
Una delle più grandi sfide per le Reti Neurali Convoluzionali, soprattutto ora che vengono utilizzate ampiamente in svariati contesti, è la mancanza di training set adeguati per sessioni di training robuste e meno prone ad overfitting. In certi casi può risultare difficile raccogliere dati a sufficienza per complessi compiti di classificazione in campi quali quello medico e simili a causa dell’assenza di casi di cui sia possibile effettuare un campionamento o delle vaste conoscenze necessarie per la corretta classificazione dei dati ottenuti. Negli ultimi anni proprio per queste ragioni il data augmentation, differentemente da altri metodi, ha avuto particolare successo, principalmente grazie alla semplicità ed efficacia che metodi semplici quali rotazione o specchiamento per le immagini o rumore gaussiano aggiunto ad altre sorgenti riescono ad ottenere. Vogliamo qui proporre vari metodi, classici e di nuova ideazione, comparati utilizzando ResNet50 e MobileNetV2 su tredici diversi dataset. I risultati mostrano che ensemble costruiti combinando i vari metodi analizzati riescono ad ottenere risultati che possono superare perfino i migliori approcci specifici nei diversi dataset.
Tecniche di image augmentation per reti neurali convoluzionali
BRAVIN, RICCARDO
2021/2022
Abstract
One of the biggest challenges for Convolutional Neural Networks (CNNs), now that they are vastly used in many contexts, is the lack of adequate training sets for robust training sessions and no overfitting. In some cases, it can be difficult to gather enough data for hard classification tasks in the medical field and similar, due to the lack of cases to document or the vast knowledge needed in order to correctly classify the images in a dataset. In the last few years for these reasons data augmentation, rather than other methods, has been under the spotlight mostly thanks to the simplicity and effectiveness that simple methods, such as rotation and flipping for images or random noise added to other sources, have. Here we want to propose various classic and newer methods compared using ResNet50 and MobileNetV2 on thirteen different datasets. Results show that ensembles built combining a variety of the analyzed methods can achieve state of the art performance or even surpass the best-known approaches on various tasks.File | Dimensione | Formato | |
---|---|---|---|
Bravin_Riccardo.pdf
accesso aperto
Dimensione
1.34 MB
Formato
Adobe PDF
|
1.34 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/32212