Storing and transmitting data has become more and more challenging due to its increasing memory occupation. In the last few decades, several algorithms have been developed to solve the task of data compression, but they are not focused on the specific structure of the data to compress. This is not efficient when - for example - the dataset consists of similar and redundant samples. In situations like this, an AI algorithm might be appropriate to learn and take advantage of these redundancies to perform compression more easily and more efficiently. As for images, Deep Learning and particularly Convolutional Neural Networks are considered to be the state-of-art approach to image classification, because they are able to automatically extract high level features from raw pixels and represent them in a very compact form. There are numerous works in the literature which describe how to compress images with Convolutional Autoencoders, getting great results in terms of performances and quality of reconstructed outputs. With this goal in mind, the purpose of this Master Thesis was to implement an architecture to compress high-resolution images produced by machine vision systems at MERMEC for the detection of anomalies and defects on railway tracks. Not only did the solution have to preserve quality of output images, but it had also to meet some performance constraints, e.g. CPU load, compression ratio, throughput and type of compression (lossy/lossless). Given that a Deep Learning model was the optimal solution to approach the described task due to the high similarity among the images in the dataset (i.e. recurrent patterns), many variations on the theme were explored, trying to find the one which could better combine all project requirements.
Memorizzare e/o trasmettere dati è diventato sempre più complesso a causa della loro crescente occupazione di memoria. Diversi algoritmi sono stati sviluppati per risolvere il task di compressione dei dati, ma nessuno di loro è focalizzato sullo specifico dato da comprimere. Ciò non è efficiente quando, per esempio, il dataset è costituito da campioni molto simili e ridondanti. In questi casi, un algoritmo di AI potrebbe risultare utile per imparare e sfruttare queste ridondanze per comprimere in modo più facile ed efficiente. Per le immagini, il Deep Learning e le Reti Neurali Convoluzionali sono considerate lo standard nella classificazione delle immagini, perché sono in grado di estrarre automaticamente features ad alto livello partendo dai pixel e rappresentare il tutto in una forma compatta. Numerosi lavori in letteratura descrivono come comprimere dataset di immagini con Autoencoder Convoluzionali, ottenendo ottimi risultati in termini di performance e qualità degli output ricostruiti. Con queste basi, l'obiettivo di questa tesi magistrale è stato quello di implementare un'architettura per comprimere immagini ad alta risoluzione acquisite dai sistemi di visione di MERMEC per il riconoscimento di anomalie e difetti sulle tratte ferroviarie. La soluzione proposta non doveva solo preservare la qualità delle immagini, ma anche venire incontro a requisiti di performance, come il carico di CPU, il rapporto di compressione, il throughput e il tipo di compressione (lossy/lossless). Assodato che un modello di Deep Learning fosse la soluzione ottima per il task descritto data la grande similarità tra le immagini presenti nel dataset (con pattern spesso ricorrenti), molte variazioni sul tema sono state esplorate, provando a trovare quella che meglio riuscisse a conciliare tutti i requisiti di progetto.
APPROCCI INNOVATIVI ALLA COMPRESSIONE DI IMMAGINI ATTRAVERSO IL DEEP LEARNING
MALITESTA, DANIELE
2018/2019
Abstract
Storing and transmitting data has become more and more challenging due to its increasing memory occupation. In the last few decades, several algorithms have been developed to solve the task of data compression, but they are not focused on the specific structure of the data to compress. This is not efficient when - for example - the dataset consists of similar and redundant samples. In situations like this, an AI algorithm might be appropriate to learn and take advantage of these redundancies to perform compression more easily and more efficiently. As for images, Deep Learning and particularly Convolutional Neural Networks are considered to be the state-of-art approach to image classification, because they are able to automatically extract high level features from raw pixels and represent them in a very compact form. There are numerous works in the literature which describe how to compress images with Convolutional Autoencoders, getting great results in terms of performances and quality of reconstructed outputs. With this goal in mind, the purpose of this Master Thesis was to implement an architecture to compress high-resolution images produced by machine vision systems at MERMEC for the detection of anomalies and defects on railway tracks. Not only did the solution have to preserve quality of output images, but it had also to meet some performance constraints, e.g. CPU load, compression ratio, throughput and type of compression (lossy/lossless). Given that a Deep Learning model was the optimal solution to approach the described task due to the high similarity among the images in the dataset (i.e. recurrent patterns), many variations on the theme were explored, trying to find the one which could better combine all project requirements.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/103065