Deep learning in detecting defects on printed fabrics

Textile industry is one of the industries that has brought prestige to our territory. In recent years thanks to the advent of synthetic fibers and the explosion of fashion industry it has welcomed numerous innovations in terms of technologies and processes becoming a fast evolving environment. Within textile production industries quality assessment represents one of the most critical processes. Its final aim is to avoid products that do not meet certain requirements be put on the market, as it would have significant costs in terms of money and reputation. Automation in quality assessment of printed fabrics is a particular and yet little explored topic. Based on a real industrial open problem presented by a company, MaGyc, this thesis works on exploring and designing a solution for quality assessment, i.e. printing defects detection, directly on production pipeline. It describes a very particular anomaly detection method working in the particular one-class learning paradigm and dealing with images as rows of 3D time series. Our method is based on deep and machine learning models: a sequence to sequence model and a one-class support vector machine. The former is a deep learning neural model, primarily employed in sentences translation, and thus properly repurposed. Sequence to sequence model aim is to reconstruct inputs to extract useful features out of reconstruction error statistics. The latter is a machine learning model particularly suited for anomaly detection in a one-class classification paradigm. It is responsible to investigate feature vectors built from previous model results and actually classify 3D time series as regular or defective. This solution reaches notable performances in terms of precision and recall, and shows ability to meet industrial requirements of a flexible and zero setup time system.represents one of the most critical processes. Its final aim is to avoid products that doesn’t meet certain requirements be put on the market, as it would have huge costs in terms of money and reputation. Automation in quality assessment of printed fabrics is a particular and yet little explored topic. Based on a real industrial open problem presented by a company, MaGyc, this thesis works on exploring and designing a solution for quality assessment, thus printing defect detection, directly on production pipeline. It describes a very particular anomaly detection method working in the particular one- class learning paradigm and dealing with images as rows of 3D time series. Our method is based on deep and machine learning models: a sequence to sequence model and a one-class support vector machine. The former is a deep learning network, primarily employed in sentences translation, and properly repurposed. Sequence to sequence model aim will be to reconstruct inputs to extract useful features out of reconstruction error statistics. The latter is a machine learning model particularly suited for anomaly detection in a one-class classification paradigm. It will be responsible to investigate feature vectors from previous model and actually classify inputs as regular or defective. This solution reaches notable performances in terms of precision and recall, and shows ability to meet industrial requirements of a flexible and zero setup time system.

L'industria tessile è uno dei settori che ha dato lustro al nostro territorio. Recentemente grazie all'avvento delle fibre sintetiche e l'esplosione dell'industria della moda, ha accolto numerose innovazioni in termini di tecnologie e processi, diventando un industria in forte evoluzione. All'interno del processo di produzione tessile, il controllo qualità rappresenta una delle fasi più critiche. Il suo scopo è quello di evitare che prodotti che non rispettano determinate caratteristiche vengano messi sul mercato, poichè questo avrebbe enormi costi in termini economici e di reputazione. L'automazione nel controllo qualità di tessuti stampati è un argomento particolare e ancora poco esplorato. Partendo da un reale problema industriale presentato da un'azienda, MaGyc, questa tesi vuole esplorare e creare una soluzione per automatizzare la fase di controllo qualità, e quindi l'individuazione di difetti di stampa, direttamente sulla catena di produzione. Descriveremo un sistema di individuzione di difetti basato su un particolare paradigma chiamato one-class learning cheelabora immmagini sotto forma di righe di serie temporali tridimensionali. Il nostro sistema è implementato da modelli di deep e machine learning: un modello sequence to sequence e una one-class support vector machine. Il primo e un modello di deep learning largamente utilizzato nella traduzione di testi e propriamente re-ingegnerizzato. Il compito del modello sequence to sequence sarà di ricostruire l'input per estrarre features utili dall'errore di ricostruzione. Il secondo è un modello di machine learning particolarment adatto per compiti di individuazione di anomalie nei dati nel paradigma one-class learning. Il modello one-class support vector machine sarà responsabile di classificare come difettosi o regolari i vettori di features provenienti dal modello sequence to sequence precedente. La soluzione descritta risulta essere flessibile e rapida da mettere in opera, raggiungendo prestazioni notevoli per quanto riguarda precisione (precision) e recupero (recall).