Pattern Recognition for challenging Computer Vision Applications

Pattern Recognition is the study of how machines can observe the environment, learn to distinguish patterns of interest from their background, and make sound and reasonable decisions about the patterns categories. Nowadays, the application of Pattern Recognition algorithms and techniques is ubiquitous and transversal. With the recent advances in computer vision, we now have the ability to mine such massive visual data to obtain valuable insight about what is happening in the world. The availability of affordable and high resolution sensors (e.g., RGB-D cameras, microphones and scanners) and data sharing have resulted in huge repositories of digitized documents (text, speech, image and video). Starting from such a premise, this thesis addresses the topic of developing next generation Pattern Recognition systems for real applications such as Biology, Retail, Surveillance, Social Media Intelligence and Digital Cultural Heritage. The main goal is to develop computer vision applications in which Pattern Recognition is the key core in their design, starting from general methods, that can be exploited in more fields, and then passing to methods and techniques addressing specific problems. The privileged focus is on up-to-date applications of Pattern Recognition techniques to real-world problems, and on interdisciplinary research, experimental and/or theoretical studies yielding new insights that advance Pattern Recognition methods. The final ambition is to spur new research lines, especially within interdisciplinary research scenarios. Faced with many types of data, such as images, biological data and trajectories, a key difficulty was to nd relevant vectorial representations. While this problem had been often handled in an ad-hoc way by domain experts, it has proved useful to learn these representations directly from data, and Machine Learning algorithms, statistical methods and Deep Learning techniques have been particularly successful. The representations are then based on compositions of simple parameterized processing units, the depth coming from the large number of such compositions. It was desirable to develop new, efficient data representation or feature learning/indexing techniques, which can achieve promising performance in the related tasks. The overarching goal of this work consists of presenting a pipeline to select the model that best explains the given observations; nevertheless, it does not prioritize in memory and time complexity when matching models to observations. For the Pattern Recognition system design, the following steps are performed: data collection, features extraction, tailored learning approach and comparative analysis and assessment. The proposed applications open up a wealth of novel and important opportunities for the machine vision community. The newly dataset collected as well as the complex areas taken into exam, make the research challenging. In fact, it is crucial to evaluate the performance of state of the art methods to demonstrate their strength and weakness and help identify future research for designing more robust algorithms. For comprehensive performance evaluation, it is of great importance developing a library and benchmark to gauge the state of the art because the methods design that are tuned to a specic problem do not work properly on other problems. Furthermore, the dataset selection is needed from different application domains in order to offer the user the opportunity to prove the broad validity of methods. Intensive attention has been drawn to the exploration of tailored learning models and algorithms, and their extension to more application areas. The tailored methods, adopted for the development of the proposed applications, have shown to be capable of extracting complex statistical features and efficiently learning their representations, allowing it to generalize well across a wide variety of computer vision tasks, including image classication, text recognition and so on.

La Pattern Recognition è lo studio di come le macchine osservano l'ambiente, imparano a distinguere i pattern di interesse dal loro background e prendono decisioni valide e ragionevoli sulle categorie di modelli. Oggi l'applicazione degli algoritmi e delle tecniche di Pattern Recognition è trasversale. Con i recenti progressi nella computer vision, abbiamo la capacità di estrarre dati multimediali per ottenere informazioni preziose su ciò che sta accadendo nel mondo. Partendo da questa premessa, questa tesi affronta il tema dello sviluppo di sistemi di Pattern Recognition per applicazioni reali come la biologia, il retail, la sorveglianza, social media intelligence e i beni culturali. L'obiettivo principale è sviluppare applicazioni di computer vision in cui la Pattern Recognition è il nucleo centrale della loro progettazione, a partire dai metodi generali, che possono essere sfruttati in più campi di ricerca, per poi passare a metodi e tecniche che affrontano problemi specifici. Di fronte a molti tipi di dati, come immagini, dati biologici e traiettorie, una difficoltà fondamentale è trovare rappresentazioni vettoriali rilevanti. Per la progettazione del sistema di riconoscimento dei modelli vengono eseguiti i seguenti passaggi: raccolta dati, estrazione delle caratteristiche, approccio di apprendimento personalizzato e analisi e valutazione comparativa. Per una valutazione completa delle prestazioni, è di grande importanza collezionare un dataset specifico perché i metodi di progettazione che sono adattati a un problema non funzionano correttamente su altri tipi di problemi. I metodi su misura, adottati per lo sviluppo delle applicazioni proposte, hanno dimostrato di essere in grado di estrarre caratteristiche statistiche complesse e di imparare in modo efficiente le loro rappresentazioni, permettendogli di generalizzare bene attraverso una vasta gamma di compiti di visione computerizzata.