Deep Models Based Recognition of Student Identifiers

Prepoznavanje rukom pisanih znamenaka popularno je od samih ambicioznih početaka strojnog i dubokog učenja. Ovaj rad opisuje sustav koji osim prepoznavanja vrši i lokalizaciju, odnosno vrši detekciju pojedine rukom pisane znamenke. Sustav na ulazu prima sliku rukom pisanih identifikatora koju transformira u binarnu i vrši se podebljanje kontura. Na temelju visine, širine i odnosa između graničnih okvira vrši se spajanje i filtriranje pronađenih komponenti. U fazi postprocesiranja na temelju širine okvira, sigurnosti i entropije odnosno zbunjenosti klasifikatora pronalaze se problematični granični okviri i metodom kliznog prozora pokušavaju se razriješiti isti. Analiza se vrši koristeći dinamičku veličinu prozora te se na temelju zbunjenosti modela vrši podjela okvira. Sustav je ukomponiran u aplikaciju s jednostavnim i intuitivnim grafičkim korisničkim sučeljem. Problematični granični okviri vizualno se ističu kako bi korisnik lakše uočio i ispravio eventualne pogreške sustava. Omogućeno je uređivanje pronađenih graničnih okvira i pripadnih klasifikacija. Sustav tako omogućava generiranje sve većeg skupa za učenje i ponovno treniranje mreže na novoizgrađenom skupu rukom pisanih identifikatora čime klasifikator postaje još robusniji.Recognition of handwritten digits is popular from the ambitious beginnings of the machine and deep learning. This paper describes a system that apart from recognition, also performs localization, or performs detection of handwritten digits. The system at the entrance receives an image of handwritten identifiers that is transformed into binary and contours of the identifiers are bolted. Based on the height, width and relationship between the boundary boxes, merging and filtering of the found boxes are performed. In the postprocessing phase based on the width of the box, classifier’s confidence and perplexity (entropy), problematic boundary boxes are found and the sliding window method is applied to resolve these boxes. The analysis is done using the dynamic size of the sliding window and based on the perplexity of the model the division of the boundary box is performed. The system is wrapped in an application with a simple and intuitive graphical user interface. The problematic bounding boxes are visually highlighted so that the user can easily detect and resolve possible system errors. Through the graphical interface, it is possible to edit the found boundaries and the corresponding classifications. The system thus enables the generation of an even larger learning dataset and retraining the network on a newly built dataset of handwritten identifiers that make the classifier even more robust