Ithimba labacwaningi abavela eNyuvesi yasePadua (Italy) kanye naseNyuvesi yaseDelft (Netherlands) bashicilele indlela yokusebenzisa umshini wokufunda ukuze kwakhiwe kabusha ikhodi ye-PIN eqoshiwe eqoshwe ngevidiyo yendawo yokufaka embozwe ngesandla ye-ATM. . Uma ufaka ikhodi ye-PIN enamadijithi angu-4, amathuba okubikezela ikhodi efanele alinganiselwa ku-41%, kucatshangelwa ukuthi kungenzeka ukwenza imizamo emithathu ngaphambi kokuvimbela. Kumakhodi ephinikhodi anezinhlamvu ezi-5, amathuba okuqagela abe ngu-30%. Ukuhlolwa okuhlukile kwenziwa lapho amavolontiya angu-78 azama ukubikezela ikhodi ye-PIN kumavidiyo afanayo arekhodiwe. Kulokhu, amathuba okubikezela okuphumelelayo ayengu-7.92% ngemva kwemizamo emithathu.
Uma umboza iphaneli yedijithali ye-ATM ngesandla sakho, ingxenye yesandla okwenziwa ngayo okokufaka ihlala ingamboziwe, okwanele ukubikezela ukuchofoza ngokushintsha indawo yesandla nokugudluza iminwe engamboziwe ngokuphelele. Lapho kuhlaziywa okokufaka kwedijithi ngayinye, uhlelo lususa okhiye abangakwazi ukucindezelwa kucatshangelwa indawo yesandla esisibekelayo, futhi lubala izinketho ezingase zibe khona zokucindezela ngokusekelwe endaweni yesandla sokucindezela esihlobene nendawo yezinkinobho. . Ukuze kwandiswe amathuba okuthola okokufaka, umsindo wezinkinobho ungarekhodwa ngokungeziwe, ohluke kancane kukhiye ngamunye.
Ukuhlolwa kusebenzise isistimu yokufunda yomshini esuselwe ekusetshenzisweni kwenethiwekhi ye-neural ye-convolutional (CNN) kanye nenethiwekhi ye-neural ephindelelayo esekelwe kuzakhiwo ze-LSTM (Inkumbulo Yesikhathi Esifushane Eside). Inethiwekhi ye-CNN ibe nesibopho sokukhipha idatha yendawo yozimele ngamunye, futhi inethiwekhi ye-LSTM yasebenzisa le datha ukuze ikhiphe amaphethini ashintsha isikhathi. Imodeli yaqeqeshwa kumavidiyo abantu abahlukene abangu-58 abafaka amakhodi e-PIN besebenzisa izindlela zekhava zokufaka ezikhethwe umhlanganyeli (umhlanganyeli ngamunye ufake amakhodi ahlukene ayi-100, okungukuthi, izibonelo zokufaka ezingu-5800 zisetshenziswe ekuqeqeshweni). Phakathi nokuqeqeshwa, kwavezwa ukuthi abasebenzisi abaningi basebenzisa enye yezindlela ezintathu eziyinhloko zokumboza okokufaka.
Ukuze uqeqeshe imodeli yokufunda yomshini, kusetshenziswe iseva esekelwe kuphrosesa ye-Xeon E5-2670 eno-128 GB we-RAM namakhadi e-Tesla K20m amathathu anememori engu-5GB ngalinye. Ingxenye yesoftware ibhalwe ngePython kusetshenziswa umtapo wezincwadi wakwaKeras kanye nenkundla yeTensorflow. Njengoba amaphaneli okokufaka e-ATM ehlukile futhi umphumela wokubikezela uncike ezicini ezifana nosayizi wokhiye kanye ne-topology, ukuqeqeshwa okuhlukene kuyadingeka ohlotsheni ngalunye lwephaneli.
Njengezinyathelo zokuvikela indlela yokuhlasela ehlongozwayo, kunconywa, uma kungenzeka, ukusebenzisa amakhodi e-PIN anezinombolo ezi-5 esikhundleni sika-4, futhi uzame ukuvala isikhala esiningi sokufaka ngesandla sakho (indlela ihlala isebenza kahle uma cishe u-75% wendawo yokufaka imbozwe ngesandla sakho). Abakhiqizi be-ATM batuswa ukuthi basebenzise izikrini ezikhethekile zokuzivikela ezifihla okokufaka, kanye nokungewona ama-mechanical, kodwa amaphaneli okufakwayo okuthinta, indawo yezinombolo ezishintsha ngokungahleliwe.
Source: opennet.ru