Ukuthuthukisa Ukusetshenziswa Amandla Kwemojuli Yekhamera Yeselula kanye nokusebenza Ngokubona Kwesigcawu se-AI

Enkathini yanamuhla yokuthuthukiswa kobuchwepheshe bokuthwebula izithombe ze-smartphone, ukuqashelwa kwesigcawu se-AI sekuphenduke amandla ayinhloko okuthuthukisa ukusetshenziswa kwamandla nokusebenza kwemojuli yekhamera. Ngokuhlonza ngokuqondile indawo yokudubula kanye namasu okulungisa ngokushintshashintshayo kanye ne-algorithm, abakhiqizi bama-smartphone abakwazi nje ukuthuthukisa ikhwalithi yesithombe kodwa futhi banciphise kakhulu ukusetshenziswa kwamandla, bandise impilo yebhethri. Lesi sihloko sizohlaziya indima ebalulekile yesigcawu se-AI ekwenzeni ngcono umakhalekhukhwini amakhamera kusuka kubukhulu obuthathu: i-hardware architecture, i-software algorithms, kanye ne-system synergy.

Iphrosesa Ephumelelayo: Ukwamukela i-chip ye-AI yokusheshisa (njenge-Rockchip RV1126) inganikeza i-2.0Tops yokusekelwa kwamandla ekhompuyutha i-INT8/INT16 i-computing-precision exubile, iqinisekisa ukusebenza kwesikhathi sangempela ngenkathi kunciphisa ukusetshenziswa kwamandla. Lolu hlobo lokucubungula luhlanganisa injini yombhalo wekhodi wevidiyo esebenza kahle kakhulu, esekela ukucutshungulwa kokusakaza okuningi kwe-4K H.264/H.265, ihlinzeka ngesiqinisekiso samandla ekhompyutha sokuhlaziywa kwesigcawu esiguqukayo.

Ukuphathwa kwamandla kanye nokukhethwa kwe-Hardware: Ukunikezwa kwamandla kagesi we-DCDC kuqala, uma kuqhathaniswa ne-LDO yendabuko, ukusebenza kahle kwamandla kagesi kwe-DCDC kukhuphuka ngaphezu kuka-30%, ikakhulukazi ekunikezeni amandla enzwa yokulungiswa okuphezulu;

Ukujwayela inzwa ye-Dynamic, khetha inzwa efanelekile ngokuya ngobunkimbinkimbi bendawo, njengamamodeli anamandla aphansi ezigcawu ezimile, shintshela kumamojula azwela kakhulu izigcawu eziguqukayo, ukulinganisa ikhwalithi yesithombe nokusetshenziswa kwamandla.

Ukuthuthukiswa Kwesilinganiso Sebhithi Enamandla: Ngokuhlaziywa kwe-AI kobunzima besigcawu kanye nesilinganiso se-ROI (Isifunda Senzalo), amapharamitha wombhalo wekhodi alungiswa ngesikhathi sangempela. Indawo eyinhloko (efana ne-portrait) ikwikhwalithi yesithombe, futhi indawo engeyona ye-ROI igcina ikhwalithi eyisisekelo, nenani elincane lokonga izikhathi ezingaphezu kwezingu-20;Ihlanganiswe nobuchwepheshe be-HEVC, ikhwalithi yesithombe ingcono kunezinhlelo zendabuko ngaphansi kwesilinganiso esincane esifanayo, sehlisa ingcindezi yokudlulisela kanye nesitoreji.

Ukulawulwa Okuhle Kwezinhlaka Ezisebenzayo: ukufanisa ukulungiswa kanye nezinga lozimele, ukushintsha okuzenzakalelayo kokuxazulula (okufana ne-1080P → 720P) ngokuya ngesidingo sendawo, nokulawulwa kozimele ebangeni le-15-30fps, kunciphisa imvamisa yewashi le-VFE; Vala imisebenzi engafuneki, ikhubaza i-ZSL (imodi yokukhipha amandla engu-Zero inganciphisa ukusetshenziswa kwedatha ye-Zero, i-10mA ye-Log yasemuva inganciphisa ukusetshenziswa kwedatha ukungabi namsebenzi.

I-Scene Semantic Segmentation Technology: Ubuchwepheshe bokuhlukaniswa kwesithombe se-AI semantic obusetshenziswa ama-chips e-MediaTek Dimensity bungabolisa isithombe sibe izifunda ezizimele njengezitshalo eziluhlaza esibhakabhakeni esiluhlaza, kanye nokuma ngobude, kuthuthukisa ukugqama, umbala, nokucija. Lobu buchwepheshe, ngokunciphisa izibalo ezingafuneki, kunciphisa isidingo samandla ekhompuyutha ngo-50, futhi busekela ukupakishwa kwama-algorithms amaningi (njengokulandelela okunamandla kwesigcawu sasebusuku).

Ukulungiswa Kwepharamitha Eguquguqukayo: I-Huawei AI Photo Master ifanisa ngokuzenzakalelayo indawo (njengokudla, umbhalo) ngokufunda imikhuba yabasebenzisi, ukwenza ngcono ibhalansi emhlophe kanye nesinxephezelo sokuchayeka. Idatha yokuhlola ibonisa ukuthi ngemva kokunika lo msebenzi amandla, ukucindezelwa kwezithombe ezinhle kakhulu nokugcinwa okumnyama kukhushulwa ngo-40%, futhi ukusetshenziswa kwamandla okuhlola kuqala kwehliswa ngo-15%.

I-Deep Synergy ye-ISP ne-AI: Ama-ISP azithuthukisile (njengochungechunge lwe-Apple, ama-chips e-Huawei Kirin) ayangenelela ekwenziweni kwe-denoising kanye nokwenza kahle kwebanga elishintshashintshayo ekuqaleni kwepayipi lokucabanga ngokuqashelwa kwesimo se-hardware-level, okunciphisa ukusetshenziswa kwamandla ekhompuyutha esigabeni sakamuva. Idatha yokuhlola ibonisa ukuthi ama-ISP ahlanganiswe ne-AI angathuthukisa isivinini sokucubungula semodi yasebusuku izikhathi ezi-2 futhi anciphise amandla ngo-35%.

Ibhalansi Eshisayo Neyokusebenza: Ezimeni zomthwalo ophezulu (ezifana nokuqoshwa kwevidiyo ye-4K), imvamisa ye-CPU/GPU iyalungiswa, ihlanganiswe namasu okulawula okushisayo ahlakaniphile ukuze kugwenywe ukwehla kwe-thermal. Isibonelo, ubuchwepheshe bokuphatha ukushisa kwe-MediaTek's AI bungabikezela iziqongo zokushisa futhi bunciphise ukusetshenziswa kwamandla kwamamojula angabalulekile kusengaphambili.

Ukuhlaziywa Kokuchithwa Kwezinto Zokusetshenziswa Amandla: Ngokuqhathanisa inani lebhentshimakhi lesiqhudelana naye, imodeli "yenkundla yesisekelo sokusetshenziswa kwamandla+isikrini+imojula+i-algorithm" iyasungulwa ukuze kutholwe imojuli yokusetshenziswa kwamandla okungavamile. Isibonelo, imodeli ethile i-algorithm yobuhle ebangele ukwanda okungama-45% ekusetshenzisweni kwamandla okuhlola kuqala ngokubola, okuthuthukiswe kwehle kwaba phakathi kuka-±5% yenani lokumakisha.

Ukulingisa Isimo: Kuhlanganiswe nedatha yokuziphatha komsebenzisi (njengesilinganiso esingu-60% sokuthwebula ividiyo emfushane), izimo zefrikhwensi ephezulu zithuthukiswa ngokukhethekile. Ukuhlola kubonisa ukujwayela okuguquguqukayo kwezinga lozimele kanye nokulungiswa kwezimo zokusakaza bukhoma kunganweba impilo yebhethri ngamahora angu-1.5.

Ukuqashelwa kwesigcawu se-AI kuqhuba inguquko ye-smartphone ukusuka "ku-hardware stacking" kuya "enguqulweni ehlakaniphile". Ngokusungulwa kwe-hardware architecture, ukwenziwa ngcono kwe-algorithm okujulile, kanye nokusebenzisana kwesistimu, ama-smartphones esikhathi esizayo azofinyelela ekugcineni "kokusetshenziswa kwamandla okuphansi kanye nekhwalithi ephezulu yesithombe". Ngokuthuthukiswa okuqhubekayo kwamandla ekhompuyutha ye-AI esemaphethelweni, ubuchwepheshe bokubona indawo buzodlulela ezinkambini ezinjengokuhlanganiswa okubonakalayo nokwakhiwa kabusha kokuxazulula okuphezulu, ukulungisa kabusha umuzwa weselula wokuthwebula.