Kima Mai Zurfi na Hashen Waje: Hanyar Tushen CNN daga Hotunan LDR Guda

1. Gabatarwa

Dawowa daidai hashen wuri daga hoto guda matsala ce ta asali kuma ba ta da tsari a cikin hangen nesa na kwamfuta, mai mahimmanci ga aikace-aikace kamar haƙiƙanin ƙari (AR), gyaran hoto, da fahimtar wuri. Takarda "Kima Mai Zurfi na Hashen Waje" ta magance wannan ƙalubale musamman ga yanayin waje. Hanyoyin gargajiya sun dogara da bayyanannun alamu kamar inuwa ko suna buƙatar ingantacciyar ƙididdigar lissafi, waɗanda galibi ba su da aminci. Wannan aikin yana ba da shawarar hanyar da ta dogara da bayanai, mafita ta ƙarshe-zuwa-ƙarshe ta amfani da Cibiyoyin Jijiyoyin Convolutional (CNN) don koma bayan sigogin hashen waje mai girma-dynamic range (HDR) kai tsaye daga hoton low dynamic range (LDR) guda.

2. Hanyar Aiki

Babban ƙirƙira ba kawai yana cikin tsarin CNN ba, amma a cikin bututun wayo don ƙirƙirar babban bayanan horo inda gaskiyar hashen HDR ba ta da yawa.

2.1. Ƙirƙirar Bayanai & Dacewar Tsarin Sama

Marubutan sun kewaye rashin bayanan LDR-HDR masu haɗin gwiwa ta hanyar amfani da babban bayanan panorama na waje. Maimakon amfani da panoramas kai tsaye (waɗanda suke LDR), sun dace da ƙirar sama mai ƙarancin girma, tushen zahiri—Tsarin Hošek-Wilkie—zuwa yankunan sama da ake iya gani a kowane panorama. Wannan tsari yana matsawa hadadden hashen siffa zuwa ƙaƙƙarfan sigogi (misali, matsayin rana, turbidity na yanayi). An cire hotuna da aka yanke, iyakacin filin duba daga panoramas, ƙirƙirar babban bayanan (hoton LDR, sigogin sama) biyu don horarwa.

2.2. Tsarin CNN & Horarwa

An horar da CNN don koma daga shigar hoton LDR zuwa sigogin tsarin sama na Hošek-Wilkie. A lokacin gwaji, cibiyar sadarwa tana hasashen waɗannan sigogi don sabon hoto, waɗanda ake amfani da su don sake gina cikakken taswirar yanayi na HDR, yana ba da damar ayyuka kamar shigar abubuwa na zamani masu kama da na gaske (kamar yadda aka nuna a Hoto 1 na PDF).

3. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Tsarin sama na Hošek-Wilkie yana tsakiya. Yana bayyana haske $L(\gamma, \theta)$ a wani batu a cikin sama, idan aka ba da nisan kusurwa daga rana $\gamma$ da kusurwar zenith $\theta$, ta hanyar jerin sharuɗɗan gwaji:

$L(\gamma, \theta) = L_{zenith}(\theta) \cdot \phi(\gamma) \cdot f(\chi, c)$

inda $L_{zenith}$ shine rarraba hasken zenith, $\phi$ shine aikin watsawa, kuma $f$ yana lissafin duhu kusa da rana. CNN yana koyon hasashen sigogin ƙirar (kamar matsayin rana $\theta_s, \phi_s$, turbidity $T$, da sauransu) waɗanda ke rage bambanci tsakanin fitowar ƙirar da kuma saman panorama da aka lura. Aikin asara yayin horarwa yawanci haɗuwa ne na asarar L1/L2 akan vector sigogi da asarar fahimta akan hotunan da aka yi amfani da hasken da aka hasashen.

4. Sakamakon Gwaji & Kima

4.1. Kima na Ƙididdiga

Takardar tana nuna mafi girman aiki idan aka kwatanta da hanyoyin da suka gabata akan bayanan panorama da kuma wani rukuni na taswirorin yanayi na HDR da aka kama. Ma'auni mai yiwuwa sun haɗa da kuskuren kusurwa a cikin matsayin rana da aka hasashen, RMSE akan sigogin tsarin sama, da ma'auni na tushen hoto (kamar SSIM) akan zane-zane na abubuwa da aka haskaka tare da hasken da aka hasashen da na gaskiya.

4.2. Sakamako na Halitta & Shigar Abubuwa na Zamani

Mafi ƙarfi shaida shine na gani. Hanyar tana samar da ingantattun siffofi na HDR daga nau'ikan shigarwar LDR guda daban-daban. Lokacin da aka yi amfani da su don haskaka abubuwa na zamani da aka shigar cikin hoton asali, sakamakon yana nuna daidaitaccen inuwa, inuwa, da haske na musamman waɗanda suka dace da wurin, sun fi dacewa da fasahohin da suka gabata waɗanda galibi suna haifar da haske mara kyau ko mara daidaituwa.

5. Tsarin Bincike: Fahimta ta Asali & Kwararar Hankali

Fahimta ta Asali: Hazakar takardar ita ce mafita mai amfani ga matsalar "Babban Bayanai" a cikin hangen nesa. Maimakon aikin da ba zai yiwu ba na tattara miliyoyin haɗin gwiwar (LDR, binciken HDR) na zahiri, sun haɗa kulawa ta hanyar haɗuwa da babban amma cikakken bayanan panorama na LDR tare da ƙaƙƙarfan ƙirar sama ta zahiri mai banbanta. CNN ba ya koyon fitar da pixels na HDR na sabani; yana koyon zama "mai juyawa mai ƙarfi" don takamaiman ƙirar zahiri da aka fayyace. Wannan aiki ne mafi ƙarfi, mai iya koyo.

Kwararar Hankali: Bututun yana da kyau sosai: 1) Injin Bayanai: Panorama -> Dace da ƙira -> Cire Amfanin ƙasa -> (Hoto, Sigogi) Biyu. 2) Koyo: Horar da CNN akan miliyoyin irin waɗannan biyu. 3) Ƙididdiga: Sabon Hoto -> CNN -> Sigogi -> Tsarin Hošek-Wilkie -> Cikakken Taswirar HDR. Wannan kwararar tana amfani da ƙirar zahiri a matsayin mai matsawa bayanai don horarwa kuma mai zane don aikace-aikace. Yana maimaita nasarar irin waɗannan hanyoyin "koyon tushen ƙira" da aka gani a wasu fagage, kamar amfani da na'urori na kimiyyar lissafi masu banbanta a cikin injiniyoyin mutum-mutumi.

6. Ƙarfi, Kurakurai & Fahimta Mai Aiki

Ƙarfi:

• Girma & Amfani: Hanyar ƙirƙirar bayanai tana da hazaka kuma mai girma, tana mai da albarkatun da ake samu (panoramas) zuwa ingantaccen bayanan horo.
• Yiwuwar Zahiri: Ta hanyar koma bayan sigogin ƙirar zahiri, fitarwa sun fi dacewa kuma ana iya gyara su fiye da fitarwar "akwatin baki" na HDR.
• Sakamako Mai Ƙarfi: Bayyanannen fifikon hanyoyin da suka gabata akan ayyukan zahiri kamar shigar abu shine tabbataccen ingancinsa.

Kurakurai & Iyakoki:

• Dogaro da Ƙira: Hanyar tana da iyaka ta asali ta hanyar bayyanar da ƙirar Hošek-Wilkie. Ba za ta iya dawowa da fasalin hasken da ƙirar ba za ta iya wakilta ba (misali, hadaddun tsarin gajimare, fitattun hanyoyin haske kamar fitilun titi).
• Dogaro da Sama: Yana buƙatar yanki na sama da ake iya gani a cikin hoton shigar. Aiki yana raguwa ko ya gaza don matakin ƙasa ko yanayin cikin gida-waje tare da iyakacin kallon sama.
• Haɗawa zuwa Hashen da ba na Sama ba: Kamar yadda aka lura a cikin PDF, mayar da hankali ne kan hasken sama. Hanyar ba ta ƙirƙira ƙwanƙwasa na biyu ko hasken ƙasa ba, waɗanda zasu iya zama masu mahimmanci.

Fahimta Mai Aiki:

1. Ga Masu Aiki (AR/VR): Wannan mafita ce kusa da shirye-shiryen samarwa don shigar abu na AR na waje. Bututun yana da sauƙin aiwatarwa, kuma dogaro da daidaitaccen ƙirar sama yana sa ya dace da injunan zane na gama-gari (Unity, Unreal).
2. Ga Masu Bincike: Babban ra'ayi—amfani da sauƙaƙan ƙirar gaba mai banbanta don samar da bayanan horo da tsara fitarwar cibiyar sadarwa—yana da ɗaukar hoto sosai. Yi tunani: kima sigogin kayan aiki tare da mai zane mai banbanta kamar Mitsuba, ko sigogin kamara tare da ƙirar rami. Wannan shine mafi dadewa gudunmawar takardar.
3. Matakai na Gaba: Bayyanannen juyin halitta shine haɗa wannan hanyar. Haɗa ƙirar sama ta sigogi tare da ƙaramin CNN na ragowar da ke hasashen "taswirar kuskure" ko ƙarin abubuwan da ba na sigogi ba don sarrafa gajimare da hadaddun hasken birane, matsawa fiye da iyakokin ƙirar yayin riƙe da fa'idodinta.

7. Aikace-aikace na Gaba & Jagororin Bincike

• Haƙiƙanin Ƙari: Sigar lokaci-lokaci, akan na'ura don AR ta wayar hannu, yana ba da damar haɗa abubuwan dijital cikin kowane hoton waje ko rafin bidiyo.
• Daukar Hoto & Bayan Samarwa: Kayan aikin atomatik don ƙwararrun masu daukar hoto da masu shirya fina-finai don daidaita haske tsakanin hotuna ko shigar abubuwan CGI cikin sauƙi.
• Tsarin Mulkin Kai & Injiniyoyin Mutum-Mutumi: Samar da ingantaccen fahimtar hasken wuri don ingantaccen fahimta, musamman don hasashen inuwa da haske.
• Zane na Jijiya & Juyawa Graphics: Yin aiki azaman ingantaccen tsarin kima haske a cikin manyan bututun "rushewar wuri" waɗanda kuma suke kima lissafi da kayan aiki, kama da faɗaɗa aikin daga MIT CSAIL akan rushewar hoto na ciki.
• Yanayi & Ƙirar Muhalli: Bincika manyan tarin hotunan waje na tarihi don kima yanayin yanayi (turbidity, matakan aerosol) akan lokaci.

8. Nassoshi

1. Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., & Lalonde, J. F. (2018). Kima Mai Zurfi na Hashen Waje. A cikin Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
2. Hošek, L., & Wilkie, A. (2012). Ƙirar Nazari don Cikakken Haske na Sama-Dome. ACM Transactions on Graphics (TOG), 31(4), 95.
3. Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hotuna zuwa Hotuna marasa Haɗin gwiwa ta amfani da Cibiyoyin Adawa na Tsarin Zagayowar. A cikin Proceedings of the IEEE International Conference on Computer Vision (ICCV). (CycleGAN, a matsayin misalin koyo ba tare da bayanan haɗin gwiwa ba).
4. Barron, J. T., & Malik, J. (2015). Siffa, Haske, da Haske daga Inuwa. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 37(8), 1670-1687. (Misalin hanyoyin hoto na ciki na gargajiya).
5. MIT Computer Science & Artificial Intelligence Laboratory (CSAIL). Hotunan Ciki a cikin Daji. http://opensurfaces.cs.cornell.edu/intrinsic/ (Misalin bincike da bayanai masu alaƙa).