基于单幅图像的快速去雾算法

L x A 1 erd x
(2)
ࡉᔿ(1)㻛᭩߉Ѫ
H x
F
x
§¨1 ©
L
x
A
· ¸ ¹

L
x
(3)
ਟ㿱, ᡁԜ䴰㾱Ӿമ‫ ۿ‬H (x) ѝՠ䇑ࠪ⧟ຳ‫ݹ‬ L(x) ઼‫ޘ‬ተབྷ≄‫ ݹ‬A , ԕᚒ༽ࠪ␵ᲠⲴᰐ䴮മ‫ۿ‬ F(x) .
2.2 զ䇗⧥ູ‫ݿ‬
ሩᔿ(9)ਣ‫ח‬䘋㹼൷٬└⌒
averagesa
§¨1 ©
M x
Ao
· ¸ ¹
1 averagesa M x
Ao
¦ M y
1
y: x
Ao sa2
(10)
ަѝ, sa 㺘皮൷٬└⌒デਓⲴቪረ, :(x) Ѫ‫ۿ‬㍐ (x) Ⲵ sa u sa 䛫ฏ. ൷٬└⌒ਾⲴ㔃᷌㜭৽ᓄ t(x) Ⲵབྷ 㠤䎻࣯, նоⵏᇎⲴ t(x) ⴨ᐞаᇊⲴ㔍ሩ٬, ഐ↔ሩ
Key Words: Haze removal, Real-time, Physical-based model, Visibility

1 ᕋ䀶
൘㿶仁ⴁ᧗, ሬ㡚᧗ࡦ, ⴞḷ䐏䑚, 䚕ᝏ઼䖖䖶㠚 ѫ傮傦ㅹᢰᵟ亶ฏ, 䜭⎹৺ࡠമ‫ۿ‬༴⨶↕僔. 䘉Ӌᇔ ཆമ‫ۿ‬༴⨶㌫㔏Ⲵᙗ㜭ᶱ᱃ਇࡠᚦ࣓ཙ≄Ⲵᖡ૽. ൘ 䴮, 䵮ㅹᚦ࣓ཙ≄л, ⢙փⲴ৽ሴ‫ݹ‬Պਇࡠབྷ≄ѝᛜ ⎞仇㋂Ⲵ੨᭦઼ᮓሴ, 㠤֯㻛㧧ਆⲴമ‫ۿ‬ਟ㿶ᙗᐞ, 㓶㢲⁑㋺, 㢢ᖙ␑ॆ, ѕ䟽ᖡ૽ਾ㔝മ‫ۿ‬༴⨶㔃᷌. ഐ ↔, മ‫ۿ‬৫䴮㇇⌅Ⲵ⹄ウ䶎ᑨާᴹᇎ䱵᜿ѹ. ᖸཊᇎ 䱵ᓄ⭘㌫㔏ሩᇎᰦᙗ㾱≲ᶱ儈, ᘛ䙏৫䴮㇇⌅Ⲵ⹄ウ ᶱѪ䟽㾱.
[0 255] , ࡉ䴰㾱൘≲ਆ M (x) ѝᡰᴹ‫ݳ‬㍐Ⲵ൷٬ਾ
⴨ᓄ㕙ሿ255‫ ؽ‬.⭡Ҿ mav
ᓖ࠶ᐳᛵߥᴹ‫Ⲵޣ‬٬, ֯ᗇ
ᱟањ઼৏മ‫Ⲵۿ‬ᮤփ⚠
G ާᴹ㠚ࣘ䈳㢲৫䴮ਾമ
‫ۿ‬ᮤփӞᓖⲴ֌⭘. ৏മ‫ۿ‬䎺᳇, mav ٬䎺վ, 䘿ሴ⦷
⴨ᓄ໎བྷ, 䘉֯ᗇ৫䴮ਾⲴമ‫ۿ‬ӞᓖнՊ䗷Ҿ᳇␑,
ᖸཊᯠⲴᘛ䙏৫䴮㇇⌅ⓀҾሩ᮷[7]઼᮷[8]ѝ㇇ ⌅Ⲵ᭩䘋. ᮷[9]‫ࠪ≲ݸ‬䜘࠶‫ۿ‬㍐⛩Ⲵ䘿ሴ⦷٬, ❦ਾ ࡙⭘ཊ‫ݳ‬㓯ᙗഎᖂ≲ࠪᮤᑵമ‫Ⲵۿ‬䘿ሴ⦷, ԕ࣐ᘛ㇇ ⌅䙏ᓖ. ն䘉⿽ᯩ⌅≲ࠪⲴ䘿ሴ⦷䖳Ѫᒣ━, ਚ䘲⭘ ҾᲟ␡ਈॆᒣ㕃Ⲵ⢩ᇊᓄ⭘൪ਸ, ༴⨶Ჟ⢙␡ᓖ੸⧠ ࢗ⛸䐣ਈⲴമ‫ࡉۿ‬᭸᷌䖳ᐞ. ᮷[10]࡙⭘ᘛ䙏ৼ䗩└ ⌒ሩ䘿ሴ⦷䘋㹼㓶ॆ, ԓᴯ᮷[7]ѝⲴ䖟ᣐ↕僔. ᮷[11] ሩമ‫ۿ‬䘋㹼ⲭᒣ㺑ਾ, ≲ਆ仌㢢䙊䚃ᴰሿ٬֌Ѫ⧟ຳ ‫Ⲵݹ‬㋇ՠ䇑, ❦ਾ䘋㹼䗩⭼‫؍‬ᤱ└⌒ᶕ㓶ॆ⧟ຳ‫ݹ‬.
മ‫(ۿ‬Haze image); F ᱟᖵᚒ༽Ⲵᰐ䴮മ‫(ۿ‬Haze-free
image); r 㺘皮བྷ≄ᮓሴ㌫ᮠ; d ԓ㺘Ჟ⢙␡ᓖ; A ᱟ
‫ޘ‬ተབྷ≄‫(ݹ‬Atmospheric light), 䙊ᑨᛵߥл‫ٷ‬䇮Ѫ‫ޘ‬
ተᑨ䟿, оオ䰤඀ḷ (x) ᰐ‫ޣ‬.
ᔿ (1) ѝ , F (x)erd (x) 㻛 〠 Ѫ ⴤ ᧕ 㺠 ߿ (Direct
ީ䭤䈃: ৫䴮, ᇎᰦ, ⢙⨶⁑ර, ਟ㿶ᙗ
Fast Haze Removal from a Single Image
Qian Liu, Maoyin Chen, Donghua Zhou
Department of Automation, Tsinghua University, Beijing 100084, P. R. China E-mail: zdh@mail.tsinghua.edu.cn
2 ᵢᮽⲺ㇍⌋
2.1 䴴ཟഴ‫܅‬䘶ौ⢟⨼⁗ශ
൘䇑㇇ᵪ㿶㿹亶ฏ, 䙊ᑨ⭘л䘠⢙⨶⁑රᶕ᧿䘠
䴮䵮ㅹᚦ࣓ཙ≄ᶑԦሩമ‫ۿ‬䙐ᡀⲴᖡ૽. 䈕⁑රवᤜ
㺠߿⁑ර઼⧟ຳ‫⁑ݹ‬රє䜘࠶, ⁑ර㺘䗮ᔿѪ
H x F x erd x A 1 erd x
(1)
ަѝ, x Ѫമ‫ۿۿ‬㍐Ⲵオ䰤඀ḷ; H ᱟ㿲ሏࡠⲴᴹ䴮
൘สҾঅᑵമ‫Ⲵۿ‬৫䴮㇇⌅ѝ, Tan[5]ԕ໎ᕪമ ‫ۿ‬ሩ∄ᓖѪⴞⲴ, ᨀࠪҶа⿽਼ᰦ䘲⭘Ҿ⚠ᓖ઼ᖙ㢢
↔亩ᐕ֌ᗇࡠഭᇦ973亩ⴞ(亩ⴞ㕆ਧ: 2ቤተ መጻሕፍቲ ባይዱ10CB731800, 2009CB320602) ઼ഭᇦ㠚❦、ᆖส䠁(亩ⴞᢩ߶ਧ: 61210012, 61021063, 61290324)䍴ࣙ.
Abstract: Images of outdoor scenes show poor visibility in presence of haze. This results in poor performance of many outdoor images progressing systems. In this paper, we present a fast haze removal method using a single color or gray level image. Based on the analyses of physical-based model, we estimate airlight and atmospheric light utilizing the simple average filter. Our method is very simple and efficient and achieves real-time computation. Experiments demonstrate that our results achieve as good visibility as a few state-of-art algorithms’.
t x 1 G M ave x
(13)
Ao
ሩ G Ⲵਆ٬䘋㹼࠶᷀ྲл, ᖃ G ਆ٬䖳ሿᰦ, 䘿ሴ⦷
ਆ٬䖳བྷ, ᴰਾᚒ༽ࠪⲴമ‫⮉↻ۿ‬䴮≄䖳ཊ, മ‫ۿ‬ᮤ
փ‫ ;ⲭٿ‬ᖃ G ਆ٬䖳བྷᰦ, 䘿ሴ⦷ਆ٬䖳ሿ, ᴰ㓸ᚒ༽
ࠪⲴമ‫ۿ‬ᮤփ䖳Ѫ᳇␑. ѪҶ䱢→৫䴮ਾമ‫⧠ࠪۿ‬ᮤ
փ⭫䶒䗷Ҿ‫ٿ‬᳇ᡆ䗷Ҿ‫ٿ‬ӞⲴᛵߥ, 䘉䟼䇮㖞
ะӄঋᑻഴ‫Ⲻ܅‬ᘡ䙕ৱ䴴㇍⌋
ࡎٟ, 䱾㤸䬬, ઞђ঄
␵ॾབྷᆖؑ᚟ᆖ䲒㠚ࣘॆ㌫, ेӜ 100084 E-mail: zdh@mail.tsinghua.edu.cn
᪎ 㾷: 䴮䵮ㅹཙ≄Պሬ㠤ᇔཆമ‫ۿ‬ਟ㿶ᙗ䱽վ, 䘋а↕ሬ㠤ᇔཆമ‫ۿ‬༴⨶㌫㔏ᙗ㜭л䱽. ᵜ᮷㔉ࠪа⿽สҾঅ ᑵᖙ㢢ᡆ⚠ᓖമ‫Ⲵۿ‬ᘛ䙏৫䴮㇇⌅. สҾ⢙⨶⁑ර࠶᷀, ӵ࡙⭘൷٬└⌒ሩ⧟ຳ‫ޘ઼ݹ‬ተབྷ≄‫ݹ‬䘋㹼ՠ䇑. ᵜ᮷ ㇇⌅ㆰঅᴹ᭸, 㜭ཏ⭘Ҿᇎᰦ䇑㇇. ᇎ傼㺘᰾, оަԆ㇇⌅⴨∄䖳, 㔃᷌ާᴹᖸྭⲴਟ㿶ᙗ.
മ‫Ⲵۿ‬৫䴮㇇⌅. Fattal൘[6]ѝสҾᲟ⢙㺘䶒䱤ᖡо ӻ䍘䘿ሴ⦷ާᴹተ䜘㔏䇑н⴨‫ޣ‬ᙗ䘉а‫ٷ‬䇮, ᨀࠪа ⿽ਚ㜭༴⨶ᖙ㢢മ‫Ⲵۿ‬㇇⌅. He[7]࡙⭘᳇৏㢢‫ݸ‬傼 ᶕ≲ਆ䘿ሴ⦷, ❦ਾ࡙⭘䖟ᣐ㇇⌅䘋㹼‫↓؞‬, ᴰ㓸ᚒ ༽ࠪᰐ䴮മ‫ۿ‬. к䘠㇇⌅䜭ਆᗇҶᶱྭⲴ৫䴮᭸᷌, ަѝHeⲴ㇇⌅৫䴮᭸᷌ቔѪᱮ㪇. ն䘉Ӌ㇇⌅Ⲵ৫ 䴮䙏ᓖ䜭нཏ⨶ᜣ, ྲ, ᮷[5]༴⨶བྷሿѪ600h400Ⲵ മ‫ۿ‬䴰㾱བྷ㓖5ࡠ7࠶䫏(double processors of Pentium 4 and 1 GB memory), [6]༴⨶аᑵ512h512Ⲵമ‫ۿ‬ᰦ ⭘ᰦ㓖35。(C++, a 1.6 GHz Intel Pentium dual core processor), [7]Ⲵ㇇⌅༴⨶аᑵ600h400‫ۿ‬㍐Ⲵമ‫ۿ‬ 䴰 㾱 བྷ 㓖 10 ࡠ 20 。 (a 3.0 GHz Intel Pentium 4 Processor). Tarel[8]䙊䗷ሩ⧟ຳ‫ݹ‬ਆ٬㤳തⲴ࠶᷀, ᨀ ࠪҶа⿽ᘛ䙏৫䴮㇇⌅, 䈕㇇⌅ሩമ‫ۿ‬䘋㹼ⲭᒣ㺑ਾ, ࡙⭘ཊ⅑ѝ٬└⌒ՠ䇑⧟ຳ‫ݹ‬. ն䈕㇇⌅Ⲵ䘀㇇䙏ᓖ ҏᴹᖵ䘋а↕ᨀ儈.
䈕൷٬└⌒Ⲵ㔃᷌ᕕ㺕ањ‫〫ٿ‬٬, ᒦ䇠
M ave x averagesa M x
(11)
Ҿᱟˈᗇࠪ䘿ሴ⦷Ⲵ㋇ՠ䇑٬
t x 1 M ave x M M ave x
(12)
Ao
Ao
кᔿѝਣ‫ח‬ᴰᵛ亩Ѫᕕ㺕Ⲵ‫〫ٿ‬٬, ф 0 d M d 1.
䇠G 1M , ࡉ
attenuation), 㺘皮Ჟ⢙㺘䶒Ⲵ৽ሴ‫ݹ‬൘ӻ䍘ѝՐ᫝ᰦ
ഐᮓሴㅹ֌⭘㘼㺠߿; A(1 erd (x) ) 㻛〠Ѫ⧟ຳ‫ݹ‬
(Airlight)ᡆབྷ≄‫ݹ‬ᒅ(Atmospheric veil), ᆳሬ㠤മ‫ۿ‬
㢢ᖙ઼ӞᓖⲴ‫〫ٿ‬.
ሩҾᔿ(1), ⭘ L(x) ԓᴯ A(1 erd (x) ) , ণ
M x min H c x
(7)
c^r ,g ,b`
> @ A Ao Ao Ao T
(8)
ࡉᔿ(6)᭩߉Ѫ Ao (1 t(x)) d M (x) , ণ
t x t1 M x
(9)
Ao
⭡ᔿ(4)⸕, t(x) 䲿⵰Ჟ⢙␡ᓖ d Ⲵ໎བྷ㘼߿ሿ.
t(x) ٬䎺ሿ㺘皮‫ۿ‬㍐ (x) ༴Ჟ⢙৽ሴ‫ݹ‬㺠߿Ⲵ〻ᓖ
G U mav
(14)
2013 25th Chinese Control and Decision Conference (CCDC)
3781
ަѝ, U Ѫਟ䈳㢲Ⲵ৲ᮠф 0 d U d 1 mav mav ᱟ M (x) ѝᡰᴹ‫ݳ‬㍐Ⲵ൷٬. ⌘᜿, 䴰㾱㓖ᶏ mav Ⲵ٬ ൘[0 1] ѻ䰤, 㤕䗃‫ޕ‬മ‫Ⲵۿ‬㤳തѪ[0 1] , ࡉ mav ⴤ ᧕ਆ M (x) ѝᡰᴹ‫ݳ‬㍐Ⲵ൷٬; 㤕䗃‫ޕ‬മ‫Ⲵۿ‬㤳തѪ
䎺བྷ. ᖃ t(x) Ѫ0ᰦ, 㺘皮Ჟ⢙৽ሴ‫ݹ‬㺠߿Ѫ0, ↔ᰦ
H (x) A , ᖃ t(x) Ѫ1ᰦ, 㺘皮Ჟ⢙৽ሴ‫ݹ‬ᆼ‫⋑ޘ‬ᴹ
㺠߿, ↔ᰦ H (x) F (x) . 䘿ሴ⦷ t(x) ާᴹ઼Ჟ⢙␡
ᓖ⴨լⲴ䎻࣯, ণ൘␡ᓖ㕃ਈ༴ᒣ━, ൘␡ᓖн䘎㔝
༴䲿ѻ䐣ਈ.
ᵜ᮷䙊䗷ԕлᯩᔿ䘁լՠ䇑䘿ሴ⦷ t(x) .
ᖸཊᆖ㘵ሩ䴮ཙമ‫␵ۿ‬Რॆ㇇⌅䘋㹼Ҷ⹄ウ. ྲ, Narasimhan൘[1]ѝᨀࠪа⿽สҾ⢙⨶⁑රⲴ৫䴮㇇ ⌅, 䈕㇇⌅Ӿ਼а൪Ჟ൘н਼ཙ≄ᶑԦлⲴཊᑵമ‫ۿ‬ ѝ㧧ਆᲟ⢙Ⲵ㔃ᶴؑ᚟; ൘᮷[2]ѝ, ShwartzㅹӪ࡙⭘ ཊᑵ‫ٿ‬ᥟമ‫ۿ‬ᶕ䘋㹼⧟ຳ‫Ⲵݹ‬ՠ䇑; ᮷[3]㔉ࠪа⿽ สҾ⭘ᡧӔӂⲴ㇇⌅; ᮷[4]࡙⭘ᐢ⸕ⲴᲟ⢙␡ᓖ⁑ ර䘋㹼৫䴮. ቭ㇑䘉Ӌ㇇⌅䜭㧧ᗇҶᖸྭⲴ৫䴮᭸᷌, նᱟ䜭䴰㾱䗃‫ޕ‬অᑵമ‫ۿ‬ԕཆⲴؑ᚟, ֯ᗇᇎ䱵ᓄ⭘ ਇ䲀.
⭘ӻ䍘䘿ሴ⦷(Medium transmission) t(x) ᶕᴯ ᦒᤷᮠ㺠߿亩 erd (x) , ণ
t x erd x
(4)
ࡉᔿ(1)ㆰॆѪ
H (x) F (x)t x A 1 t x
(5)
⭡ᔿ(5), ᗇࠪ
A 1 t x d H x
(6)
ሩᔿ(6)є‫≲ח‬ਆйњ仌㢢䙊䚃Ⲵᴰሿ٬, ᒦ䇠
ᵜ᮷ᰘ൘‫؍‬䇱аᇊ৫䴮᭸᷌Ⲵส⹰к, ቭਟ㜭䱽 վ৫䴮㇇⌅Ⲵ༽ᵲᓖˈ֯㇇⌅㜭ཏ┑䏣ᇎᰦᓄ⭘㌫㔏 Ⲵ㾱≲. ᵜ᮷ሶ䙊䗷ሩ䴮ཙമ‫ۿ‬䘰ॆ⢙⨶⁑රⲴ࠶᷀, 䙀↕࠶᷀ࠪ䘿ሴ⦷Ⲵՠ䇑ᯩ⌅, ӵ࡙⭘а⅑ㆰঅⲴ൷ ٬└⌒, 㔉ࠪ⧟ຳ‫ޘ઼ݹ‬ተབྷ≄‫Ⲵݹ‬ՠ䇑ᯩᔿ. фᵜ ㇇⌅㜭ཏ㠚ࣘ䈳㢲৫䴮ਾമ‫Ⲵۿ‬Ӟᓖ, 䱢→৫䴮ਾമ ‫⧠ࠪۿ‬䗷ᓖ‫ٿ‬᳇ᡆ‫ٿ‬ӞⲴᛵߥ.
ª1º
A
1 2
§ ©¨
max
©§¨
max
c^r ,g ,b`
Hc x
¹·¸ max
M ave x
· ¹¸
««1»» «¬1»¼
(19)
䘉⿽ՠ䇑ᯩᔿᰒ䚯‫ݽ‬Ҷⴤ᧕䘹ਆ䗃‫ޕ‬മ‫ۿ‬ѝᴰ Ӟ⛩Ⲵ⚠ᓖ٬֌Ѫབྷ≄‫ݹ‬٬Ⲵнਸ⨶, ৸ެ亮Ҷ㇇⌅ Ⲵᘛ䙏ᙗ.
978-1-4673-5534-6/13/$31.00 c 2013 IEEE
3780
к䘠ᯩ⌅ѝⲴ└⌒ಘӽ❦䖳Ѫ㙇ᰦ, ᘛ䙏ᙗᴹᖵᨀ儈. ᵜ᮷㔉ࠪⲴ㇇⌅ሶӵ࡙⭘а⅑ㆰঅⲴ൷٬└⌒ᶕՠ 䇑⧟ຳ‫ޘ઼ݹ‬ተབྷ≄‫ݹ‬, ާᴹᴤѪᘛ䙏Ⲵᢗ㹼䙏ᓖ. ᵜ᮷㇇⌅ᱟਇࡠ᮷[8]ѝሩ⧟ຳ‫ݹ‬ਆ٬㤳ത࠶᷀Ⲵ੟ ਁ㘼ᨀࠪⲴ, նᵜ᮷㇇⌅о᮷[8]⴨∄൘ᘛ䙏ᙗᯩ䶒 ᴹᶱѪᱮ㪇ᨀ儈, ਼ᰦ, ᵜ㇇⌅Ⲵ৫䴮᭸᷌о᮷[8]Ⲵ ৫䴮᭸᷌⴨ᖃ, ⭊㠣ᴤՈҾ᮷[8]Ⲵ৫䴮᭸᷌.