Deep Learningã§ã©ãã©ã¤ãï¼ãã£ã©ãèå¥ãã
Â
ãã®ã¨ããDeep Learningãç¸å½æµè¡ã£ã¦ããããã§ãã»ã¨ãã©è³ãã¨ããã§è©±é¡ã«ãªã£ã¦ããã®ãè¦ã¾ãã
Deep Learningã¯æ·±å±¤å¦ç¿ã¨ãå¼ã°ãããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã®å±¤ãããã¾ã§ããæ·±ããã¦æ©æ¢°å¦ç¿ãè¡ãææ³ã§ãï¼ã ããã§ãï¼ã
ç»åèªèã³ã³ãã¹ãã§ä»ã®æ¹æ³ã¨æ¯ã¹ã¦é常ã«é«ã精度ã示ãã¦ããã以åã¯äººã®æã§è¡ã£ã¦ããç¹å¾´ã®æ½åºã¾ã§è¡ãã¾ãã
以åã§ããã°è»ãèªèããã«ã¯è»ã¯ã©ã®ãããªç¹å¾´ãæã£ã¦ãããã人ãã¢ãã«åãã¦å
¥åãã¦ããããã§ããããã®ç¹å¾´ãå
¥åç»åã¨ä¸ããããã©ãã«ãããã¥ã¼ã©ã«ãããã¯ã¼ã¯ãæãã¦ããã¾ãã詳ãããã¨ã¯Deep Learningã§æ¤ç´¢ãã¦åºã¦ããè¨äºãã¹ã©ã¤ããåç
§ã®ãã¨ã
Deep Learningèªä½ã¯å®¹æã«å®è£
å¯è½ãªãã®ã§ã¯ãªãããã§ãããå¤ãã®ç 究ã°ã«ã¼ããDeep Learningãè¡ãããã®ã½ããã¦ã§ã¢ããªã¼ãã³ã½ã¼ã¹ã«ãã¦ãããããæ°è»½ã«è©¦ããããã«ãªã£ã¦ãã¾ãã
å
æ¥éå¬ããã87ã§SIG2Dã¨ãããµã¼ã¯ã«ãDeep Learningã®å
¥éè¨äºãè¼ããæ¬*1ãé å¸ãã¦ããããããèªãã§ãªãã¨ãªãããæ°ã«ãªã£ãã®ã§ãã®è¨äºããªãã£ã¦èªåã§ãç»ååé¡å¨ãä½ã£ã¦ã¿ã¾ããã
é¡æã¯ã©ãã©ã¤ãï¼ã®ãã£ã©ã¯ã¿ã¼ã§ããä¸ããããç»åã«å¯¾ããã©ã®Î¼'sã®ã¡ã³ãã¼ãç»åã®ä¸ã«ããããããããã®ç²¾åº¦ã§å¤å®ããã®ãç®æ¨ã§ããã¾ããä»ã®ã¢ãã¡ã®ãã£ã©ãμ'sã¨å¤å¥ããªãããã«ãã¾ãï¼é«ªã®è²ã ãã§å¤å®ãããã¯ã§ããªãï¼ã
Â
å®æåã¯ãããªæãã§ã
Â
Â
Â
ãã¼ã¿åé
æåã«è¡ããªããã°ãªããªãã®ãç»åã®åéã§ããã©ãã©ã¤ãã®ç»åèªä½ã¯ãããã§ãè¦ã¤ããã®ã§ãããDeep Learningã¯ããç¨åº¦ã®ææ°ã®ç»åããªãã¨ç²¾åº¦ãåºãªãå°è±¡ã§ãããããã®å
¥åç»åã®ä¸ã«ã¯èª°ãããã®ãã¨ããã©ãã«ãä»ããªããã°ãªãã¾ããã
Deep Learningã試ãããã ãã«ã©ãã«ãä¸æä¸ææã§ä»ããã®ã¯ãã¾ãã«ãã¼ãã«ãé«ãã®ã§ãä»åã¯æ¢ã«ç»åã«ã¿ã°ãä»ãã¦ãããµã¤ããã¯ãã¼ã«ãã¾ãã
ã¾ããã¿ã°ãä»ãã¦ãã¦ãç»åã«äºäººä»¥ä¸ã®Î¼'sã¡ã³ãã¼ãããå ´åã¯ã©ãã誰ãªã®ãå¦ç¿åã«æããªããã°ãªããªãã®ã§ãä¸äººã®ã¡ã³ãã¼ã ããæããã¦ããç»åã®ã¿ã使ç¨ãã¾ãã
å®éã«äººããã£ã©ã¯ã¿ã¼ãèªèããéã«ã¯é¡ã髪åãä½å½¢ãæè£
ãªã©æ§ã
ãªè¦ç´ ãèæ
®ãã¾ãããé¡ã ããæããã¦ãã¦ãä½è£ã§èª°ãèªèã§ããäºãï¼èªåã®ï¼çµé¨çã«ç¥ããã¦ããã®ã§Deep Learningã«ã¯é¡å¨è¾ºã ããè¦ã¦ãããã¾ãã
ããã§å¥ã®ç»åèªèã¿ã¹ã¯ãå ããã¾ããã2次å
ãã£ã©ã®é¡ã®é åãèªèããªãã¨ç»ååé¡å¨ã«æ¸¡ãã¾ãããããããã£ã¡ã®åé¡ã®æ¹ãé£ããã®ã§ã¯â¦
ãã®åé¡ã«1ããåãçµãããã«ããããªãã®ã§ãããã§ã¯ãã®éã®ããã«ããæ¢åã®ã©ã¤ãã©ãªã使ãã¾ããOpenCV+AnimeFaceã§ãã
nagadomi/lbpcascade_animeface · GitHub
Â
åºæ¬çãªæé ã¨ãã¦ã¯ä»¥ä¸ã®ããã«ãªãã¾ãã
1. ç»åãåé
2. é¡é åãèªèãã¦åãåã
3. ã¿ã°ã¨åãåã£ãç»åãæ¯è¼ãã¦ãã¿ãæé¤
Â
AnimeFaceã®é¡èªèã¯ããªã精度ãè¯ãã®ã§ãããè¨å¤§ãªç»åãå ¥åããã°å½ç¶é¡ã§ãªãé¨åã誤èªèãã¦ãã¾ãã¾ããç»åã«ä»ãã¦ããã¿ã°ã常ã«æ£ãã訳ã§ã¯ããã¾ããããã®ã¾ã¾å ¥åç»åã¨ãã¦ä½¿ãã¨ä½ãå¦ç¿ãããæãã®ã§èª¤èªèãããç»åã¯åãé¤ãã¾ãããã®ç®çã«ã¯ä¸è¬çãªäººéãå©ç¨ãã¾ããå¹³åçãªäººéï¼èªåï¼ã§ããã°1åå½ãã100-200æç¨åº¦ãæ£èª¤ã«åé¡ã§ãã¾ãã1æéããã6000æ以ä¸ãªã®ã§ãããªã«é ãããã§ããªããããéãæã£ã¦äººãéãã¨ã¹ã±ã¼ã«ãã¾ããAmazon Mechanical Turkã¨ãã«æãã¦ã¿ããæ°åããã¾ããã
ãããã¦éã¾ã£ã6600æã®Î¼'sã®ç»åã¨Î¼'sã§ã¯ãªãç»å24000æã§å¦ç¿ãè¡ãã¾ãï¼ããç¨åº¦è² ä¾ãå
¥ããªãã¨Î¼'s以å¤ã®ãã£ã©ãé¡ã§ãããªãç©ä½ãμ'sã¨å¤å®ããå²åãé«ããªãã¾ããããã«ãã¦ãå¤ããããããããªãï¼ãå¦ç¿ç¨ãã¼ã¿ã«ã¯ç»åã®85%ãå©ç¨ããæ®ãã®15%ããã¹ããã¼ã¿ã«ãã¾ãã
Â
Caffeã®å¦ç¿æºå
ä»åã¯SIG2Dã«å¾ã£ã¦Caffeã使ã£ã¦å¦ç¿ãã¾ããCaffeã¯GPUã使ãã¨çéãªãããªã®ã§AWSã®EC2ã§GPUã¤ã³ã¹ã¿ã³ã¹ãåãã¾ããä»åå©ç¨ããç¯å²ã§ã¯ã¹ãããã¤ã³ã¹ã¿ã³ã¹ã¯1æé0.06ãã«ãã¤ã¾ã7ååå¾ã§ããäºãå¤ãã£ãã®ã§ä½¿ãããã¦ãç ´ç£ã¯ããªãã§ãããã
ã»ããã¢ããã®æ¹æ³ã¯SIG2D'14ã«å
¨é¨æ¸ãã¦ããã®ã§ãã¡ããåç
§ãã¦ä¸ãããæ¢ãã°ä¼¼ããããªã»ããã¢ããè¨äºãããã¨æãã¾ãã
å¦ç¿ã®æ¹ã¯Caffeã«å梱ããã¦ããcifar10ã®ãããã¯ã¼ã¯ãå©ç¨ãã¾ãããã©ã¡ã¼ã¿ã¼ãé©å½ã«èª¿æ´ãã¦ç»åãçªã£è¾¼ãã¨ã¢ãã«ãçæãã¦ãããåªããã®ãæ
å ±éç¡ãããã®ç« ãããªãã§ããã
Â
å¦ç¿çµæ
å¦ç¿ä¸ã®éç¨ã¯ãããªæãã§ãã縦軸ãæ£ççã横軸ãå復åæ°ã
Â
Â
4000åã®å復ã§ãã¹ããã¼ã¿ã«å¯¾ãã¦90ï¼ ã®ç²¾åº¦ãåºããå¾ã¯ãã£ããã¨å¦ç¿ãã40000åã®å復ã§96%ã®ç²¾åº¦ã«éãã¦ãã¾ããããããå ã¯è¨ç·´ãã¼ã¿ã¸ã®æ£ççã¯åãã¦ãããã®ã®ããã¹ããã¼ã¿ã¸ã®æ£ççã¯ã©ããå¤ãããªãããã«è¦ããã®ã§70000åã®å復ãçµãã£ãå¦ç¿çµæãå©ç¨ãããã¨ã«ãã¾ãã
ã¨ããã§ããã®ç²¾åº¦ã¯Î¼'sãèå¥ããã®ã«ã¯ãã¾ãããææ¨ã«ãªãã¾ãããμ'sã®ç»åã«æ¯ã¹ã¦ãã以å¤ã®ç»åã®ææ°ãå¤ãããã®ã§ã精度ã®è©ä¾¡ãμ'sã§ã¯ãªãç»åã®ç²¾åº¦ã«å¼ãããããäºã«ãªãã¾ããã ã£ããæåãããã¹ããã¼ã¿ã調æ´ãã¦ãããã¨ããæãã§ãããä»ããé ãã®ã§ããå°ã詳細ã«ç²¾åº¦ã調ã¹ãããã«æ··åè¡åãè¦ã¦ã¿ã¾ãã
Â
Â
縦ãçã®ã¯ã©ã¹ã§æ¨ªãã¢ãã«ã«ããåé¡çµæã§ãã
ç姫ã¡ããã96%ã¨ãã精度ãåºãã¦ããä¸æ¹ãæµ·æªã¡ããã¯85%ã¨è¥å¹²å³ããæãã§ããããã®ä»ã®ã¡ã³ãã¼ã¯90%ã®å¨ãããµãã¤ãã¦ãã¾ããçã®ã¯ã©ã¹ãμ'sã«éå®ããæã®å¹³å精度ã¯90%ã§ãã
大é¨åã®èª¤èªèã¯Î¼'sããã®ä»ã®ã¯ã©ã¹ã¸å¤å®ããäºã«ãã£ã¦èµ·ãã¦ãã¾ããã¡ã³ãã¼éã®èª¤èªèã¯çµµéãããç©ä¹æã¨å¤å®ã2%ãåã¡ãããç©ä¹æããã¨ãã¡ããã¨æ··åããããã1%ç¨åº¦ã§ããã
ä¸æ¹ããã®ä»ããμ'sã¸ã®èª¤èªèã¯å
å·®ã§ç¢æ¾¤ããããã§ããä¸çã®YAZAWAã ããããä»æ¹ãªããã
Â
Webã¢ããª
以ä¸ã®æé ã§å¾ãããç»ååé¡å¨ã使ã£ã¦webã¤ã³ã¿ã¼ãã§ã¹ãä½ãã¾ãã
è¨ç·´ã«ä½¿ç¨ããã®ã¯1人ããåã£ã¦ããªãç»åã§ãããå®ç¨ä¸ã¯Î¼'sã®éååçã«å¯¾ãã¦åé¡å¨ã使ç¨ãããã®ã§é¡èªèãå©ç¨ãã¦å¾ãããããããã®é¡é åãåé¡å¨ã«å
¥åããåé¡çµæãå
ã®ç»åã«ä¹ãã¦è¿ãã¾ãã
ãããOpenCVã¨AnimeFaceã§ãæ軽ã«å®ç¾ã§ãã¾ãã
ãã¡ãããã¢ã®URLã§ãã
ãµã¼ãã¼ã®ãå¼è¶ãã¨å ±ã«æ¶ãã¾ãããã¢ã¼ã¡ã³
Â
ç»åããã©ãã°ï¼ããããããã¨å¤å®çµæã¨ã¹ã³ã¢ãç»åã«ä¹ãã¦è¿ãã¾ããåè§å½¢ã¯å®éã«Deep Learningãè¦ã¦ããé¡é åã表ãã¦ãã¾ãã
Â
çæ³çã«åé¡ãããã¨ãããªç»åãè¿ã£ã¦ãã¾ã
Â
ããããè±é½ã®èªèã¯é£ãã
Â
Â
ãã¹ããã¼ã¿ã§ã¯åé¡å¨ã®ç²¾åº¦ã¯90%ç¨åº¦ã ã£ãã®ã§ã9人ãåã£ã¦ããç»åã ã¨å¤§ä½1人ã¯èª¤èªèãè¡ãããäºã«ãªãã¾ããå®éã«ã¯ç²¾åº¦ã¯ããããè½ã¡ãããã§ãã
è¨ç·´ãã¼ã¿ã«ã¯ä¸äººã§åã£ã¦ããç»åããå©ç¨ãã¦ããããæ£é¢ãåãã¦ããç»åãå¤ããªããã¡ã§ããã2人以ä¸ã®ç»åã«ãªãã¨å¥ã®æ¹åãåãã¦ãããé¡ã®é åã交差ããäºãå¤ããªãããã¾ãå¤å®ã§ããªããªãã¾ããé¡ãå¾ãã¦ããã¨ã¾ãã§ãã¤ãã§ãã
å ãã¦ã¢ãã¡ã®ç»åã¯ã»ã¨ãã©è¨ç·´ã«ä½¿ã£ã¦ããªãã®ã§ãã¡ãã®ç³»çµ±ã®ç»åãè¦æã§ããã
Â
åé¡å¨ãæ··ä¹±ãã¦ãç»åã«ã¯ãããªã®ãããã¾ãã
Â
髪ãä¸ãããçµµéããã¯çããããã¹ã³ã¢ãããªãè½ã¡ãäºéæ²ã¨è¿·ã£ã¦ããã
Â
Â
髪åã®æ å ±ã大é¨åæ¶ãã¦ããã®ã§ç¢æ¾¤ãã©ããå¤æãä»ããªãæ§åãå¸ã¯æ¨ªãåãã¦ããã®ã§é¡èªèã失æãã¦ã¾ããâ¦
Â
Â
ãã¡ãã¯æçµçã«ã¯ä½¿ã£ã¦ããªãã¢ãã«ã®çµæã§ãããããã£ã½ãééãæ¹ããã¦ãã¾ãã
ãã®é«ªåã®æµ·æªã¡ããã¯ãã¾ãã«çãããç¸å½è¦ããã§ããã
Â
Â
確ãã«ç¢æ¾¤ã£ã½ãã
Â
Â
Â
ã©ãè¦ã¦ãçµµéããã
Â
ã©ãè¦ã¦ãçµµéããã ãééãã¦ããã
Â
ãªãã¨ãªãç³ãéãã¦ããã¨èª¤å¤å¥ãå¤ããªãå°è±¡ããã¾ããç³ãã©ããããå½±é¿ãã¦ãããè¦ã¦ã¿ããã®ã§çµµéããã«èµ¤ã®ã«ã©ã¼ã³ã³ã¿ã¯ããä»ãã¦ãããã¨ï¼
Â
2ã¤åã®ç»åã¨ã¯ç³ã®è²ããéããªãã®ã«ããããªããããã¹ã³ã¢ãæ¥è½ãã¦ãã¾ãããç³ã®è²ã£ã½ãä½ããå¼·ãç¹å¾´ã¨ãã¦æãã¦ãã¦ãç³ã®è²ãæ¯è¼çè¿ããã¨ãã¡ããã¨è±é½ã®æåãå ¥ã£ã¦ããããã§ãã
å®éã«Deep Learningã«å ¥åããã®ã¯ç»åã¨èª°ãåã£ã¦ãããã¨ããã©ãã«ã ããªã®ã§ããã®ãããªãã£ã©ã¯ã¿ã¼ãèå¥ããã®ã«æå¹ãªç¹å¾´ã人ã®æãåããèªåã§ç²å¾ãã¦ããããã§ãã
Â
æå¾ã«å¥ã¸ã£ã³ã«ã®ã¢ã¤ãã«ãã¹ã¿ã¼ããä¸æã§ãã
13人ãã¦Î¼'sã£ã½ãã®ã¹ã³ã¢ãåºãã®ã¯åæ©ã®ã¿ãå¤ãããªãä½ã0.13ãªã®ã§ãã®2ã¤ã®ã°ã«ã¼ãã®èå¥ã¯å¤§ä½ã§ãã¦ãã¾ããã¢ã¤ãã«ãã¹ã¿ã¼ã¯å¦ç¿ãã¼ã¿ã«å ¥ã£ã¦ããªãã®ã§ãDeep Learningã¯åæ©ã®äºã¯ç¥ããªãã¨ãæµ·æªã¡ããã£ã½ãã妥å½ã«è©ä¾¡ã§ããããã§ããã
Â
 çµè«
Â
ä»åã¯Deep Learningã®ä¸èº«ã«ã¯å ¨ã触ããã«éãã§ã¿ãã ãã§ãããããã§ãçµæ§ã¾ã¨ããªåé¡å¨ãä½ããããã§ããä»åã¯é¡èªèã«OpenCVã¨AnimeFaceã使ã£ã¦ãã¾ããããã¡ããDeep Learningã«é¢é£ããä½ãã§å®ç¾ã§ãããé¢ç½ããã§ããã
è¨ç·´ãã¼ã¿ã®ä¸è¶³ã«ã¤ãã¦ã¯ã¢ãã¡ã使ããã¨ã§ããç¨åº¦æ¹åã§ãã¾ããã¢ãã¡ã¯å·®åç»åã®å®åº«ãªã®ã§å ¨ãã¬ã¼ã ãç»åã«åãåºããããã§ä½ã£ãåé¡å¨ã§ã¿ã°ãä»ãã¦å度å¦ç¿ãããã¼ãã¹ãã©ããã³ã°ã«ããè¨ç·´ãã¼ã¿ã®ä¸è¶³ã解æ¶ããã精度ã®é«ãåé¡å¨ãä½ãäºãã§ãããã§ãã
ããã§ä½ã£ãã¢ãã«ã¨ã³ã¼ãã¯ãã®ãã¡å ¬éãã¾ãããã¶ãã
Â
Â
*1:Hazuki, Tachibana., Kotori, Otonashi., Yuuma, Kanoue., and Asuna. Proceedings of the 3rd Interdimensional Conference on 2D Information Processing. SIG2D'14