• Àüü
  • ÀüÀÚ/Àü±â
  • Åë½Å
  • ÄÄÇ»ÅÍ
´Ý±â

»çÀÌÆ®¸Ê

Loading..

Please wait....

±¹³» ÇÐȸÁö

Ȩ Ȩ > ¿¬±¸¹®Çå > ±¹³» ÇÐȸÁö > µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

µ¥ÀÌÅͺ£À̽º ¿¬±¸È¸Áö(SIGDB)

Current Result Document : 11 / 11

ÇѱÛÁ¦¸ñ(Korean Title) CNN±â¹Ý ¼Ò¸® ºÐ·ù ¸ðµ¨ÀÇ °íÁ¶ÆÄ ½ÅÈ£ ÀÎ½Ä °³¼±À» À§ÇÑ ¿Àµð¿À Àüó¸® ¹æ¹ý
¿µ¹®Á¦¸ñ(English Title) Audio Pre-processing Method for Improved Harmonics Signal Recognition in CNN-based Sound Classification Model
ÀúÀÚ(Author) ±¸º»Ã¶   Bon-Cheul Koo   ¹é¹®±â   Moon-Ki Back   À̱Ôö   Kyu-Chul Lee  
¿ø¹®¼ö·Ïó(Citation) VOL 36 NO. 01 PP. 0018 ~ 0038 (2020. 04)
Çѱ۳»¿ë
(Korean Abstract)
ÃÖ±Ù ¼Ò¸® ºÐ·ù¿Í À½¼ºÀÎ½Ä ºÐ¾ß¿¡¼­ CNNÀÌ ¼º°øÀûÀ¸·Î Àû¿ëµÇ¾î º¸´Ù È¿°úÀûÀÎ ¸ðµ¨ÇнÀÀ» À§ÇÑ ÀԷ ǥÇö¿¡ ´ëÇÑ °ü½ÉÀÌ ³ô¾ÆÁ³´Ù. ´ëÇ¥ÀûÀÎ STFT ½ºÆåÆ®·Î±×·¥°ú °°Àº ´Ù¾çÇÑ ½Ã°£-ÁÖÆļö Ç¥ÇöÀ» ÅëÇÑ ¼Ò¸® ½ÅÈ£ÀÇ ½Ã°¢Àû Ç¥ÇöÀº ¿ø½ÅÈ£ÀÇ ½Ã°£¿¡ µû¸¥ ½ºÆåÆ®·³ º¯È­¸¦ °üÂûÇÒ ¼ö ÀÖµµ·Ï µ½´Â´Ù. ±×·±µ¥ ÀϺΠÀÔ·Â ¼Ò¸® ½ÅÈ£¿¡ Á¸ÀçÇÒ ¼ö ÀÖ´Â Áֱ⼺°ú ÁÖ¾îÁø ÇÁ·¹ÀÓ Å©±â¿Í º¸ÆøÀ» °¡Áö´Â STFT¿Í °°Àº ÇÁ·¹ÀÓ ÇÕ¼º°ö ±â¹ýÀÇ Ãâ·Â º¸ÆøÀº ¼­·Î Á¤È®È÷ °°À» ¼ö ¾ø´Ù. ¶§¹®¿¡ µÑ »çÀÌÀÇ Á¤·ÄÀÌ ½Ã°£ÀÇ È帧¿¡ µû¶ó ¾î±ß³ª°Ô µÇ¾î ¿ø½ÅÈ£ÀÇ ±¹¼ÒÀû Áֱ⼺¿¡ ´ëÇÑ À§»ó Á¤º¸°¡ À¯½ÇµÇ´Â Çö»óÀÌ ¹ß»ýÇÑ´Ù. º» ¿¬±¸¿¡¼­´Â ¼Ò¸® ºÐ·ùÀÛ¾÷¿¡ À¯¿ëÇÑ Æ¯¼ºÀ» ÇнÀÇϴµ¥ È¿°úÀûÀÎ ÇÁ·¹ÀÓ ±â¹Ý ½Ã°£-ÁÖÆļö Àüó¸® °úÁ¤¿¡¼­ À¯½ÇµÇ´Â À§»óÁ¤º¸¸¦ º¹¿øÇÏ¿© CNN±â¹Ý ºÐ·ù ¸ðµ¨ÀÇ ÇнÀ µ¥ÀÌÅÍ·Î È°¿ëÇÏ´Â ½ÇÇèÀ» ÁøÇàÇÏ¿´À¸¸ç, ±× °á°ú ƯÁ¤ ¼Ò¸® Áý´Ü¿¡¼­ 10% ÀÌ»óÀÇ ºÐ·ù Á¤È®µµ Áõ°¡¸¦ È®ÀÎÇÏ¿´´Ù.
¿µ¹®³»¿ë
(English Abstract)
Recently, CNN has been successfully applied in the field of sound classification and speech recognition, and interest in input expression for more effective model learning has increased. The visual representation of the audio signal through various time-frequency representations, such as the typical STFT spectrogram, helps to observe the spectral changes over time of the original signal. However, the periodicity that may exist in some input audio signals and the output stride length of a frame-based convolutional method such as STFTs with given frame size and stride, cannot be exactly the same. Therefore, the alignment between the two are shifted over time, resulting in a loass of phase information about the local periodicity of the original signal. In this paper, we conducted an experiment to restore the phase information lost in the frame-based time-frequency pre-processing, which is effective for learning useful properties for sound classification task, and use it as training data for the CNN-based classification model. And as a result, we confirmed an increased in classification accuracy of more than 10% in a specific sound group.
Å°¿öµå(Keyword) ½Åȣ󸮠  ȯ°æÀ½ ¼Ò¸®ºÐ·ù   CNN   µö·¯´×   Signal processing   environmental sound classification   convolutional neural network   deep learning  
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå