lazy photographer 黃彥翔 張嫚家 林士涵 黃彥翔...
DESCRIPTION
Four approaches ‧ remove over exposure ‧ remove blur image ‧ remove (denote) duplication ‧ clustering the photos by sceneTRANSCRIPT
Lazy Lazy PhotographerPhotographer
黃彥翔黃彥翔黃彥翔 張嫚家 林士涵黃彥翔黃彥翔 張嫚家黃彥翔 張嫚家黃彥翔 張嫚家黃彥翔 張嫚家黃彥翔 張嫚家黃彥翔 林士涵黃彥翔 林士涵黃彥翔 林士涵黃彥翔 張嫚家 林士涵黃彥翔 張嫚家 林士涵黃彥翔 張嫚家 林士涵黃彥翔 張嫚家 林士涵黃彥翔
Motivation
‧ A lazy photographer‧ after traveling, we want to see photos as soon as possible.
Four approaches‧ remove over exposure ‧ remove blur image‧ remove (denote) duplication‧ clustering the photos by scene
Over exposure‧ use Lab color space‧ split photo to blocks‧ use L value and the distance ab to (0,0)‧ set various thresholds to detect‧ “Correcting Over-Exposure in Photographs “ (must read)
Blur‧ deal with vibration or defocused‧ Use gradient magnitude + gradient direction as a feature vector‧ take 100 blur photos and 100 non-blur to train a model by SVM‧ “Blurred Image Detection and Classification” (must read)
Duplication‧ compare photo with SIFT feature‧ compare with the next n photos
Clustering‧ based on SIFT feature‧ union similar photos‧ set thresholds to detect
Result‧ dataset : 120 photos with
23 over exposure43 blur photos
‧ dataset2 : 60 photos in a single trip
DEMO TIMEDEMO TIME
Result (cont.)‧ Blur detection Recall 83.72% (36/43) Precision 76.60% (36/47)‧ Over exposure Recall 82.60% (19/23) Precision 79.17% (19/24)
Conclusion & future‧ Auto photo adjustment based on our system, fast and convenient‧ replace or correct over exposure parts‧ deblur‧ user-friendly UI
Thank you!Thank you!