quocnh/Blurry_Detection_DeepLearning

Description

There were lots of photos with excessive amounts of blurring, which are taken from users. This project will classify good/bad images by using deep learning.

Table of Contents

• Dataset description
  • File and columns description
  • Folder structure
• Evaluation
• Submission format

Dataset description

File and columns description

• Some researches indicate that the top four main cause of blurry photos are:

  • Out Of Focus [img]
  • The subject moves while the shutter is open [img]
  • The camera moves while the shutter is open [img]
  • Depth Of Field is too shallow [img] 

• We define train & test datasets followed by four metrics above (1692 OK and 13798 NG images):

  • OK: clear focused-objects, not blurry, good contract
  • NG: blurry (out_of_focus, object_move, camera_move, depth_of_field)

Dataset Link: https://goo.gl/yp1XS6

Folder structure

Dataset
        |__train
        |    |__OK (1692)
        |    |__NG (13798)
        |       |__camera_move
        |       |__depth_of_field_shallow
        |       |__object_move
        |       |__out_of_focus
        |
        |__test
            |__OK (49)
            |__NG (49)
                |__camera_move
                |__depth_of_field_shallow
                |__object_move
                |__out_of_focus

Evaluation

Metrics

* binary_crossentropy

Submission format

Output

• {'OK': 0, 'NG': 1}
• Check the result-logs in the result_logs.md.

Changes log

2018/05/20 10:00: Added more type of images into train dataset.

2018/05/15 14:00: Removed motion and out-of-focus images from NG dataset. (because this type of images does not macth with our dataset standard)

2018/05/15 10:00: Init document