Underwater image restoration based on residual dense learning (review)

ABSTRACT

1. efficiency and quality of underwater image restoration is hindered due to degradation factors like scattering,color shift,suspended particles,etc

2. To reduce increase the efficiency and quality this artical introduces the FloodNet which uses residual dense learning for restoring underwater images from wide variety of degraded underwater images.

INTRODUCTION

1. It is important to have degraded free images

2. DNN good at restoring the images

3. DNN learn restoration parameters directly from data distribution which makes them more effective in underwater image restoration

4. CNNs and CycleGAN were used but they were unable to maintain sharpness of image

5. FloodNet is fully CNN that consist of 3 modules

   1. LOW LEVEL FEATURE EXTRACTION (LLFE)
   2. RESIDUAL DENSE BLOCKS (RDBs)
   3. GLOBAL FEATURE FUSION (GFF)

6. LLFE ->

   1. extracts features from degraded underwater image and makes them available to all
      layers of RDB & CFF.
   2. Above is done because it promotes structural similarity.
7. RDB ->
   1. densely connected by skip connections , where output of preceding RDB is available
      to all succeding RDBs.
   2. All featured properties extracted are passed to GFF
8. GFF->
   1. Finally obtains the restored image

src->sciencedirect.com/science/article/abs/pii/S..