Real image denoising based on multi-scale residual dense blocks and block connected cascaded u-net

Catalog

One 、 introduction

Two 、 Method

2.1、 Multiscale residual dense network (MRDN)

2.2、MRDB Cascaded U-Net with Block-Connection(MCU-Net)​ edit

2.3、 Noise arrangement

3、 ... and 、 experimental result ​

---

One 、 introduction

        Traditional image denoising research mainly focuses on the removal of sRGB Noise in data , In recent years, research has found that Bayer The denoising ratio in the original data will be sRGB Denoising in data is more effective . therefore , This paper focuses on Bayer The problem of image denoising .

Two 、 Method

2.1、 Multiscale residual dense network (MRDN)

Just look at the picture ：

chart b Medium conv Rate 6 and conv Rate 12 Respectively, the expansion rate is 6 and 12 Of 3×3 Expansion convolution ( Cavity convolution ).

2.2、MRDB Cascaded U-Net with Block-Connection(MCU-Net)

2.3、 Noise arrangement

         Data enhancement is an effective method for neural network to avoid over fitting problem . And then because of Bayer Special properties of data , Traditional data enhancement methods such as image flipping will be due to Bayer Pattern mismatch leads to poor image quality . To solve this problem , This paper introduces a new data arrangement method using real noise from real noisy images . By changing the spatial distribution of real noise , Generate more training samples with real content and noise ：

Pictured 4 Shown , The first step of this method is to groundtruth The image is subtracted from the corresponding noise image , Generate noise image data . For noise data , The noise clustering process divides the data into N A cluster of . then , Arrange randomly in each cluster , Exchange the positions of these noises . After arrangement , Generate a new synthetic noise image , And add it back to the corresponding groundtruth Images , Generate a new synthetic noise image .

3、 ... and 、 experimental result

​​​​​​​