Multi level wavelet CNN for image restoration

Catalog

1. Abstract interpretation

2. Introduce

3. Related work

4. Method

5. experiment

6. Conclusion

---

1. Abstract interpretation

The motivation of the article ： Ordinary CNN The receptive field is usually expanded at the expense of computational costs . In the process of solving the above problems for the existing extended filtering , The receptive field affected by the grid effect is the love with only checkerboard patterns, such as sparse sampling of images . In this paper, multilevel wavelet is proposed CNN, To better balance the size of receptive field and calculation efficiency .

Network implementation ： Through the revised U-Net framework , Wavelet transform is introduced to reduce the size of the characteristic graph in the shrinking sub network . Besides , Another convolution layer is further used to reduce the channel of the characteristic graph . In the extended subnet , Then deploy the inverse wavelet transform to reconstruct the high-resolution feature map .MWCNN It can also be interpreted as the generalization of extended filtering and secondary sampling , And it can be applied to many image restoration tasks .

effect ： The experimental results clearly show MWCNN In image denoising 、 Single image super-resolution and JPEG Effectiveness of image artifact removal .

2. Introduce

The way of writing ： Image restoration development --->CNN Two explanations ------>CNN Feel the wild problem -----> This paper puts forward the method and the mechanism of solving the receptive field problem -----> Contribution of this paper ：

• A novel MWCNN Model , Used to expand the receptive field , Achieve a better balance between efficiency and recovery performance .
• because DWT Good time-frequency positioning , Promising detail retention .
• In image denoising 、SISR and JPEG The most advanced performance in image deblocking .

3. Related work

A brief review of image denoising 、SISR、JPEG Image artifact removal and other image restoration tasks CNN The development of . say concretely , To expand the receptive field and will DWT Included in the CNN More discussions on the related work of .

• Image denoising
• Single image super-resolution
• JPEG Image artifact removal
• Other recovery tasks （ The main landing points are the existing wavelet transform and CNN Combined method development ）

4. Method

• First of all, we introduce multilevel wavelet packet transform （WPT）.
• Then it introduces the multi-level WPT Driven MWCNN, And describes its network architecture .
• Last , Discussed and analyzed MWCNN Connection with expansion filtering and secondary sampling .
• 

 

5. experiment

Experimental setup （ Data sets 、 Network training details ）

Quantitative and qualitative evaluation （ Image denoising 、 Image super-resolution 、JPEG Artifact removal performance test ）

Ablation Experiment （MWCNN Comparison of variants ）

6. Conclusion

This paper presents a multilevel wavelet for image restoration -CNN（MWC-NN） framework , It consists of a shrinking sub network and an expanding sub network . The shrink subnet consists of multiple levels DWT and CNN Block composition , The extended subnet consists of multiple levels IWT and CNN Block composition . because DWT Reversibility of 、 Frequency and position characteristics ,MWCNN Sub sampling can be performed safely without losing information , And it can effectively restore detailed texture and sharp structure from degraded observation . therefore ,MWCNN Can expand the receptive field , Achieve a better balance between efficiency and performance . A lot of experiments have proved that MWCNN Effectiveness and efficiency in three recovery tasks , That is, image denoising 、SISR and JPEG Compression artifact removal .

In the future work , We're going to expand MWCNN For more general recovery tasks , For example, image deblurring and blind deconvolution . Besides , our MWCNN It can also be used to replace CNN Pooling operations in the architecture , To complete advanced visual tasks such as image classification .