Edge and texture aware image denoising using median noise residue U-net with hand-crafted features
Image denoising is a complex task that always yields an approximated version of the clean image. Unfortunately, the existing works have focussed only on the peak signal to noise ratio (PSNR) metric and have shown no attention to edge features in a reconstructed image. Although fully convolution neur...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
PeerJ Inc.
2025-01-01
|
| Series: | PeerJ Computer Science |
| Subjects: | |
| Online Access: | https://peerj.com/articles/cs-2449.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Image denoising is a complex task that always yields an approximated version of the clean image. Unfortunately, the existing works have focussed only on the peak signal to noise ratio (PSNR) metric and have shown no attention to edge features in a reconstructed image. Although fully convolution neural networks (CNN) are capable of removing the noise using kernel filters and automatic extraction of features, it has failed to reconstruct the images for higher values of noise standard deviation. Additionally, deep learning models require a huge database to learn better from the inputs. This, in turn, increases the computational complexity and memory requirement. Therefore, we propose the Median Noise Residue U-Net (MNRU-Net) with a limited training database without involving image augmentation. In the proposed work, the learning capability of the traditional U-Net model was increased by adding hand-crafted features in the input layers of the U-Net. Further, an approximate version of the noise estimated from the median filter and the gradient information of the image were used to improve the performance of U-Net. Later, the performance of MNRU-Net was evaluated based on PSNR, structural similarity, and figure of merit for different noise standard deviations of 15, 25, and 50 respectively. It is witnessed that the results gained from the suggested work are better than the results yielded by complex denoising models such as the robust deformed denoising CNN (RDDCNN). This work emphasizes that the skip connections along with the hand-crafted features could improve the performance at higher noise levels by using this simple architecture. In addition, the model was found to be less expensive, with low computational complexity. |
|---|---|
| ISSN: | 2376-5992 |