Deep Learning Based Switching Filter for Impulsive Noise Removal in Color Images

Abstract

Noise reduction is one of the most important and still active research topics in low-level image processing due to its high impact on object detection and scene understanding for computer vision systems. Recently, we observed a substantially increased interest in the application of deep learning algorithms. Many computer vision systems use them, due to their impressive capability of feature extraction and classification. While these methods have also been successfully applied in image denoising, significantly improving its performance, most of the proposed approaches were designed for Gaussian noise suppression. In this paper, we present a switching filtering technique intended for impulsive noise removal using deep learning. In the proposed method, the distorted pixels are detected using a deep neural network architecture and restored with the fast adaptive mean filter. The performed experiments show that the proposed approach is superior to the state-of-the-art filters designed for impulsive noise removal in color digital images.

Keywords: deep learning; deep neural networks; image denoising; image enhancement; impulsive noise; switching filter.

Figure 1

A general scheme of a switching filter. Only the pixels identified as corrupted are being restored and the remaining pixels are retained.

Figure 2

The architecture of the proposed network for impulsive noise detection.

Figure 3

Training of the proposed Impulse Detection Convolutional Neural Network (IDCNN) detector.example

Figure 4

Example images from (a) Berkeley segmentation dataset (BSD500) [61] used for training purposes and (b) our test dataset accessible from [36] as supplementary material.

Figure 5

Exemplary distributions of the IDCNN outputs for the test image contaminated with impulsive noise with intensity $\rho =0.1$ and 0.5. For the contamination $\rho =0.1$, 89.59% of pixels are assigned $\pi <0.1$ and 10.14% reaches a value $\pi >0.9$. For noise intensity $\rho =0.5$, the respective frequencies are 50.20% and 49.44%.

Figure 6

Repeatability of the training procedure on BSD500 dataset.

Figure 7

Impact of the patch size p used in the training procedure on the network’s weighted detection accuracy.

Figure 8

Example images from GoogleV4 and PASCAL VOC2007 datasets [63].

Figure 9

Impact of the type of dataset used in the training and its size on the average wACC of the proposed IDCNN.

Figure 10

Impact of the noise density used during training on the final network performance.

Figure 11

Representative test images from benchmark dataset [36] for which numerical results were calculated.

Figure 12

Box plots presenting the distributions of the obtained results for the analyzed methods using the test dataset [36].

Figure 13

Visual comparison of the filtering efficiency using a part of the PEPPERS image ($\rho =0.4$).

Figure 14

Visualization of the denoising efficiency of the proposed IDCNN_G in comparison to the ideal impulse detector, which correctly identifies all impulses in the analyzed image ($\rho =0.4$). The right column shows the difference between the restored and clean image.

Figure 15

Denoising result of the proposed IDCNN filter on real noisy images: a part of an image of the fresco “The Condemned in Hell” by Luca Signorelli (top) and cDNA image (bottom). Detected impulses are annotated using cyan color.

Figure 16

Diagrams that show what portion of the MAE error was caused by the improper decision of the used filter from classification perspective. These diagrams were obtained for the PEPPERS image ($\rho$ = 0.4).

Figure 17

Execution time dependency of the IDCNN on the number of image pixels N.