A Comparative Study of Removal Noise from Remote Sensing Image

• Remote sensing analyzes information gathered by a sensor without physically touching the target, area, or phenomenon.Source: §1 · p. 1
• Satellite imagery is used for environmental applications including resource tracking, mapping, crop prediction, urban growth, weather, and flood and fire control.Source: §1 · p. 1
• Space-image applications include recognizing and analyzing objects in images from deep-space probe missions.Source: §1 · p. 1

2. Image Noise

• Noise in satellite images can distort the image and make its content harder to interpret, motivating the use of filters to limit or reduce noise.Source: §2 · p. 1

2.1 Types of Noise

• The section introduces three common noise types, then characterizes RVIN, salt-and-pepper noise, and speckle noise.Source: §2.1 · p. 1, §2.1.1 · p. 2, §2.1.2 · p. 2

2.1.1 Random Variation Impulsive Noise (RVIN)

• RVIN, also called Gaussian or normal noise, occurs randomly as white intensity values and is described by a Gaussian distribution.Source: §2.1.1 · p. 2

2.1.2 Salt & Pepper Noise (SPN)

• Salt-and-pepper noise contains randomly occurring black and white intensity values and is often caused by a noise-image threshold.Source: §2.1.2 · p. 2

2.1.3 Speckle Noise (SPKN)

• Speckle is described as multiplicative noise added to an image, limiting interpretation of optical coherence in remote-sensing images.Source: §2.1.3 · p. 2

3.1 Mean Filter (MF)

• Mean filtering replaces each pixel with the average of its neighborhood, including itself, to smooth intensity variation.Source: §3.1 · p. 2

3.2 Standard Median Filter (SMF)

• The standard median filter applies a two-dimensional mask at each pixel and selects the median brightness among covered pixels.Source: §3.2 · p. 2

3.3 Adaptive Wiener Filter (AWF)

• AWF adapts to image statistics within its window; adaptive filters are described as typically outperforming non-adaptive filters at the cost of added complexity.Source: §3.3 · p. 2

3.4 Gaussian Filter (GF)

• The Gaussian low-pass filter smooths through two-dimensional convolution to remove noise and blur, while its design avoids overshoot to a step-function input.Source: §3.4 · p. 2

3.5 Adaptive Median Filter (AMF)

• AMF varies its surrounding window according to the current-window median and expands the window when that median is an impulse.Source: §3.5 · p. 2

4.1 Testing Procedure

• The Saturn image was tested under SPN, RVIN, and SPKN corruption at probabilities from 10% to 60%, using five filters.The filters were MF, median filter, adaptive filter, Gaussian low-pass filter, and AMF.Source: §4.1 · p. 2

4.2 Simulation Results

• Filtering performance was evaluated with PSNR and MSE, using the noisy and filtered images in the stated metric calculations.M and N denote the image’s horizontal and vertical pixel dimensions; g is the noisy image and f the filtered image.Source: §4.2 · p. 2
• The 60% SPN figure juxtaposes the original and noisy image with filtered outputs labeled AWF, GF, MF, and SMF.The figure caption also identifies AMF in its description of the filtering.Source: Figure 2 · p. 3
• Restoration tables report PSNR and MSE results for SPN, RVIN, and speckle noise.The supplied table titles identify separate PSNR and MSE tables for these noise types.Source: Table 1 · p. 3, Table 2 · p. 3, Table 3 · p. 3

5. Conclusion

• For SPN, AMF is reported to remove noise without leaving blurring, while SMF is described as good below 40% noise density.The conclusion also allows AMF for small-density SPN.Source: §5 · p. 4
• AMF is reported as best for SPN, while MF, AWF, and GF are ranked best in that order for low-density RVIN, where SMF and AMF fail.The conclusion also says AMF can be used for small-density SPN and SMF is good for SPN below 40% density.Source: §5 · p. 4
• For SPKN, MF and AWF are reported as the best filters, with the other filters described as not bad.The conclusion does not quantify the differences among these filters.Source: §5 · p. 4