Improved Line Tracing Methods for Removal of Bad Streaks Noise in CCD Line Array Image—A Case Study with GF-1 Images

Bo Wang, Jianwei Bao, Shikui Wang, Houjun Wang, Qinghong ShengSensors 2017 · DOI

TL;DR

This paper introduces an improved method for removing "bad streaks" noise from high-resolution satellite imagery, specifically from the GF-1 satellite. By combining a novel line tracing technique for accurate noise detection with the Lagrange polynomial algorithm for restoration, the method effectively cleans up images while significantly reducing processing time compared to traditional approaches.

What Problem Does This Paper Address?

High-resolution remote sensing images are invaluable for applications like agricultural and water resource surveys. However, these images are often plagued by "bad streaks" noise, which manifests as sudden drops in pixel brightness values, often appearing as thin, elongated lines. This noise renders the images unsuitable for analysis and can even lead to misinterpretations, such as mistaking noise for linear features. Existing methods for noise removal have limitations, including being time-consuming, inaccurate, or not well-suited for the specific characteristics of bad streaks noise in CCD linear array images. This paper aims to address this by developing a more efficient and accurate method for detecting and removing this specific type of noise.

Methods

The proposed method consists of two main steps:

  1. Line Tracing Method for Noise Detection: This novel approach focuses on identifying the abrupt changes in Digital Number (DN) values that characterize bad streaks.
  2. It starts by examining pixel data column by column, looking for significant drops in DN values that exceed a defined threshold ($T_1$). These points are marked as potential noise edges.
  3. To ensure accuracy and avoid false positives, these initial points are verified by comparing neighboring pixels and checking for consistent patterns across a small window.
  4. The method then determines the extent of the bad streak by analyzing the distance between these verified edge points.
  5. Finally, it searches for the end point of the streak by looking for where the DN value changes stabilize, effectively defining a rectangular region of noise.
  6. Lagrange Polynomial Algorithm for Noise Restoration: Once the bad streak regions are identified, this algorithm is used to fill in the missing or corrupted pixel data.
  7. The Lagrange polynomial algorithm constructs a smooth curve that passes through known data points above and below the identified noise region.
  8. This curve is then used to interpolate and estimate the correct DN values for the pixels within the bad streak, effectively replacing the corrupted data with plausible values. The paper highlights that this method is preferred over neighborhood mean methods because it creates a smoother interpolation, better preserving image information and aligning with visual observation habits.

Datasets

The study utilizes Gaofen-1 (GF-1) satellite images for its case study. Specifically, it uses grayscale images with existing bad streaks from GF-1. The paper provides examples of three types of terrain: farmlands, mountains, and islands, to test the algorithm's robustness across different landscapes. No new benchmark datasets were introduced; rather, existing GF-1 imagery with known noise issues was used for evaluation.

Key Insights and Ideas

  • Characterizing Bad Streaks: The paper correctly identifies that bad streaks in CCD linear array images are characterized by discontinuous, parallel, elongated lines with abrupt DN value changes, often dropping to zero. This specific characteristic is key to the proposed detection method.
  • Line Tracing Efficiency: The line tracing method is a significant improvement because it focuses on the mutation points caused by the noise, rather than processing the entire image or large neighborhoods. This targeted approach drastically reduces computational load.
  • Lagrange Polynomial for Smoothness: The choice of the Lagrange polynomial algorithm is crucial. It's not just about filling gaps, but doing so in a way that creates a visually pleasing and information-preserving result, unlike simpler methods that can introduce blockiness or artifacts.
  • Combined Approach Synergy: The power of the method lies in the synergy between accurate detection (line tracing) and sophisticated restoration (Lagrange polynomial). This combination allows for precise targeting of noise and its effective removal without damaging surrounding image data.

Key Results

  • Detection Accuracy: The line tracing method achieved a very low omission ratio (0.6%) and a false detection ratio of 0%, indicating high accuracy in identifying bad streaks.
  • Time Efficiency: The proposed method significantly reduced processing time, saving 55.9% compared to traditional methods. For example, the line tracing method took 26.9 seconds, while the traversal full map method took 61.1 seconds.
  • Restoration Quality: Qualitative observation showed that the bad streaks were effectively removed, making the images suitable for observation. Quantitative analysis using image entropy indicated that the Lagrange polynomial algorithm restored the image entropy closer to the original values compared to the neighborhood mean algorithm, especially in complex terrain.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the accuracy of the algorithm is improved, it doesn't drastically increase in terms of overall accuracy compared to some methods, but rather improves the detection of nearly all points, including subtle changes.
  • The study focuses on GF-1 images; its performance on other satellite sensors with different noise characteristics might require further investigation.
  • The paper mentions that the threshold ($T_1$) for DN value mutation needs to be carefully set, and its optimal value might vary depending on the image.
  • Future Work:
    • The paper implicitly suggests that further research could explore optimizing the threshold selection for different sensor types or noise levels.
  • While not explicitly stated as future work, the success in complex terrains suggests potential for application in more challenging remote sensing scenarios.

Summary updated

Abstract

Remote sensing images could provide us with tremendous quantities of large-scale information. Noise artifacts (stripes), however, made the images inappropriate for vitalization and batch process. An effective restoration method would make images ready for further analysis. In this paper, a new method is proposed to correct the stripes and bad abnormal pixels in charge-coupled device (CCD) linear array images. The method involved a line tracing method, limiting the location of noise to a rectangular region, and corrected abnormal pixels with the Lagrange polynomial algorithm. The proposed detection and restoration method were applied to Gaofen-1 satellite (GF-1) images, and the performance of this method was evaluated by omission ratio and false detection ratio, which reached 0.6% and 0%, respectively. This method saved 55.9% of the time, compared with traditional method.
BibTeX
@inproceedings{wang2017improved,
  author = {Bo Wang and Jianwei Bao and Shikui Wang and Houjun Wang and Qinghong Sheng},
  title = {{Improved Line Tracing Methods for Removal of Bad Streaks Noise in CCD Line Array Image—A Case Study with GF-1 Images}},
  booktitle = {Sensors},
  year = {2017},
  doi = {10.3390/s17040935},
}
Figure