TL;DR
This paper introduces the Fourier-Assisted Correlative Denoising (FACD) method to tackle a specific type of low-frequency, periodic stripe noise in infrared linear array images caused by electromechanical interference. FACD effectively removes this noise by leveraging signal correlation and Fourier analysis, achieving a significant improvement in image quality and maintaining real-time performance.
What Problem Does This Paper Address?
Infrared linear array detectors, especially those used in 360-degree scanning applications, are prone to a unique type of periodic stripe noise. This noise is characterized by its vertical orientation, low frequency, and long periodicity, often appearing as subtle, gradually varying stripes in the spatial domain. Unlike high-frequency periodic noise, conventional filtering techniques, such as Gaussian notch filters, are ineffective and can even degrade image quality by removing important low-frequency details. This noise stems from electromechanical interference during image capture and poses a significant challenge for applications requiring high-resolution, clear infrared imagery.
Methods
The core of the proposed method, Fourier-Assisted Correlative Denoising (FACD), lies in its correlation-centric approach. The process begins by capturing a pure background image during the detector's non-uniformity correction phase, which contains the characteristic periodic noise.
- Noise Period Estimation: The algorithm first transforms this background image into the frequency domain to extract a one-dimensional, single-cycle noise signal. By analyzing the spectrum, the dominant noise period (P̂) is estimated.
- Image Segmentation: The infrared image to be denoised is then segmented into blocks. The width of these blocks is set to match the estimated noise period.
- Signal Correlation: For each segmented block, a one-dimensional signal is computed. The algorithm then calculates the cross-correlation between this block's signal and the extracted single-cycle noise signal. This step helps determine the phase offset of the noise within each block.
- Noise Profile Extraction: Based on the correlation results, the noise profile for each segment is ascertained.
- Spatial Domain Denoising: Finally, this noise profile is used to perform denoising in the spatial domain across the entire image.
A key aspect is the robustness of the noise period estimation, which is refined using statistical analysis of pixel values around the median to account for variations and ensure accurate period determination.
Datasets
The paper utilizes two main types of data:
- Pure Background Noise Images: These are captured during the non-uniformity correction phase of the infrared detector, often using a blackbody, to isolate the periodic noise.
- Simulated Infrared Images: These are created by superimposing the captured periodic noise onto clean infrared images. The paper mentions using 20 infrared line-scanning images (1024 × 4096 resolution) for experiments.
- CAMEL Dataset: For simulation experiments, 50 infrared frames from the CAMEL dataset [24] were used. These images, originally 480 × 640, were resized to 1024 × 1365 to match the noise image dimensions.
Key Insights and Ideas
- Low-Frequency Periodic Noise is Different: The paper highlights that the specific periodic noise in infrared linear array detectors is low-frequency and has a long periodicity, making traditional high-frequency noise filters ineffective.
- Correlation is Key: The core idea is that by understanding the pattern of the periodic noise (its period and phase), we can effectively remove it by correlating it with a known noise signature.
- Background Noise is Informative: Capturing a pure background image during non-uniformity correction is crucial as it provides a clean sample of the periodic noise without scene interference.
- Block-Based Processing: Segmenting the image into blocks, with widths matching the noise period, allows for localized correlation and denoising, which is more effective than global filtering.
- Robustness to Scene Content: The FACD method is designed to be less affected by the actual image content, focusing on the predictable nature of the periodic noise.
Key Results
- PSNR Improvement: The FACD method achieved an approximate 2.5 dB increase in Peak Signal-to-Noise Ratio (PSNR) compared to contemporary denoising techniques.
- Qualitative Denoising: Visual results show that FACD effectively removes stripe noise, producing images that are virtually indistinguishable from noise-free images, while preserving image details.
- Real-Time Performance: The algorithm demonstrates excellent real-time capabilities, with significantly lower processing times compared to other methods, making it suitable for high-resolution, real-time scanning applications. For example, for a 1024 × 4096 image, FACD takes approximately 0.07 seconds, while other methods can take minutes.
- Robustness: Experiments showed that the algorithm maintains stable performance even with increasing levels of Gaussian noise, indicating its robustness to varying image complexities.
- Consistent Noise Characteristics: The study found that the periodic noise in the infrared detector maintains consistent characteristics (period and amplitude) over extended operation periods, validating the approach of using a single-period noise signal extracted at power-up.
Limitations and Future Work
- Preliminary Step Requirement: The FACD algorithm requires a preliminary step of collecting pure noise images against a blackbody or baffle. This might be a limitation for imaging systems that lack these capabilities or are used in environments where such collection is difficult (e.g., indoors).
- Future Work:
- Refining the preliminary step to potentially allow for simultaneous acquisition of periodic noise frequencies without compromising real-time operation.
- Exploring the integration of neural network frameworks, acknowledging the current limitations of CNNs in handling high-resolution, real-time denoising and the lack of high-resolution training datasets.
- Addressing deployment complexities and optimizing for real-time processing.
Summary updated