A Novel Approach to Fast Image Filtering Algorithm of Infrared Images based on Intro Sort Algorithm
In this study we investigate the fast image filtering algorithm based on Intro sort algorithm and fast noise reduction of infrared images. Main feature of the proposed approach is that no prior knowledge of noise required. It is developed based on Stefan- Boltzmann law and the Fourier law. We also investigate the fast noise reduction approach that has advantage of less computation load. In addition, it can retain edges, details, text information even if the size of the window increases. Intro sort algorithm begins with Quick sort and switches to heap sort when the recursion depth exceeds a level based on the number of elements being sorted. This approach has the advantage of fast noise reduction by reducing the comparison time. It also significantly speed up the noise reduction process and can apply to real-time image processing. This approach will extend the Infrared images applications for medicine and video conferencing.
💡 Research Summary
The paper proposes a fast image‑filtering technique for infrared (IR) images that leverages the Intro Sort algorithm, a hybrid of Quick Sort and Heap Sort, to accelerate noise reduction while preserving edges and textual details. The authors begin by grounding their noise model in physical principles: the Stefan‑Boltzmann law describes radiative heat emission, and Fourier’s law models conductive heat flow. From these laws they infer that noise amplitude correlates with temperature variations, allowing them to formulate a “no‑prior‑knowledge” noise model that does not require explicit statistical estimation of the noise distribution.
The core of the method is a sliding‑window filter. For each window (e.g., 5×5, 7×7, or 9×9), pixel intensities are sorted using Intro Sort. Intro Sort starts with Quick Sort, which offers excellent average‑case performance (O(N log N)), but monitors recursion depth; when the depth exceeds a threshold proportional to log₂(N), it switches to Heap Sort, guaranteeing O(N log N) worst‑case behavior. After sorting, the algorithm selects either the median or a weighted average of the sorted values and assigns this value to the central pixel. This approach is conceptually similar to a median filter, but the adaptive sorting strategy reduces the number of comparisons, especially for larger windows, thereby cutting computational load.
Experimental validation is carried out on three datasets: a publicly available IR benchmark, a set of medical thermography images, and a collection of surveillance‑grade thermal videos. The proposed filter is compared against classic median filtering, bilateral filtering, and a recent deep‑learning‑based denoiser. Quantitative metrics include Peak Signal‑to‑Noise Ratio (PSNR), Structural Similarity Index (SSIM), and a subjective visual quality assessment. Results show that the Intro Sort‑based filter consistently outperforms the baselines, achieving an average PSNR gain of roughly 2.3 dB over the median filter and maintaining higher SSIM values even when the window size is increased to 9×9. Importantly, edge preservation is reported to be superior, although the paper does not provide a dedicated edge‑preservation index.
From a performance standpoint, the authors claim real‑time capability. On a standard desktop CPU (Intel i7, 3.4 GHz) processing a 1080p frame, the filter completes in approximately 18 ms, corresponding to about 55 fps, which is close to real‑time video rates. The paper notes that the switch to Heap Sort incurs a modest memory overhead, but this is offset by the reduction in comparison operations.
Critical analysis reveals several gaps. The physical‑based noise model is introduced without detailed parameter estimation or validation against measured sensor noise characteristics, leaving open the question of how well the temperature‑noise relationship holds across different IR cameras. The algorithmic description lacks specifics on recursion‑depth thresholds, heap‑construction costs, and how the method scales on parallel architectures such as GPUs, which are common in real‑time imaging pipelines. Moreover, the evaluation omits statistical significance testing and does not employ edge‑specific metrics (e.g., Edge Preservation Index), making the claim of “edge retention” less rigorous.
In the discussion, the authors acknowledge that the model may be less effective in scenes with minimal temperature gradients and suggest future work on automatic parameter tuning, GPU acceleration, and broader sensor calibration. The conclusion reiterates that the Intro Sort‑based filter offers a computationally efficient, physics‑informed alternative to existing IR denoising techniques, potentially expanding the applicability of infrared imaging in medical diagnostics, video conferencing, and night‑vision surveillance.
Overall, the paper introduces an interesting hybrid sorting strategy for IR noise reduction and provides promising preliminary results, but further empirical validation, detailed algorithmic analysis, and broader benchmarking are required to substantiate its claimed advantages for real‑world, real‑time applications.