SPICE: Simple and Practical Image Clarification and Enhancement

TL;DR

SPICE addresses challenging image enhancement tasks—low-light and hazy conditions—using a light-weight, non-deep-learning approach based on reverse filtering. The method inverts a distortion filter (with a haze-simulating or darkening filter) by transforming the image, e.g., $v -> v / (f(v)^p + r)$ with $0 < p <= 1$, where $f$ is a smoothing filter such as the guided filter, and extends to dehazing via AMF inversion using fixed-point iterations like $x_{k+1} = x_k * ( y / AMF(x_k) )$ and the simple variant $x = y / sqrt(AMF(y))$. The authors implement the scheme as MATLAB code (imBeam) and demonstrate strong performance on extremely dark images and hazy conditions, outperforming several state-of-the-art LLIE and dehazing methods and even aiding downstream tasks such as object detection with YOLOX. The results suggest that a compact, principled inverse-filter framework can rival complex deep-learning pipelines while retaining interpretability and easy customization, with potential applications in medical imaging and beyond.

Abstract

We introduce a simple and efficient method to enhance and clarify images. More specifically, we deal with low light image enhancement and clarification of hazy imagery (hazy/foggy images, images containing sand dust, and underwater images). Our method involves constructing an image filter to simulate low-light or hazy conditions and deriving approximate reverse filters to minimize distortions in the enhanced images. Experimental results show that our approach is highly competitive and often surpasses state-of-the-art techniques in handling extremely dark images and in enhancing hazy images. A key advantage of our approach lies in its simplicity: Our method is implementable with just a few lines of MATLAB code.

Figures (7)

• Figure 1: MATLAB script for image brightening imBeam and simple post-processing scheme pp.
• Figure 2: Visual comparison of the proposed method with several popular LLIE schemes. While these schemes show excellent performance for not-so-dark images, they are not able to match the performance of imBeam for extremely dark images.
• Figure 3: Images from the Dark Face collection used in various LLIE papers: Dark Face 8 CoLIE_eccv24, Dark Face 9, LightenDiffusion_eccv24, Dark Face 11, Zero-DCE_cvpr20, Dark Face 101 Li-etal_pami22, Dark Face 1353 Niu_TCI2025, Dark Face 1462 SCI_cvpr22, and Dark Face 5720 LMT_eccv24) are shown in the upper row and filtered with the proposed imBeam in the lower row. The reader is invited to compare our results with those presented in the above-mentioned papers.
• Figure 4: Visual comparison of several SOTA prior-based image clarification schemes and their improved version, followed by object detection. '+im$/\!\sqrt{\texttt{GF}}$ ' indicates that the image was enhanced by im$/\!\sqrt{\texttt{GF}}$ . Object detection was performed using YOLOX Ge2021yolox.
• Figure 5: Visual comparison of SOTA prior-based image clarification schemes and the proposed improvement for enhancing a challenging image with heavy fog. Notice how combining im$/\!\sqrt{\texttt{GF}}$ with other methods further improves dehazing results. This is particularly apparent for the results obtained by SLP.