FCDFusion: RGB-Based Real-Time Image Fusion

• FCDFusion is a visible-and-infrared image fusion method that preserves the natural color of the visible image by directly scaling its RGB vectors without using conventional color-space transformations.
• It utilizes a gamma-corrected scaling factor computed from infrared brightness cues to enhance detail and contrast while operating at only 7 FLOPs per pixel for real-time performance.
• The method maintains color fidelity by geometrically preserving the direction of the RGB vectors, resulting in near-zero color deviation compared to traditional methods.

FCDFusion is a visible-and-infrared image fusion (VIF) method designed to preserve the color of the visible image while injecting the brightness and detail cues of the infrared image, but without the HSV, HSI, YIQ, or YUV color-space transformations that are commonly used in fast color-preserving fusion pipelines (Li et al., 2024). Its central premise is that fusion can be implemented directly in RGB space as a scaling operation on a visible pixel’s three-dimensional color vector, thereby retaining color direction while modifying brightness. The method incorporates gamma correction at little extra cost, aims at low color deviation and good contrast, and is presented as a real-time-friendly alternative to transform-based VIF schemes (Li et al., 2024).

1. Problem setting and design objective

Visible-and-infrared image fusion addresses a standard complementarity: infrared imagery is useful for revealing objects in dark, backlit, or low-visibility scenes, whereas visible imagery provides natural color. Many prior color-preserving VIF methods convert RGB images into a color space such as HSV, HSI, YIQ, or YUV, fuse only a brightness-like component, and then convert back to RGB. In FCDFusion, these conversions are treated as the computational bottleneck in otherwise fast pipelines (Li et al., 2024).

The method is therefore framed around four explicit objectives: low color deviation, good contrast, very low computation, and real-time friendliness. Rather than decomposing hue, saturation, and intensity explicitly, FCDFusion reasons directly in RGB vector space. This design choice is motivated by the observation that if the operational goal is to preserve visible-image color while modulating brightness using infrared information, then expensive color-space transforms are not strictly necessary (Li et al., 2024).

A common assumption in fast color-preserving VIF is that color fidelity requires explicit hue-saturation preservation through auxiliary color spaces. FCDFusion departs from that assumption. This suggests a reformulation of the VIF problem in which chromatic preservation is enforced geometrically, by maintaining vector direction, rather than analytically, by transforming into a color space whose channels are interpreted separately.

2. RGB-vector formulation

The key conceptual step in FCDFusion is to treat each RGB pixel as a three-dimensional color vector,

$$\mathbf{c}= \begin{bmatrix} r\ g\ b \end{bmatrix}.$$

In this representation, the direction of the vector corresponds to color characteristics, described as hue/saturation-like information, while the vector length corresponds to brightness (Li et al., 2024).

Under this interpretation, preserving color while changing brightness reduces to a uniform scaling of all three channels by the same scalar $k$: $\mathbf{c}_f = k\,\mathbf{c}_v.$ Because scaling leaves direction unchanged in the ideal case, color is retained while intensity is adjusted. FCDFusion therefore formulates fusion not as channel mixing, but as vector scaling of the visible pixel under the control of the infrared signal (Li et al., 2024).

This is the main distinction from direct RGB averaging. RGB averaging is computationally cheap, but it averages each visible channel with the infrared gray value, which the paper associates with dulled saturation, blurred details, and color washout. By contrast, transform-based methods such as YIQ-AVG and HSV-AVG preserve color by separating brightness from chromatic components, but incur the cost of forward and inverse transforms. FCDFusion seeks the color-preserving behavior of those methods while remaining in native RGB coordinates (Li et al., 2024).

3. Fusion rule and scaling-factor construction

For each aligned visible/infrared pair, the visible pixel is written as

$\mathbf{c}_v = [r_v, g_v, b_v]^T,$

and the infrared pixel as a single-channel grayscale value $v_i$. FCDFusion computes a scaling factor $k$ from the infrared intensity and the visible pixel’s maximum RGB component, then outputs

$$\mathbf{c}_f= \begin{bmatrix} \min(k r_v, 255)\ \min(k g_v, 255)\ \min(k b_v, 255) \end{bmatrix}.$$

Operationally, the procedure is: estimate a brightness multiplier from the infrared image, apply that multiplier to the visible RGB vector, and clip the result to $[0,255]$ (Li et al., 2024).

The infrared-driven factor is

$$\alpha = \left(\frac{v_i}{255}\right)^\gamma \in [0,1].$$

The method chooses $\gamma=2$, giving

$k$0

Gamma correction is used because larger $k$1 emphasizes bright regions and suppresses dark-region contrast; the paper further states that this helps reduce noise in dark areas and improves object enhancement. The specific choice $k$2 is presented as a compromise that is visually similar to common display gamma values such as 2.2 while remaining computationally cheap because squaring is faster than general exponentiation (Li et al., 2024).

Visible-pixel brightness is approximated by

$k$3

Using the maximum channel avoids a more expensive luminance computation. The maximum allowable scale is then defined as

$k$4

which allows the largest channel to expand up to $k$5 while avoiding division by zero. The method forms

$k$6

but does not use $k$7 directly. Instead it averages this amplified scale with the identity scale: $k$8 so that

$k$9

The paper also reports the equivalent form

$\mathbf{c}_f = k\,\mathbf{c}_v.$0

An implementation detail noted in the paper is that $\mathbf{c}_f = k\,\mathbf{c}_v.$1 can be computed by a right shift, reducing cost relative to a true division (Li et al., 2024).

The averaging with 1 in $\mathbf{c}_f = k\,\mathbf{c}_v.$2 is introduced to prevent over-aggressive brightening. The paper states that without this averaging, object enhancement may become stronger, but background detail can also become more distorted or blurred; the averaged form gives a better balance for general scenes (Li et al., 2024).

4. Color preservation and the color deviation metric

FCDFusion’s color-preservation argument is geometric. If the fused vector $\mathbf{c}_f = k\,\mathbf{c}_v.$3 remains collinear with the visible vector $\mathbf{c}_f = k\,\mathbf{c}_v.$4, then the angle between them is zero. On this basis, the paper proposes the color deviation (CD) metric: $\mathbf{c}_f = k\,\mathbf{c}_v.$5 The overall image-pair score is the average CD over all pixels (Li et al., 2024).

The purpose of CD is to evaluate the ability of a VIF method to preserve color in settings where only the visible image carries color and the infrared image is monochrome. Existing VIF metrics are described as focusing on contrast, gradient strength, structural similarity, or similarity to both inputs; the paper argues that such metrics are not tailored to color-preserving fusion because they may reward hue or saturation distortion, or may compare fused color against a monochrome infrared reference in an unhelpful way. CD instead isolates the deviation of the fused color vector from the visible color vector (Li et al., 2024).

The geometric interpretation of competing fast methods is expressed in the paper as follows. RGB averaging mixes the visible color vector with the infrared diagonal vector and therefore causes large angle changes. YIQ averaging changes only the luminance-like channel, so the angle change is smaller. HSV averaging preserves hue and saturation directly, so angle deviation is near zero. FCDFusion also preserves direction well because it uses scaling; its slight disadvantage relative to HSV is attributed to clipping at $\mathbf{c}_f = k\,\mathbf{c}_v.$6, which can perturb direction slightly. The paper summarizes this ordering as

$\mathbf{c}_f = k\,\mathbf{c}_v.$7

This suggests that FCDFusion’s principal innovation is not merely computational efficiency, but a reframing of color preservation as directional stability in RGB space (Li et al., 2024).

5. Computational profile and empirical evaluation

FCDFusion is explicitly presented as a low-cost method. The paper’s per-pixel accounting reports RGB-AVG at 0 FLOPs for fusion itself, YIQ-AVG at 18 FLOPs, HSV-AVG at 14 FLOPs, and FCDFusion at 7 FLOPs. It further summarizes FCDFusion as

$\mathbf{c}_f = k\,\mathbf{c}_v.$8

This is the basis for the claim that the method operates at roughly half the cost of HSV-based fast color-preserving methods (Li et al., 2024).

A separate FLOP comparison in the same description reports RGB-AVG: 23.76K, HSV-AVG: 2.33M, YIQ-AVG: 2.99M, and FCDFusion: 1.16M. The paper interprets this comparison as showing that FCDFusion is much cheaper than HSV/YIQ fast color-preserving methods, while still more computationally involved than plain RGB averaging and substantially better in visual quality and color fidelity (Li et al., 2024).

The evaluation is conducted on the VIFB benchmark, using 21 visible/infrared pairs with color visible images and dim and backlit scenes. The compared methods are RGB-AVG, YIQ-AVG, HSV-AVG, MST-SR, CNN, PIAFusion, and SeAFusion. The reported findings are that FCDFusion achieves higher contrast than the three averaging methods, preserves color better than most non-HSV methods, gives the best user study score among all compared methods, is much faster than the deep-learning and multi-scale methods, and has 0 parameters (Li et al., 2024).

For the CD metric, the paper reports that FCDFusion and HSV-AVG obtain the smallest values among the fast methods:

• FCDFusion: CD $\mathbf{c}_f = k\,\mathbf{c}_v.$9
• HSV-AVG: CD $\mathbf{c}_v = [r_v, g_v, b_v]^T,$0

This result is used to support the claim that FCDFusion attains near-HSV-level color preservation without any color-space conversion (Li et al., 2024).

6. Assumptions, limitations, and practical scope

FCDFusion assumes that visible and infrared images are already registered or aligned. The infrared image is treated as a single-channel grayscale signal, conceptually replicated across RGB, and the visible image’s chromatic direction is assumed to be the quantity that should be preserved while brightness is modified (Li et al., 2024).

The method is especially positioned for settings where real-time performance matters, hardware resources are limited, color fidelity is important, and a simple, interpretable fusion rule is preferred. The paper therefore identifies embedded or lightweight systems as a natural use case, particularly when the overhead of HSV or YIQ conversion is undesirable (Li et al., 2024).

The principal limitation stated in the paper is that direct scaling followed by clipping can saturate channels at 255, causing slight color-direction distortion; this is one reason the CD is not exactly zero. The paper also notes that, like most pixelwise methods, FCDFusion does not explicitly model complex spatial structure beyond local brightness scaling (Li et al., 2024).

A common misconception would be to classify FCDFusion primarily as a color-space simplification of HSV-style fusion. A more precise characterization is that it replaces explicit color decomposition with an RGB-vector geometry in which chromatic preservation is encoded by directional invariance. This suggests a broader methodological implication for VIF: in some color-preserving settings, the essential operation may be scalar modulation of a color vector rather than transformation into an auxiliary representation (Li et al., 2024).