Central Difference Convolution (CDC)

• Central Difference Convolution (CDC) is a technique that fuses intensity-based feature extraction with explicit local gradient cues using a tradeoff parameter θ.
• It enhances performance in tasks like face anti-spoofing and action recognition by improving sensitivity to fine-grained patterns and robustness to illumination changes.
• Variants such as cross-pattern CDC and 3D CDC extend its application to spatio-temporal data while reducing parameter count and computational cost.

Central Difference Convolution (CDC) is a generalization of standard convolutional operators that augments intensity-based feature extraction with explicit local gradient (finite-difference) cues. Originating in the context of face anti-spoofing, it has been validated to enhance fine-grained pattern recognition, robustness to illumination variance, and generalization across imaging domains. The CDC formulation is parameter- and computation-efficient, can be seamlessly integrated into convolutional neural network (CNN) architectures, and has been extended to sparse, cross-pattern variants and to spatio-temporal (3D) domains.

1. Mathematical Formulation and Properties

Let $x$ be a 2D input feature map and $w$ a $K \times K$ convolutional kernel, with $\mathcal{R}$ the kernel spatial support. The vanilla convolution outputs at position $p_0$:

$$y_{\text{conv}}(p_0) = \sum_{p_n \in \mathcal{R}} w(p_n) \, x(p_0 + p_n)$$

The central-difference (gradient) operator aggregates local differences relative to $x(p_0)$:

$$y_{\text{CD}}(p_0) = \sum_{p_n \in \mathcal{R}} w(p_n) \big[ x(p_0 + p_n) - x(p_0) \big]$$

Central Difference Convolution blends these via a hyperparameter $\theta \in [0, 1]$:

$$y_{\text{CDC}}(p_0) = (1-\theta)\sum_{p_n \in \mathcal{R}} w(p_n)x(p_0+p_n) + \theta\sum_{p_n \in \mathcal{R}} w(p_n)[x(p_0+p_n)-x(p_0)]$$

which simplifies to:

$$y_{\text{CDC}}(p_0) = \sum_{p_n \in \mathcal{R}} w(p_n) x(p_0 + p_n) - \theta\, x(p_0) \sum_{p_n \in \mathcal{R}} w(p_n)$$

No extra learnable parameters are introduced beyond the vanilla kernel. When $\theta = 0$, CDC reverts to standard convolution; $\theta = 1$ yields a weighted local gradient. The parameter $\theta$ governs the tradeoff between intensity and gradient sensitivity (Yu et al., 2020, Yu et al., 2021, Ma et al., 2023).

2. Motivation and Theoretical Rationale

CDC targets the limitations of vanilla convolution in environments where local high-frequency artifacts (edges, moiré, reflection, print-pixelation) rather than global semantic context are discriminative. Vanilla convolutions primarily aggregate local intensities, leading to poor sensitivity to subtle spatial variations and susceptibility to illumination changes. The central difference term introduces explicit sensitivity to local gradients and edges, resulting in improved robustness to background or domain shifts and illumination variance, which is crucial for tasks such as face anti-spoofing and fine-grained motion perception (Yu et al., 2020, Yu et al., 2021, Ma et al., 2023).

3. CDC Variants: Cross-Pattern and 3D Extensions

Cross Central Difference Convolutions (C-CDC)

Aggregating central-difference signals from all eight neighbors can introduce parameter redundancy and directional competition. Decoupled cross-pattern variants are introduced:

• C-CDC(HV): Horizontal and vertical cross ($$\mathcal{S}_{HV} = \{(-1,0), (0,-1), (0,0), (0,1), (1,0)\}$$)
• C-CDC(DG): Diagonal cross ($$\mathcal{S}_{DG} = \{(-1,-1),(-1,1),(0,0),(1,-1),(1,1)\}$$)

For each, the CDC operator is computed on the reduced support $\mathcal{S} \subset \mathcal{R}$:

$$y_{\text{C-CDC}}(p_0) = (1-\theta) \sum_{p_n \in \mathcal{S}} w(p_n)x(p_0+p_n) + \theta\sum_{p_n \in \mathcal{S}} w(p_n)[x(p_0+p_n)-x(p_0)]$$

C-CDC decreases parameter count and FLOPs by $5/9$ relative to CDC (for a $3 \times 3$ kernel), e.g., $1.25$M vs $2.25$M parameters on typical face-anti-spoofing backbones, and achieves higher efficiency without performance compromise (Yu et al., 2021).

3D Central Difference Convolution (CDC-3D)

CDC is generalized to 3D for spatio-temporal data:

• CDC-ST (spatio-temporal): Full spatio-temporal cube, fusing all neighbor cube gradients.
• CDC-T (temporal-only): Central differences only along the temporal axis.

The 3D extension for a voxel $p_0$ in cube $\mathcal{Z}$:

$$y_{3D-\text{CDC-ST}}(p_0) = \theta\sum_{p_n\in\mathcal{Z}}w(p_n)[x(p_0+p_n)-x(p_0)] +(1-\theta)\sum_{p_n\in\mathcal{Z}}w(p_n)x(p_0+p_n)$$

For CDC-T, only temporal neighbors contribute to the central-difference term. These designs yield fine-grained spatio-temporal feature extraction, especially beneficial for human action and gesture recognition (Ma et al., 2023).

4. Network Architectures and Practical Integration

CDC is implemented as a drop-in replacement for standard convolutional layers. On 2D, operational efficiency is maintained by decomposing CDC into a vanilla convolution and a $1 \times 1$ convolution:

class CDC(nn.Module): def __init__(self, in_ch, out_ch, kernel=3, padding=1, theta=0.7): super(CDC, self).__init__() self.vani = nn.Conv2d(in_ch, out_ch, kernel_size=kernel, padding=padding) self.theta = theta def forward(self, x): out_v = self.vani(x) kd = self.vani.weight.sum(dim=2).sum(dim=2) kd = kd.view(kd.size(0), kd.size(1), 1, 1) out_cd = F.conv2d(x, kd, padding=0) return out_v - self.theta * out_cd

For CDCN-like architectures, all $3 \times 3$ convolutions are replaced by their CDC analogs with a fixed or learnable $\theta$ (empirically, $\theta=0.7$ on anti-spoofing; $\theta=0.6$ on action recognition). Integration into network backbones is systematic and computationally light, requiring only one additional $1 \times 1$ convolution per layer (Yu et al., 2020, Ma et al., 2023).

In the CDCN framework, input images are mapped to low-resolution depth maps. Architectures are constructed from sequential blocks: a CDC stem, multiple feature extraction stages, multiscale fusions, and a regression head (see architecture specifics in Section 5).

The CDC enhancement is equally applicable to I3D-style 3D convolutional backbones for spatio-temporal recognition (Ma et al., 2023).

5. Architectural Advances: NAS, Attention, and Dual-Stream Designs

Neural Architecture Search (CDCN++)

CDCN++ utilizes a DARTS-style bi-level optimization over a CDC-rich operator search space, allowing independent architectures for low-, mid-, and high-level “cells.” Channel-doubling and output node softmax attention further increase modeling capacity. CDCN++ achieves $\simeq$3M parameters with increased expressive power and generalization (Yu et al., 2020).

Multiscale Attention Fusion Module (MAFM)

To exploit multi-level CDC features, MAFM imbues each level with spatial attention calculated via vanilla convolutional attention kernels (conv7 for low, conv5 for mid, conv3 for high). Correct kernel sizing is crucial; vanilla conv outperforms CDC within attention modules (Yu et al., 2020).

Dual-Cross Stream + CFIM

The Dual-Cross CDC Network (DC-CDN) implements two parallel C-CDC streams (HV and DG), each extracting a distinct subset of gradient cues. The Cross Feature Interaction Module (CFIM) adaptively fuses features between these streams by learning spatial scalar affinities at each pyramid level, promoting interaction without parameter explosion. Late-stage fusion concatenates all levels before final regression (Yu et al., 2021).

Patch Exchange (PE)

PE data augmentation randomly exchanges patches and corresponding pseudo-labels between samples within a batch, incentivizing the model to attend to localized spoof artifacts rather than global context. PE consistently improves generalization, particularly in partial and cross-domain designation scenarios (Yu et al., 2021).

6. Empirical Benchmarks and Comparisons

Extensive experimental validations have been reported:

• Face Anti-Spoofing (CDCN/CDCN++/DC-CDN):
  • On OULU-NPU Protocol 1: CDCN baseline ACER 1.0%, CDCN++ 0.2% (state-of-the-art), DC-CDN+PE 0.4% (Yu et al., 2020, Yu et al., 2021).
  • On SiW Protocol 1: CDCN and CDCN++ 0.12%, outperforming previous methods.
  • Cross-dataset CASIA→Replay: CDCN HTER 15.5%, CDCN++ 6.5%, DC-CDN+PE 6.0%.
  • DC-CDN reduces parameter count (1.25M vs 2.25M) and computational cost (0.37× vanilla) while outperforming on ACER and HTER metrics.
• RGB-D Action and Gesture Recognition (CDC-3D-Stem):
  • On NTU RGB-D (Cross-Subject): vanilla 3D conv (90.4% RGB acc), 3D-CDC-ST ($\theta=0.6$) 92.1% (+1.7%), 3D-CDC-T ($\theta=0.6$) 93.8% on depth (+2.6%) (Ma et al., 2023).
  • CDC-Stem consistently enhances low-level temporal gradient sensitivity and discriminative spatial encoding.

Visualization studies (feature maps, t-SNE) show CDC-enhanced architectures provide superior clustering of live/spoof classes, highlight fine patterns such as moiré and lattice artifacts, and exhibit stable generalization under domain shift (Yu et al., 2020).

7. Context, Generalization, and Future Implications

CDC introduces a principled, parameter-free mechanism to directly encode local gradients within standard convolutional frameworks. This approach is broadly beneficial wherever high-frequency, locally discriminative features determine target classification. Core properties—parameter-efficient gradient fusion, adaptability to spatial/temporal axes, and empirical robustness across domains—render CDC and its variants widely applicable, with demonstrated utility in anti-spoofing and spatio-temporal recognition. The directionality control in cross-pattern variants further accentuates selectivity and computational efficiency.

A plausible implication is that CDC-style operators, especially in 3D or multi-stream variants, may serve as competitive modules in architectures aiming for interpretable, efficient, and robust local feature extraction in domains beyond facial biometrics, including medical imaging and surveillance. Extensions likely include learnable or data-dependent $\theta$ scheduling, dynamic pattern selection, and hybridization with transformer-style global context modules (Yu et al., 2020, Yu et al., 2021, Ma et al., 2023).