Central-Peripheral Vision-Inspired Framework (CVP)

• CVP is a vision framework inspired by the human retina’s central–peripheral structure, integrating high-acuity central and broad-field peripheral processing.
• It employs dual-pathway models, log-polar transformations, and transformer-based mechanisms to dynamically fuse detailed and contextual visual information.
• Empirical studies demonstrate that CVP improves recognition accuracy, reduces processing latency, and enhances resource efficiency in vision tasks.

The Central-Peripheral Vision-Inspired Framework (CVP) encapsulates a biologically and computationally motivated paradigm for vision system design, analysis, and modeling. It draws foundational inspiration from the human retina’s spatially inhomogeneous photoreceptor distribution and the central–peripheral dichotomy in visual processing. CVP frameworks formalize and exploit these anatomical and perceptual insights to optimize visual representation, attention, perception, resource allocation, and spatial reasoning in both biological and artificial systems. CVP has been concretely instantiated in computational neuroscience, computer vision, quality assessment, multimodal reasoning, reinforcement learning, and transformer-based recognition architectures.

1. Biological Foundations and Theoretical Rationale

The CVP framework is rooted in pronounced biological structure. The foveal (central) region of the human retina exhibits a cone photoreceptor density of approximately 200,000 cones/mm² at the center, rapidly falling to ~10,000 cones/mm² at 10° eccentricity (Zhaoping, 24 Mar 2025, Guo et al., 2018). This gradient underlies two fundamental visual subfunctions:

• Central Vision: Supports high-acuity, detailed recognition (e.g., faces, text) and conscious report (“seeing”).
• Peripheral Vision: Prioritizes lower-acuity, broad-field monitoring for salient events and saccadic guidance (“looking”).

This anatomical asymmetry maps directly onto computational bottlenecks: the retina and V1 receive vast information bandwidth (∼10⁷ bits/s), while downstream decoding is constrained to ~10² bits/s (Zhaoping, 24 Mar 2025). Peripheral vision is specialized for rapid, low-bandwidth, saliency-driven selection of new fixation targets; central vision exploits feedback and high-fidelity coding for fine discrimination and recognition. This dichotomy is formalized in the Central–Peripheral Dichotomy (CPD) theory and integrated into resource allocation schemes throughout the CVP literature (Zhaoping, 24 Mar 2025, Guo et al., 2018).

2. Core CVP Architectures and Computational Instantiations

2.1 Dual-Pathway and Gated-Mixture Models

Neural network models typically operationalize CVP by partitioning visual input into central and peripheral channels via soft or hard masks, log-polar transformations for cortical magnification mapping, and foveated rendering. Canonical instantiations process both branches with structurally identical CNN subnets, fusing their outputs via a gating or mixture-of-experts module—learned to weight central vs. peripheral features task-adaptively (Wang et al., 2017):

• For input $I(x, y)$ and masks $W_{\mathrm{c}}, W_{\mathrm{p}}$:
  • Central stream: $x_c(x, y) = W_c(x, y)\cdot I(x, y)$
  • Peripheral stream: $x_p(x, y) = W_p(x, y)\cdot I(x, y)$

Downstream, the network computes

$$g = W_g\cdot[h_c; h_p] + b_g, \quad \alpha = \mathrm{softmax}(g)$$

and fuses per-pathway predictions weighted by $\alpha_c, \alpha_p$, allowing dynamic task-dependent integration (Wang et al., 2017).

2.2 Transformer-Based and Multimodal CVP

In transformer architectures, PerViT introduces a multi-head peripheral attention (MPA) mechanism, injecting radial distance–dependent positional biases into the attention map to enforce central-peripheral partitioning at the level of attention heads (Min et al., 2022). Mathematically, at each head $h$, position-based attention $P$ alters the softmax map as:

$A = \mathrm{softmax}_\mathrm{row}(L_c + \log P)$

where $L_c$ is the content-based attention and $P$ is output by a learned MLP over radial distances.

The large multimodal model-based CVP approach for spatial reasoning incorporates:

• Target-Affinity Token (central vision analog): A trainable token prepended to the sequence, producing a query vector $q$ for contrastive object retrieval in 3D environments.
• Allocentric Grid (peripheral vision analog): Discretized scene layout as text describing object presence per cell, establishing global spatial context (Chen et al., 9 Dec 2025).

3. Mathematical Modeling and Parameterization

CVP frameworks often explicitly encode human eccentricity-dependent sensitivity in model parameters. For VR image quality, quantization step size $q(\theta)$ and spatial resolution $s(\theta)$ as functions of eccentric angle $\theta$ are modeled as normalized, generalized Gaussians with asymptote:

$$\hat{q}(\theta) = \frac{1}{c_q \sqrt{2\pi} \exp\left(-(|b_q \theta|^{a_0})/(2 c_q^2)\right)} + d_q$$

(similar for $\hat{s}(\theta)$), where $a_0, b_q, c_q, d_q$ are fit on human JND data (Guo et al., 2018). Content-adaptive terms (e.g., $c_s$) are predicted via linear regressions on statistics such as spatial information and Gabor-filter response. This guides real-time per-region quality assignment aligned to psychophysical thresholds.

For multimodal models, the target-affinity mechanism is supervised by an InfoNCE contrastive loss,

$$\mathcal{L}_\mathrm{InfoNCE} = -\log\left(\frac{\sum_{e_+ \in \mathcal{E}_+} \exp(q^T e_+ / \tau)}{\sum_{e \in \mathcal{E}} \exp(q^T e / \tau)}\right)$$

where $q$ is the query and $e_+$, $e$ are embeddings for relevant and all objects, respectively (Chen et al., 9 Dec 2025).

4. Empirical Findings and Task-Specific Profiles

Empirical validation of CVP spans behavioral replication, system-level benchmarks, and ablation analysis:

• Scene Recognition: Peripheral-only input outperforms central-only at small mask radii; central vision is more area-efficient up to a crossover at ∼10.8° visual angle. Combined CVP dual-pathway models always exceed either alone (e.g., up to 95% accuracy vs. max 93%) (Wang et al., 2017). Peripheral streams often specialize in natural scenes, central in man-made.
• Object/Face Recognition: Face and object tasks are more dependent on central vision; face verification is nearly abolished (>30% drop) if only peripheral data is available (Wang et al., 2016). Peripheral streams are highly inefficient for faces.
• Multimodal 3D Reasoning: The CVP model exceeds Video-3D-LLM on spatial reasoning tasks (e.g., ScanQA: CIDEr 107.1 vs. 102.1; ScanRefer [email protected]: 62.0% vs. 58.1%). Removal of either the target-affinity (central) or allocentric grid (peripheral) components produces significant performance drops (Chen et al., 9 Dec 2025).
• Quality Assessment and Streaming: Applying region-adaptive quantization (CVP) reduces gigapixel image retrieval times by ≈90% (from ~6–8s to ~0.5–1s); users experience no mean opinion score loss compared to uniform-quality schemes (Guo et al., 2018).

5. Algorithms for Active Vision and Information Bottleneck

The V1SH–Bottleneck–CPD framework (Zhaoping, 24 Mar 2025) formalizes visual information flow as alternating “looking” (peripheral-driven, saliency map computed in V1, selection of saccade target) and “seeing” (central, feedback-mediated decoding under bottleneck constraints). Mathematically:

• Saccade target: $b(x, y) = \max_{i : (x_i, y_i)$ covers $(x, y)} r_i$, with $r_i$ the response of V1 neuron $i$
• "Feedforward–Feedback–Verify-and-reWeight" (FFVW) inference algorithm:
  1. Feedforward: initial hypothesis posteriors from compressed V1 signals
  2. Feedback: select neurons to query for information maximizing mutual information
  3. Verify and reweight hypotheses via observed vs. predicted V1 responses
  4. Iterate or terminate based on confidence/resource limits

This theoretical formalism explains robust central immunity and peripheral susceptibility to illusions (e.g., flip-tilt, reversed-depth), with immunity lost under feedback masking.

6. Extensions, Limitations, and Broader Implications

Extensions

• Active Control: Hierarchical CVP models for vergence control deploy multiple expert modules over nested perimeters (foveal, inner-, and outer-peripheral) and a gating network, demonstrating 33% reduced alignment error and 56% lower oscillation relative to non-hierarchical baselines (Zhao et al., 2021).

• Transformer Generalization: Peripheral-attention heads can be swapped into any ViT/DeiT-like architecture, and the same CVP heuristics can guide object detection, segmentation, and video analysis (Min et al., 2022).
• Resource-Efficient Streaming and Sensing: CVP-based foveated acquisition and transmission can cut network and storage requirements by an order of magnitude (Guo et al., 2018).

Limitations

• Feedforward CVP networks may fail to capture the recurrent, feedback-sensitive central vision mechanisms critical for ambiguity/noise disambiguation (Wang et al., 2016, Zhaoping, 24 Mar 2025).
• Simplified partitioning (circular masks, hard concentric zones) does not fully emulate real retinal and cortical mappings (e.g., log-polar).
• Coverage for dynamic, saccadic, and higher-order cognitive integration is an area for further advancement.

Broader Significance

CVP frameworks formalize and mechanistically instantiate the central–peripheral trade-offs fundamental to vision, providing optimality principles that inform both biological theory and artificial system architecture. This includes bottleneck-driven resource allocation, attention mechanisms, spatial context modeling, and quality-adaptive transmission. The approach sets a foundation for future research targeting:

• Foveated and saliency-driven active perception
• Recurrent vision models and feedback mechanisms
• Multimodal and reinforcement learning agents with explicit spatial reasoning
• Cross-modal extensions in audition and somatosensory domains

CVP has unified quantitative, algorithmic, psychophysical, and neurobiological research, producing falsifiable predictions and concrete performance improvements in artificial systems (Zhaoping, 24 Mar 2025, Chen et al., 9 Dec 2025, Guo et al., 2018, Min et al., 2022, Zhao et al., 2021, Wang et al., 2017, Wang et al., 2016).