HVPNet: A Bio-Inspired Network for General Salient and Camouflaged Object Detection

Abstract: In recent years, most research on multimodal salient object detection (SOD) and camouflaged object detection (COD) typically aims to improve performance through complex cross-modal feature fusion and decoding structures. However, this approach leads to an excessively large model parameter scale and often fails to deliver satisfactory detection performance due to structural redundancy. In contrast, the human visual process is able to efficiently perform salient and camouflaged object identification without such complex structures. This contrast raises an important question: Can we draw conceptual inspiration from the human visual process to achieve a simpler modeling strategy, and still realize accurate and efficient object detection? To answer this question, we propose HVPNet, a simple yet general bio-inspired computational architecture. Drawing on the multi-layered information integration of the retina as a conceptual metaphor, we designed a Retinal Integration Module (RIM), which effectively integrates multimodal features through a level-specific multi-stage integration strategy. To fully exploit these features, we further design a cortical decoder (CD) that breaks down the decoding process into low- and high-level visual stages, abstracting the hierarchical processing in the human visual cortex. Benefiting from these designs, HVPNet can readily extend to seven tasks across four modalities. Without bells and whistles, it establishes an excellent accuracy-efficiency trade-off across 22 datasets spanning these seven tasks. Our code is available at https://github.com/jiaweiXu1029/HVPNet.

• The paper introduces a bio-inspired design leveraging retinal and cortical modules to achieve efficient, multimodal detection with up to 90% parameter reduction.
• It employs a three-stage fusion and decoding pipeline that outperforms state-of-the-art methods across seven tasks and 22 datasets.
• Empirical results validate its robust performance in complex scenarios while also highlighting limitations under extreme conditions.

Bio-Inspired Generalized Object Detection: An Analysis of HVPNet

Motivation and Context

Salient object detection (SOD) and camouflaged object detection (COD) are central tasks in computer vision, each requiring high discrimination between foreground and background across challenging scenarios, including multimodal settings (RGB, RGB-D, RGB-T, video modalities). Prevailing approaches emphasize complex, parameter-heavy fusion and decoding structures, frequently resulting in inefficient architectures with limited scalability. The HVPNet framework directly addresses this by abstracting principles from the human visual system to achieve efficient, multi-modal SOD and COD without excessive architectural overhead. The method aims for accurate detection across seven tasks and 22 datasets, all via a single, structurally concise pipeline.

Architectural Overview

HVPNet is composed of three functional stages: feature extraction, fusion by the Retinal Integration Module (RIM), and hierarchical decoding via the Cortical Decoder (CD). The network explicitly leverages the multi-layered information integration of the retina and the hierarchical abstraction pipeline of the cerebral cortex.

Figure 1: HVPNet's three-stage pipeline: feature extraction, retinal integration, cortical decoding, designed for general SOD and COD across multiple modalities.

Retinal Integration Module (RIM)

The RIM mimics the retina's layered processing. It operates via three level-specific fusion stages:

• Stage 1: Low-level fusion prioritizes edge and contour information, using element-wise addition/multiplication, followed by detail extraction and dilation mechanisms for multi-scale structure capture.
• Stage 2: Mid-level fusion applies Selective Region Attention (SRA) to boost structural semantic representations, using spatial and channel-wise attention with dilation-based pooling. The SRA is stringently ablated to maximize representation quality.
• Stage 3: High-level fusion aggregates global semantics and applies channel optimization/fusion plus spatial attention for robust contextual integration, followed by dimensionality reduction.

  Figure 2: Detailed visual flow of RIM, with staged, modality-specific integration generating cross-modal features for subsequent decoding.

Cortical Decoder (CD)

Inspired by the hierarchical processing of the visual cortex (V1, V4, IT), the CD decouples semantic abstraction and detail refinement:

• High-Level Visual Decoder (HLVD): Performs multi-level fusion via dilated and separable convolutions, aligning feature resolutions, and generates original attention maps.
• Gaussian Guide Attention (GGA): Applies spatial smoothing and normalization to high-level maps, integrated with max operations for bio-inspired modulation.
• Low-Level Visual Decoder (LLVD): Fuses high-level modulation and low-level details via optimized convolutions, generating predictions with enhanced boundary accuracy.

  Figure 3: Schematic of the CD, revealing hierarchical decoding and modulation for robust detail and semantic preservation.

Empirical Results

HVPNet achieves strong results across a wide spectrum of tasks:

• The framework consistently outperforms prior SOTA solutions, not only in accuracy metrics ($E_m$, $S_m$, $F_m$, mean absolute error) but also in efficiency (90%+ reduction in params and FLOPs vs. complex Transformer-based models).
• It generalizes across seven tasks, including RGB SOD, RGB-D SOD/COD, RGB-T SOD, VSOD, and VCOD, demonstrating versatility and structural robustness.

  Figure 4: Parameter count vs. structural measure comparison on diverse datasets, showing HVPNet's superior efficiency-performance trade-off.

  

  Figure 5: PR curve comparisons on eight SOD datasets, showing HVPNet matches or exceeds heavy models across modalities.

Qualitative analyses illustrate robust handling of small, occluded, and boundary-ambiguous objects; model excels even with orders-of-magnitude less compute.

Figure 6: Extensive cross-task qualitative comparisons, highlighting boundary preservation and failure cases for various models.

Ablation studies confirm the necessity of staged RIM, the synergy of lightweight backbones, and the strict order and selection of fusion mechanisms. Heavy backbones degrade performance; staged fusion is strictly necessary for optimal results.

Figure 7: Visualized feature maps at each RIM stage, demonstrating preservation of structural details across modalities.

Limitations and Failure Modes

Despite superior overall performance, HVPNet underperforms in select extreme conditions:

• Mutual cross-modal misactivation (e.g., RGB/Thermal misleading cues) results in amplified false positives.
• Objects with transparent interiors or weak cues challenge holistic representation; fragmented predictions observed.
• Highly complex or blurred boundaries induce ambiguity, limiting accurate contour delineation.

  Figure 8: Illustrative failure cases showing limitations in suppressing modal interference and boundary ambiguity.

Practical and Theoretical Implications

The HVPNet approach demonstrates that brain-inspired, level-specific operations can achieve high accuracy and efficiency in multimodal detection, challenging the paradigm of ever-increasing architectural complexity and parameter count. The synergy of lightweight encoders and staged fusion/decoding yields performance previously seen only in resource-intensive models. The strict necessity of layer ordering and fusion specificity supports deeper exploration of biologically plausible architectures.

Given the confirmed generalizability across object detection and segmentation tasks, this work provides a scalable baseline for practical deployment in resource-constrained environments (e.g., real-time video, edge devices). Its results also point toward future developments: more advanced biological simulation, adaptive modularity, and robust handling of ambiguous input modalities.

Conclusion

HVPNet introduces a conceptual bio-inspired pipeline for SOD and COD, achieving state-of-the-art performance with minimal architectural complexity. The model’s retinal and cortical analogues enforce structural compatibility and efficient multi-level cue coordination, validated by extensive quantitative and qualitative analysis. This work confirms the value of biologically motivated, efficient architectures and lays the foundation for future explorations in multimodal detection and efficient general AI vision systems.

(2606.31496)