HVPNet: A bio-inspired network for general salient and camouflaged object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biologically Inspired Computing

Background:

• Current multimodal salient object detection (SOD) and camouflaged object detection (COD) models often rely on complex fusion and decoding structures, leading to large parameter counts and redundancy.
• This complexity contrasts with the human visual system's efficiency in identifying salient and camouflaged objects using simpler mechanisms.

Purpose of the Study:

• To investigate if conceptual inspiration from the human visual process can lead to simpler, yet accurate and efficient, object detection models.
• To propose and evaluate a novel bio-inspired computational architecture for multimodal SOD and COD.

Main Methods:

• Proposed HVPNet, a simple and general bio-inspired architecture.
• Introduced a Retinal Integration Module (RIM) to integrate multimodal features via level-specific, multi-stage integration, inspired by retinal information processing.
• Designed a cortical decoder (CD) that mimics hierarchical visual processing in the human cortex by separating decoding into low- and high-level visual stages.

Main Results:

• HVPNet demonstrated versatility, extending to seven tasks across four modalities.
• The model achieved an excellent accuracy-efficiency trade-off across 22 diverse datasets.
• The proposed architecture avoids complex, redundant structures common in prior work.

Conclusions:

• A simpler, bio-inspired approach (HVPNet) can achieve competitive performance in multimodal SOD and COD.
• The Retinal Integration Module and cortical decoder effectively process multimodal information.
• HVPNet offers a promising direction for developing efficient and accurate object detection systems by drawing parallels with human visual processing.