CoSOV1Net: A Cone- and Spatial-Opponent Primary Visual Cortex-Inspired Neural Network for Lightweight Salient Object...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Neuroscience

Background:

• Deep neural networks excel at salient object detection on high-end systems.
• Resource-limited devices face challenges in deploying high-performance models.
• Existing models often rely on pre-trained backbones, limiting adaptability.

Purpose of the Study:

• To develop a novel, lightweight salient object detection model (CoSOV1net).
• To mimic the human visual system's color and shape perception.
• To enable high-performance salient object detection on resource-constrained devices.

Main Methods:

• Proposed CoSOV1net, a neural network inspired by V1 cone and spatial-opponent processes.
• Trained the model from scratch without using pre-trained backbones.
• Evaluated performance on standard salient object detection datasets.

Main Results:

• CoSOV1net achieved competitive performance (Fβ=0.931 on ECSSD) against state-of-the-art models.
• The model boasts a low parameter count (1.14 M) and computational cost (1.4 G FLOPS).
• Achieved high inference speed (211.2 FPS) on a GPU, demonstrating efficiency.

Conclusions:

• CoSOV1net is an effective lightweight model for salient object detection.
• Its efficiency makes it suitable for mobile and resource-limited applications.
• The model demonstrates the potential of bio-inspired designs for efficient AI.