A Neuromorphic Proto-Object Based Dynamic Visual Saliency Model With a Hybrid FPGA Implementation.

Area of Science:

• Computer Vision
• Neuroscience
• Robotics

Background:

• Visual attention is crucial for efficient processing in biological and engineered systems.
• Modeling dynamic visual saliency, mimicking human attention, is challenging for artificial systems.
• Current models often require extensive training and lack real-time capabilities.

Purpose of the Study:

• To develop a biologically plausible, training-free neuromorphic dynamic saliency model.
• To evaluate the model's performance against state-of-the-art saliency prediction methods.
• To implement the model on a Field-Programmable Gate Array (FPGA) for real-time applications.

Main Methods:

• A bottom-up, feed-forward neuromorphic model based on proto-objects and neurophysiological features was designed.
• The model's accuracy in predicting human eye fixations was compared to existing dynamic saliency models.
• A hybrid FPGA implementation was created to assess real-time processing capabilities.

Main Results:

• The neuromorphic model demonstrated superior performance in predicting human eye fixations compared to state-of-the-art methods.
• The FPGA implementation achieved a 23.77x speedup over the software version, processing 112x84 frames at 18.71 Hz.
• A fixed-point version of the FPGA model produced results comparable to the software implementation.

Conclusions:

• The proposed neuromorphic dynamic saliency model offers a computationally efficient and accurate approach to visual attention.
• The FPGA implementation enables real-time visual processing, suitable for robotics and autonomous systems.
• This work bridges the gap between biological visual processing and engineered systems.