HDR vision sensor with neuro-memristive skin detection for edge computing.

Area of Science:

• Computer Vision
• Neuromorphic Engineering
• Materials Science

Background:

• Human skin classification is crucial for applications like human-machine interfaces and object tracking.
• Existing methods often require complex post-processing and are sensitive to lighting variations.
• There is a need for efficient, low-power, and robust skin detection systems.

Purpose of the Study:

• To develop a hybrid CMOS/memristor vision sensor architecture for in-sensor skin detection.
• To implement an analog memristive neural network (NN) for efficient skin classification.
• To achieve robust skin detection across a wide dynamic range and varying illuminant conditions.

Main Methods:

• A hybrid CMOS/memristor vision sensor was designed and implemented.
• In-sensor RGB to r g-chromaticity color-space conversion was performed using pixel-level automatic exposure control.
• An analog memristive neural network (NN) was utilized for direct skin/non-skin classification of analog signals.
• The performance and hardware integration of the memristive NN were studied.

Main Results:

• The proposed sensor architecture enables on-the-fly skin detection.
• The system achieves skin classification directly from analog signals without further processing.
• The memristive neural network (NN) demonstrated efficient and low-power operation.
• The output skin bitmap is robust against changes in illuminant color and intensity.

Conclusions:

• The hybrid CMOS/memristor vision sensor offers a flexible, customizable, and low-power solution for skin detection.
• In-sensor processing using memristive neural networks (NNs) significantly simplifies the classification pipeline.
• This architecture is well-suited for various machine vision applications requiring real-time skin detection.