A Compact High-Quality Image Demosaicking Neural Network for Edge-Computing Devices.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image demosaicking is crucial but challenging for image sensors.
• Deep learning (CNN) methods exist but struggle with real-time edge device constraints due to large model sizes.

Purpose of the Study:

• To develop a compact demosaicking neural network suitable for real-time edge computing.
• To improve computational efficiency and reduce model parameters without sacrificing reconstruction quality.

Main Methods:

• A UNet++ based compact demosaicking neural network is proposed.
• Incorporates densely connected blocks with depthwise separable convolutions for efficient feature extraction and parameter reduction.
• Utilizes Gaussian smoothing layers instead of down-sampling to preserve image information and expand receptive fields.

Main Results:

• The proposed network achieves a 42% speed improvement over the fastest CNN-based methods.
• Demonstrates comparable reconstruction quality on four mainstream datasets.
• Achieves high accuracy (85.83% top 5, 75.44% mAP) in subsequent inference tasks on Mobilenet v1 and SSD.

Conclusions:

• The compact network offers the highest computing efficiency and lowest parameter count among compared methods.
• It is highly suitable for real-time applications on modern edge computing devices.
• The method successfully balances speed and quality for efficient image demosaicking.