Ghost edge detection based on HED network.

Area of Science:

• Computational Imaging
• Machine Learning for Optics
• Image Processing

Background:

• Ghost imaging (GI) is an advanced optical technique for object reconstruction.
• Edge detection is crucial for image analysis and object recognition.
• Traditional methods often require significant data acquisition time and computational resources.

Purpose of the Study:

• To develop an efficient edge detection scheme using ghost imaging and a holistically-nested neural network (HED).
• To evaluate the performance of the proposed scheme against existing ghost imaging techniques.
• To demonstrate the feasibility of edge detection with sub-Nyquist sampling and reduced measurement costs.

Main Methods:

• Utilizing a holistically-nested edge detection (HED) network, combining fully convolutional neural networks (CNNs) with deep supervision.
• Training the HED network with simulated data for effective edge learning.
• Reconstructing edge information from experimental ghost imaging data.

Main Results:

• The proposed scheme achieved high-quality edge information at a compression ratio (CR) of 12.5% with sub-Nyquist sampling.
• Demonstrated superior performance compared to speckle-shifting GI (SSGI), compressed ghost edge imaging (CGEI), and subpixel-shifted GI (SPSGI).
• Maintained good signal-to-noise ratio performance even at sub-Nyquist sampling ratios above 5.45%.

Conclusions:

• The HED network trained via numerical simulations offers a promising approach for edge detection.
• This method significantly reduces measurement times and computational costs.
• The technique enables effective edge detection in ghost imaging under challenging sampling conditions.