Deep learning-based object classification through multimode fiber via a CNN-architecture SpeckleNet.

Area of Science:

• Optics
• Machine Learning
• Image Processing

Background:

• Deep learning shows significant advancements in image classification and object recognition.
• These advancements offer potential for analyzing speckle patterns in scattering media imaging.

Purpose of the Study:

• To propose and evaluate a novel deep learning model, SpeckleNet, for classifying face and nonface speckle patterns.
• To assess the performance improvement of SpeckleNet compared to traditional methods.

Main Methods:

• Utilizing a multimode fiber as scattering media to generate speckle patterns from 4000 original images.
• Developing and training SpeckleNet, a convolutional neural network (CNN) architecture, with 3600 speckle patterns.
• Integrating a support vector machine (SVM) classifier with the CNN's output layer.

Main Results:

• SpeckleNet achieved a binary classification accuracy of approximately 96% for face and nonface speckle patterns.
• This accuracy surpasses the performance of a pure SVM method.
• The trained model was tested on a separate set of 400 speckle patterns.

Conclusions:

• The combination of deep learning with optical sensing significantly improves speckle pattern classification.
• SpeckleNet demonstrates the potential to lower optical and computational costs in optical sensing applications.
• The findings suggest practical applications in the field of optics.