Joint object classification and turbulence strength estimation using convolutional neural networks.

Area of Science:

• Optical remote sensing
• Machine learning applications
• Atmospheric optics

Background:

• Atmospheric turbulence degrades image quality in remote sensing.
• Prior work used multilayer perceptron networks for object classification and turbulence estimation with prior object knowledge.
• Existing turbulence simulations may lack realism for robust model training.

Purpose of the Study:

• To enhance the realism of atmospheric turbulence simulations for training and evaluating neural networks.
• To develop a novel convolutional neural network for joint character classification and turbulence strength estimation.
• To remove the constraint of requiring prior object knowledge, broadening applicability.

Main Methods:

• Developed a more realistic atmospheric turbulence simulation model.
• Designed and implemented a new convolutional neural network architecture.
• Trained and evaluated the network on degraded imagery with varying turbulence levels.

Main Results:

• The enhanced turbulence simulation improved neural network performance.
• The joint classifier-estimator network achieved accurate character classification and turbulence estimation.
• The new method successfully eliminated the need for prior object knowledge.

Conclusions:

• The developed convolutional neural network offers a more versatile approach to analyzing imagery affected by atmospheric turbulence.
• This method significantly expands the applicability of machine learning in remote sensing, especially when direct object access is limited.
• Realistic turbulence simulation is crucial for developing robust deep learning models for optical remote sensing.