Deblurring Dynamic Scenes via Spatially Varying Recurrent Neural Networks

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Deblurring images in dynamic scenes is difficult due to spatially varying motion blurs.
• Camera shakes and object movements introduce complex blur patterns.

Purpose of the Study:

• To propose a novel spatially varying neural network for deblurring dynamic scenes.
• To develop a model capable of implicitly handling spatially varying blur kernels.

Main Methods:

• A deep learning architecture combining three Convolutional Neural Networks (CNNs) and a Recurrent Neural Network (RNN).
• The RNN acts as a spatial-aware deconvolution operator using learned spatially varying weights.
• End-to-end trainable network incorporating 1D and 2D RNN variants.

Main Results:

• The proposed spatially varying RNN effectively models deblurring with spatially varying kernels.
• Quantitative and qualitative evaluations show superior performance compared to state-of-the-art methods.
• The network successfully reconstructs deblurred images from feature maps.

Conclusions:

• The developed spatially varying neural network offers a robust solution for deblurring dynamic scenes.
• The approach demonstrates significant improvements in handling complex motion blurs.
• This method advances the field of image deblurring for real-world dynamic environments.