This work describes the development of a compact neural network for the blind deconvolution and restoration of a blurred image. A scale-based, convolutional, long-short term memory (LSTM) network is developed for image deblurring. In this network, multi-scale information is obtained through the use of dilated convolutions and this is shared between scales using recurrent connections. The proposed network is designed to be of low-parameter count and to deblur an image without the use of prior information. We show the effectiveness of this approach through evaluation with industry standard datasets (GOPRO [2] and Kohler [10]), and we compare our results with those of two other state-of-the-art blind deblurring networks. Our results show that a comparable sharp image can be recovered more efficiently with a significant reduction in the total number of network parameters.
Paper
Full text
Image Deblurring using Multi-Scale Dilated Convolutions in a LSTM-based Neural Network
Semantic Scholar · Computer Science · 2021
Abstract
This work describes the development of a compact neural network for the blind deconvolution and restoration of a blurred image. A scale-based, convolutional, long-short term memory (LSTM) network is developed for image deblurring. In this network, multi-scale information is obtained through the use of dilated convolutions and this is shared between scales using recurrent connections. The proposed network is designed to be of low-parameter count and to deblur an image without the use of prior information. We show the effectiveness of this approach through evaluation with industry standard datasets (GOPRO [2] and Kohler [10]), and we compare our results with those of two other state-of-the-art blind deblurring networks. Our results show that a comparable sharp image can be recovered more efficiently with a significant reduction in the total number of network parameters.