Image Deblurring Method Based On Simplified Gated Activation Network

The camera will produce a large number of low-quality images under extreme or improper operation conditions, image deblurring is an effective solution, but the existing SOTA methods seek to restore spatial details, ignoring the increasing computational cost. In this paper, we propose a Simplified Gated Activation Network (SGANet), which adopts the U-Net-shaped deep auto-encoder structure as a whole, and uses skip connection between encoders and decoders to transmit and fuse the feature information of different scales. In the detail part, Gated Linear Unit (GLU) is simplified by using channel split technology and depthwise separable convolution, the nonlinear expression ability of the model is enhanced at the same time, so as to achieve a high level in balancing complexity and image deblurring performance. We use two methods to verify the effectiveness of SGANet, one is to evaluate the PSNR index of the model on the baseline dataset (GoPro), and the other is to compare the performance difference between original image and deblurred image in object detection task. Therefore, we also make a blurred images dataset in semi-structured road scene for experiment. Experimental results show that the deblurring performance of our model surpasses many excellent methods, and the computational cost is greatly reduced.