EAIDGAN: Edge and Attention-Based Image Deblurring by Generative Adversarial Network

Abstract

The image is a representation of something or thought we're imagining in our heads. Since one image tells a hundred words peoples uses an image to exchange information on different communication platform while, data scientist uses the image to investigate research on object detection and classification, action recognition, image generation, image enhancement and for other computer vision applications. To use an image for exchanging information and in computer vision applications the image must have good quality. However, image quality is frequently degraded by factors such as blur. Blur is a part of artifacts that occur during the image acquisition time. An image deblurring mechanism is required before using the blurry image for computer vision applications. Image deblurring is an essential and challenging low-level vision issue that aims to get a sharp image from the blurred image. In the past several works have been conducted on image deblurring by using the generative adversarial network on the paired dataset. However, those methods cannot deliver satisfactory outcomes when the blur size is large since they rely only on local convolution filters. In this work edge and attention-based image deblurring by the generative adversarial network is proposed to mitigate these issues. The architecture utilizes a feature pyramid network as the core building block of the generator since it can simply integrate with many pre-trained backbones. To improve the execution of the network with small computational complexity EfficientNet-B7 is utilized as a feature extractor in the generator network. Further, to handle a large size blur kernel without increasing parameters the channel and self-attention mechanism was utilized in this work to capture long-range dependencies between channels and pixels instead of relying only on local convolution filters. Moreover, Edge and least square loss are introduced to preserve the edge information and to stabilize the training respectively. Broad subjective and quantitative evaluations illustrate the striking victory of the proposed system against existing methods. The peak signal to noise ratio and structural similarity index measure reaches over 26.174 and 0.8406 respectively which outperforms the other comparative works. Average human assessment study has appeared that this research exceeds expectations at 25% compared to DeblurGAN-v2. This work concludes that adding EfficientNet B7, self-attention, channel attention, edge, and least square loss does improve the execution of the deblurring network.