Dynamic Image Deblurring based on Crosshatch Attention Adversarial Network and Hybrid Loss

Image deblurring techniques that uses deep learning have shown great potential but due to low generalizability, noise immunity and the correlation among different pixels is not addressed in detail that results in unwanted artifact that appears in the deblurred image. To tackle this problem an end-to-end approach is proposed for the recovery of sharp image from blurred one without the estimation of blur kernel. A special type of attention module known as crosshatch attention is used after Residual Block of Generator model for removing noise and for the collection of correlation of different pixels in an image. Hybrid Loss function is defined which focus on different part of image and improve edges and texture details. The performance of the model for deblurring is measured on GoPro dataset. Our proposed model has slightly higher objective and subjective evaluation i-e PSNR, SSIM value and the visual results.