Fisheye lens, which is suitable for panoramic imaging, has the prominent advantage of a large field of view and low cost. However, the fisheye image has a severe geometric distortion which may interfere with the stage of image registration and stitching. Aiming to resolve this drawback, we devise a stronger stitching algorithm for fisheye images by combining the traditional image processing method with deep learning. In the stage of fisheye image correction, we propose the Attention-based Nonlinear Activation Free Network (ANAFNet) to deblur fisheye images corrected by Zhang calibration method. Specifically, ANAFNet adopts the classical single-stage U-shaped architecture based on convolutional neural networks with soft-attention technique and it can restore a sharp image from a blurred image effectively. In the part of image registration, we propose the ORB-FREAK-GMS (OFG), a comprehensive image matching algorithm, to improve the accuracy of image registration. Experimental results demonstrate that panoramic images of superior quality stitching by fisheye images can be obtained through our method.