Deep Image Stitching with Pixel Similarity Correlation

Image stitching is a task of synthesising images from different views without blur, seams and ghost effects. However, it is hard for existing methods to get all these visual problems solved. In this paper, we propose an image pre-reconstruction network to improve the similarity of images from different views in pixel level, which can eliminate ghost effects and correct some deviations in the image alignment step. Given the overlapping areas by warping two original images, we extract their features and calculate the distance map. According to the distance map, we classify the receptive field and define two kinds of loss functions to achieve different convergence effects. In addition, to obtain a seamless stitched image, we define a gradient loss to smooth the edge of overlapping area by narrowing the gradient between adjacent pixels. At last, we apply this gradient loss to all the image area to get a smooth effect for the whole image. Experiment shows that our method performs better than some existing methods on visual effect in the final composite.