Abstract
In this paper, we propose a feature extraction of panoramic images. Feature extraction directly on the sphere will cause a large amount of feature information loss. In order to solve the problem of converting panoramic images from spherical space to two-dimensional planes, the pre-processing of panoramic images is achieved by using equirectangular projection. The use of equirectangular projection can project panoramic images on the sphere to the two-dimensional plane, which is convenient for feature extraction and reduces the loss of feature information. The feature extraction method first broadens the number of channels of pre-processed panoramic images, through focus network to facilitate subsequent feature extraction. Then panoramic images are implemented with swin-transformer network for feature extraction. Which is able to achieve feature extraction by splitting images into windows and layers, exchanging information between different windows to improve the integrity of feature information, and being able to obtain three effective feature layers. Finally, feature fusion and feature information enhancement are performed on the three feature layers through the feature pyramid network to output the final feature information. The experimental results show that the proposed method has good effect on panoramic images, and the feature information of panoramic images can be extracted effectively.