Research On Fish-eye Image Correction Method Based On Generative Adversarial Network

The fisheye lens is a kind of ultra-wide angle lens with a short focal length.The image taken by fisheye lens has barrel distortion,which can not be directly used for computer vision tasks such as object recognition,image segmentation,etc.The methods for correcting fisheye images mainly include traditional methods,convolutional neural network methods,and generation methods.The skip connection in the generation method will bring a large error,which will affect the correction effect of the fisheye image.Aiming at this problem,this paper proposes a fisheye image correction method based on generative adversarial network,which realizes the end-to-end correction of fisheye image without other auxiliary information,and the content of the corrected fisheye image is not affected by the scene limitations,the main work of this article includes:Firstly,in view of the lack of fisheye image dataset,this paper chooses a polynomial model to make fisheye images.Through experiments,select the appropriate single distortion parameters to complete the production of fisheye images.The fisheye image dataset in this paper contains the fisheye image,the background image corresponding to the fisheye image,and the distortion parameters of the fisheye image.Secondly,this paper proposes a distortion parameter attention module,which is added to the up-sampling process of the decoder to guide the correction process of the fisheye image in different spatial dimensions and reduce the error caused by skip connections.A parameter prediction network is introduced into the model of this paper to predict the distortion parameters of the fisheye image.The network shares the encoder with the generation network to reduce the amount of data in the model.Finally,experiment and comparison of the model on the fisheye image dataset in this paper.Compared with the latitude and longitude correction method and the correction method based on progressive complementary network,the correction method proposed in this paper has improved the fisheye image correction effect.The correction method in this paper has improved Peak Signal to Noise Ratio(PSNR),Structural Similarity(SSIM),and Multi-scale Structural Similarity(MSSSIM),which shows the effectiveness and advancement of the correction method in this paper.