Abstract: The reconstruction of circular synthetic aperture sonar (CSAS) images using the time-domain back-projection algorithm introduces asymmetry in the pulse response, leading to reduced imaging resolution away from the imaging scene center. Theoretically, image blur caused by this asymmetry can be corrected by deconvolution with the point spread function (PSF) of the imaging system. However, the spatial variance in the PSF of CSAS, coupled with the ill-posed nature of deconvolution as an inverse problem, results in poor correction of image blur due to sensitivity to noise. To address the spatial variance in the PSF, this paper utilizes Implicit Neural Representation (INR) neural networks for deconvolution of underwater CSAS images. This approach effectively corrects high-order phase errors in reconstructed images and enhances the algorithm's robustness through improvements in the INR network. Computer simulations and experiments conducted on a lake demonstrate that this method outperforms traditional deconvolution methods, showcasing superior image enhancement capabilities.