Objective. Precise monitoring of the position and dwell time of iridium-192 (Ir-192) during high-dose-rate (HDR) brachytherapy is crucial to avoid serious damage to normal tissues. Source imaging using a compact gamma camera is a potential approach for monitoring. However, images from the gamma camera are affected by blurring and statistical noise, which impact the accuracy of source position monitoring. This study aimed to develop a deep-learning approach for estimating ideal source images that reduce the effect of blurring and statistical noise from experimental images captured using a compact gamma camera. Approach. A double pix2pix model was trained using the simulated gamma camera images of an Ir-192 source. The first model was responsible for denoising the Ir-192 images, whereas the second model performed super resolution. Trained models were then applied to the experimental images to estimate the ideal images. Main results. At a distance of 100 mm between the compact gamma camera and the Ir-192 source, the difference in full width at half maximum (FWHM) between the estimated and actual source sizes was approximately 0.5 mm for a measurement time of 1.5 s. This difference has been improved from approximately 2.7 mm without the use of DL. Even with a measurement time of 0.1 s, the ideal images could be estimated as accurately as in the 1.5 s measurements. This method consistently achieved accurate estimations of the source images at any position within the field of view; however, the difference increased with the distance between the Ir-192 source and the compact gamma camera. Significance. The proposed method successfully provided estimated images from the experimental images within errors smaller than 0.5 mm at 100 mm. This method is promising for reducing blurring and statistical noise from the experimental images, enabling precise real-time monitoring of Ir-192 sources during HDR brachytherapy.