การฟื้นฟูภาพถ่ายดิจิตอลด้วยการตรวจสอบตัวเองของการเรียนรู้เชิงลึก

Abstract

Noise occurrence and image quality loss are crucial issues in image processing. They can occur from the surrounding environment or the electric signal in the camera device. They result in the degradation of image quality and might lead the image processing to errors. In addition, rephotographing is not feasible for obtaining new usable images in some situations, such as single image problems and blind noise scenarios. Therefore, image denoising and restoration of damaged are essential for image processing tasks. Deep learning network is a successful learning-based method with excellent performance that has resulted in much research; nevertheless, it requires more training data. Therefore, it is unsuitable for applications without training data. This research aims to develop a deep learning framework that can be used in the field that lacks validation data for training. This research divides the experiments into 3 topics. The first experiment is an edge-perceptual loss for image denoising which exploits the edge loss and pixel-wise loss for training denoising networks. While this training strategy can increase the performance of the denoising network, it still requires more image data for training. The second experiment presents a self-validation Noise2Noise (SV-N2N) framework to solve the insufficient training dataset for deep learning. SV-N2N can self-train without validation of a noise-free image. The third experiment employs image downsampling and upsampling to eliminate noise and self-validation to restore image resolution. The outcomes demonstrate that the proposed approach can remove and recover image detail from noise and poor resolution.