Fig. 1: HitchLearning paradigm mechanism.

a Comparison of the traditional and HitchLearning paradigms. The traditional machine learning paradigm has two stages: the first stage (upper left) uses a large amount of training data to create a well-performing model through various learning methods, and the second stage (upper right) uses this trained model to infer a restored output. Our new paradigm, shown at the bottom, unifies these stages by incorporating inference data during training. The two-stage framework faces a non-i.i.d. problem due to the domain gap between source and target data (middle left). In HitchLearning, this problem is alleviated by using inference data to generate new source data during model training. b The architecture of the HitchLearning paradigm. Domain alignment (blue) can utilize any method to align the distribution of source and target images. The image restoration learning method can be any approach (supervised, self-supervised, or unsupervised) suitable for tasks such as image denoising, deblurring, and super-resolution. The image restoration model is a well-trained model with strong performance on a single target image. The blue dotted box illustrates our method’s domain alignment in the Fourier domain, which requires only one target image, unlike commonly used DNN methods that need multiple target data, increasing costs. The black dotted box displays results of three tasks using our paradigm, with the bar chart at the bottom showing the PSNR comparison of two main training paradigms against HitchLearning across these tasks. c A simple but effective distribution alignment method. Here, we use Fourier Domain Adaption (FDA) method to obtain domain alignment, by doing this, the source image has a similar style to the target and can be verified by the data shown in the Results.