In many visual applications, the HR is generally required and preferred for producing more detailed information inside the digital images. Therefore, this leads to the improvement of pictorial information for human analysis and interpretation and also for automatic machine perception (Köhler et al., 2016). Unfortunately, the real imaging systems may bring about several degradations or artifacts in the digital images. The main reason of such distortions is maintained in a variety of factors such as blurring, aliasing, and noise. Such factors may affect the resolution of imaging systems and produce LR images. Consequently, the LR images have a serious shortage in the stored information which, in turn, they will affect the quality of human interpretation and automatic machine perception. As a result of these factors, many researchers have implemented diverse methods in an effort to produce a high quality image based on the SR image reconstruction approaches (Laghrib, Ghazdali, Hakim, & Raghay, 2016; Nayak & Patra, 2016; Shen et al., 2016). Therefore, a general comparison between frequency and spatial domain approaches is shown in Table 1. In addition, Table 2 shows the comparative analysis of a common multi-frame SR approaches.In general, the frequency domain approaches have many problems which prohibited researchers from an advanced development, especially in case for the sensitivity of model errors and difficulty in dealing with more complex motion models (Begin & Ferrie, 2006; Hadhoud, El-Samie, & El-Khamy, 2004; Papathanassiou & Petrou, 2005; Patanavijit, 2009; Yang & Huang, 2010; Yue et al., 2016). Due to the limitations of the frequency domain approaches, the spatial domain approaches are classified as the most popular approach adopted to develop the SR image. The popularity of these approaches springs from the fact that the motion is not restricted to the translational shifts only; therefore, a more general global or non-global motion may also be integrated and managed (Park et al., 2003).The spatial domain approaches are usually split into interpolation-based approaches and regularization-based approaches (Park et al., 2003). However, interpolation-based approaches regularly generate images with several drawbacks around the object’s borders, consisting of zigzag, blurring, and aliasing edges. Therefore, regularization-based approaches are a challenging in SR image reconstruction (Hadhoud et al., 2004).
Generally, the SR image reconstruction approaches are really depicted as an ill-posed problem because of the inadequate number of LR images and the ill-conditioned blur operators. Techniques which usually manipulate to support the inversion of the ill-posed problem are identified as regularization. Regularization approaches take advantage of the prior knowledge of the unidentified HR image to resolve the SR problem (Yue et al., 2016). Thus, the regularization approach can be used as an attempt to stabilize the inversion process and compensate the missing information. Additionally, it is used to represent a prior of image, eliminate artifacts from the image, provide the prior information to generate a stable solution, enhance the convergence rate, and include artificial constraints on the solution such as smoothness and edge preserving (El Mourabit, El Rhabi, Hakim, Laghrib, & Moreau, 2017; Yue et al., 2016). Nevertheless, most of the regularization approaches that have been developed in the literatures are still suffering from an imbalance between edges preservation and noise suppression inside the reconstructed HR image (L. Wang, Lin, Deng, & An, 2017). In which, if the noise is completely eliminated from the reconstructed HR image, this leads to smoothness in the edges. Moreover, if the edges are preserved well in the reconstructed HR image, this leads to suffering from the image noise (El Mourabit et al., 2017; Kiani & Drummond, 2017; Long, Lu, Shen, & Xu, 2017; Mohan, 2017; L. Wang et al., 2017).
In this paper, a general survey of the present multi-frame SR approaches is presented over the previous three decades. The primary improvement of SR approaches can essentially be split into three phases. In the first 10 years phase, researchers shift their focus from the study of frequency domain to spatial domain approaches, especially interpolation-based approaches. In second phase, regularized SR approaches acquire a primary emphasis. Within the last phase, the Bayesian MAP construction has become the most common approach due to its great performance and flexible properties. Recently, researchers have primarily focused on SR reconstruction in several areas. Nevertheless, the comprehensive practical use of SR still continues to be described as problematic.