Abstract:
High-resolution images are a fundamental requirement of modern imaging applications. However, sensor hardware can only go so far in its ability to capture high density images of a scene. Postprocessing algorithms, such as the super-resolution technique, can be used to enhance image quality after capture, thus overcoming the limits of sensor hardware. The super-resolution technique can be approached with different methods, including interpolation, learning and reconstruction. Each of these is best suited for different regions of a scene. Interpolation performs well on smooth areas of an image, while learning is better at enhancing edges and reconstruction will be more appropriate for a heavily textured region. This research investigates the question of combining the strengths of these methods, in order to provide more accurate high-resolution approximations of a scene. The system performs feature detection in order to determine the prominent features of the input images. Depending on the presence of different features, the hybrid system will choose which super-resolution algorithms to use in the construction of the final output image. The interpolation algorithm�s fast computation speed and simplicity made it the perfect candidate for the processing of the simplest identified feature - smooth regions. The learning algorithm�s greater complexity and ability to enhance high frequency information efficiently allowed it to be used in cases where many edges were present in an image. The reconstruction algorithm�s ability to deal with high levels of noise and blur in an image made it useful when dealing with features that contained windows of rapidly changing pixel intensity. Feature detection made the combination of these algorithms possible and by analysing different algorithms� behaviour when faced with different features, different combinations of algorithms were able to be applied appropriately to maximise performance and accuracy, while minimising computational complexity. The super-resolution image reconstruction technique is capable of overcoming the inherent limitations of an imaging system and can improve the resolving quality of low-resolution images. Different methods and approaches to this technique have different practical and theoretical applications. The advancement of knowledge in this field will make it possible for image processing researchers to not only further investigate the use of multiple images in high-resolution imaging systems, but also make imaging technology affordable in a technologically advancing world.
