Study On The Artificial Neural Network Based Multi-frame Super-resolution Method

With the development of electronic products, consumers and academia are more and more pickier about the quality of digital images. Resolution is one of the most important index to evaluate the quality of pictures. With higher resolution, one picture can contain more detailed information and offer better view experience. Superresolution(SR) is the technique to enhance digital image resolution without changing hardware system. The dissertation studies on digital images superresolution by using artificial neural network(ANN), which belongs to a learning-based superresolution algorithm.The dissertation explore and discuss the key problems for Artificial Neural Network based superresolution, such as net training samples data, net architecture and net training method. In addition, optimization methods such as Genetic Algorithm(GA) and image classification are also discussed in this dissertation. The main works are as follows:① A new block search strategy which is based on Motion Estimation(ME) and local similarity is implemented. With this method, the performance of superresolution is enhanced and the computation is reduced.② A relatively ideal combination of transfer functions for every layer of the Artificial Neural Network is found.③ The reflection of different net training methods in the proposed method is found.④ The feasibility of classification in our method is discussed.⑤ Genetic algorithm is used to enhance the performance of Artificial Neural Network based superreslution.The research of this dissertation is an attempt to digital image superresolution, which has academic value and practicality significance.