Reserch On Image Super Resolution Algorithm Based On Sparse Representation

With the fast development of technology,the requirement of people for high image quality increases greatly.If the image quality is only developed through hardware,it reaches the bottleneck for further development and the cost is very high.Thus,researchers want to increase the image quality through the software areas and overcome the inherent resolution limits of imaging equipment,such as camera and cell phone.The super-resolution algorithms are active research areas in recent years to use the software to enlarge the image resolution.The current super-resolution algorithms can be classified into the following three areas,the ones based on interpolation,the ones based on reconstruction,and the ones based on learning.The ones based on the interpolation have low complexity and fast reconstruction.However,the they have low reconstruction quality and generally cannot meet people's requirement.The ones based on the reconstructions can relatively solve the shortcoming of the ones based on the interpolation and can relatively boost the performance.However,they have the shortcoming of having ringing artifacts in the reconstructed images and the visual quality is still not ideal.The ones based on the learning can greatly improve the performance based on the learning data sets.Thus,a lot of research has been conducted in this area.This paper proposes revised single image super-resolution algorithms based on the learning algorithms.This thesis is based on the super-resolution algorithms using sparse representatio n.It overcomes the shortcomings of the current algorithms and proposes effective solutio n schemes and boost the image quality.The main contribution is listed in the following.First,it proposes the sparse reconstruction method based on enlarged training set and the SM classification algorithm having scale invariant characteristics.This method considers that the image features retrieved from different angles may be different,and they may not have ideal training effects.Thus,it proposes to rotate the images in the training set by 90,180,and 270 degrees and flip them.The number of the training images is enlarged by 8 times to enrich the image features.To overcome the redundancies of the dictionaries,the training set is further classified into three categories as the smooth feature set,the sub-texture set and the texture feature set.The SM value which has the scale invariant characteristics is used as the measure for classification.Then,the K-SVD algorithm is used for each category to train the three dictionary pairs.For the online case of sparse reconstruction,each block is enlarged by using the corresponding dictionary pair corresponding to the class of the block and relative clear images can be obtained.Through a lot of experimental results,it can be seen that our method is better than the traditional sparse reconstruction algorithm.To overcome the problem that the current super-resolution algorithms cannot well keep the image texture information,we propose a texture constraint term in the optimal function for reconstruction.This constraint term first uses the relative total variatio n(RTV)to retrieve the image textures,and then minimize the difference of the downsampling initial high resolution image and the texture image from the low-resolut io n image to build the texture constraint term and provides a scheme for texture protection.On the other hand,the current methods process the image based on blocks and can have block artifact inherently.Considering that each block in the image may have differe nces with other blocks in the image,the adaptive total constraint method is proposed to reduce the block artifact and boost the quality of the images.Thus,two constraints are added to the original constraint.Experiments show that our revised method is effective and can greatly enhance the reconstruction quality of the images.