Image Resizing Algorithms And Their GPU Implementations

Image resizing is an important and basic operation in image processing, which is widely applied in military reconnaissance, medical image processing, astronomical observation, digital photographing, internet multimedia, etc. Image resizing changes image resolution via resampling the source image according to color information and other features, also known as image resampling or image interpolation.The paper firstly reviews current image resizing approaches, which includes interpolation algorithms based on kernel functions, nonlinear algorithms based on local correlations, machine learning and neural network based algorithms, triangulation based algorithms, vectorization algorithms, etc. In summary, the assessment criteria of the image resizing algorithms includes the following three aspects: the first one is to keep the continuity among neighboring pixels, the second one is to maintain the important features in the resized image such as sharp edges, detail textures etc., the last one is the efficiency. Thus the thesis dedicates to investigate high quality and efficiency image resizing algorithms via Graphics Processing Unit, i.e. GPU.Two algorithms are proposed in this paper. One is to zoom in an image 2 times via nonlinear correlation detection. It can preserve image's continuity and other features better by fully detecting image's correlation directions and considering all the potential correlations up to quadratic along the sampling directions. The second algorithm is to resize an image based on biquadratic B-spline local interpolation. It first reconstructs the image locally (4x4 pixels) by using a biquadratic B-spline function. Due to the symmetry of neighboring pixels, the algorithm avoids the phase distortion in conventional biquadratic interpolations. Finally we can obtain the resized image by resampling the locally reconstruction functions. The proposed two algorithms are local without coupling. Thus they can be accelerated via GPU. Implementation results show that GPU acceleration can achieve 1-2 orders improvements compared with their CPU counterparts.