Multi-channel Image Reconstruction Based On Quaternion Sparse Representation

Sparse representation has been widely used in image processing,such as superresolution image reconstruction and object detection.However,the traditional sparse representation models concatenate the channel images,incurring color distortion and structure degeneration.In order to solve these problems,quaternion sparse representation model was proposed.The quaternion-based methods process multichannel signal in a parallel way,which mimics the human perception of a visual environment.This thesis will extend quaternion sparse representation model into two applications-brain atlas template construction and saliency detection.The major contributions in this work are listed as follow:We proposed an atlas construction model based on quaternion sparse representation.The input of the infant brain magnetic resonance images of intensity image,gray matter map,white matter map and cerebrospinal fluid map are represented by quaternion model.It can enhance the detail of reconstructed image.Experiment shows that our algorithm has the highest consistency in the description of the structure of the brain.In addition,we proposed a saliency detection model based on quaternion sparse reconstruction residual analysis.We reconstruct current image patch by dictionaries obtain from its neighbors patch and background patch,then calculate reconstruction residual as saliency value.This method considers both local and global contrast and uses quaternion sparse model to obtain more accurate reconstruction result.By this way,our model reduces reconstruction error and improves the accuracy of saliency detection result.Experimental results show that our algorithm has a high consistency with the eye movement data and can effectively locate the salient object.