Image Reconstruction Based On Vector Sparse Model

Image reconstruction has been widely used in the field of image processing,including image quality enhancement,image superresolution,and saliency detection.Recently,the image reconstruction techniques based on sparse model have achieved good performance in many applications.However,the traditional scalar sparse representation model cannot deal with multi-channel color images.It incurred the order reduction problem during color image representation,resulting in the loss of structural relationships between channels.Consequently,color distortion and structural blurring often exist in the reconstruction results.The quaternion sparse model can encode each channel of the color image as imaginary parts of a quaternion matrix,thus avoiding the order reduction problem during color image representation.According to the experimental results,quaternion sparse model can reconstruct color images more accurately and it is superior to the scalar sparse model in various applications of color image processing.In this paper,we mainly focus on the application of vector sparse model based image reconstruction in saliency detection and image superresolution.In order to overcome the shortcomings of the traditional sparse model,we propose the following innovative works:We propose a method for saliency detection based on quaternion sparse model using ensemble color features and hierarchical quaternion features.In the traditional methods,only the low-level visual features such as color and texture are used to represent objects.However,it is difficult to distinguish background pixels and salient objects in complicated environments.Meanwhile the existing methods use a background dictionary which cannot extract salient objects from the background environment completely.To overcome these problems,we improve the ways of feature extraction,dictionary formation and saliency value fusion.Firstly multi-level convolutional features are introduced to improve the characterization of image features.At the same time,multi-channel color features and multi-level convolutional features are expressed as a pure quaternion imaginary number.This vector sparse representation preserves the correlation between each feature channel,which improves the overall accuracy of saliency detection.Secondly,we compute saliency map based on the reconstruction residuals using background dictionary and foreground dictionary,which can effectively improve the recall rate of the saliency detection algorithm.Finally,we imposed global optimization on the saliency map to achieve more reasonable detection results.We evaluate the proposed method on three large publicly available datasets and the experimental results show that this method performs well with a recall rate about 10% higher than the existing methods.Meanwhile,our method achieves the competitive results in F measure comparing with the state-of-art results.Meantime,we propose an algorithm of image superresolution based on quaternion/vector sparse model.The existing superresolution methods use a single dictionary to recover the high frequency information of the high resolution images.However,the details of the image information often exist in different scale spaces,which means they cannot be completely reconstructed using a monoscale dictionary.To overcome this problem,we use an external image dataset to train a low-high resolution dictionary pair.Multi-scale reconstruction results are merged to compensate the loss of image structure information at different scales.Moreover,the quaternion sparse model is used to represent the color channels as a quaternion matrix,which avoids order reduction problem and thus retains the relationship between the channels.The high frequency components in the high resolution images can be constructed accurately.Finally,we introduce an image patch selection method to speed up our method.Only the patches with loss will be reconstructed.Experimental results show that the proposed method can effectively improve the quality of superresolution.Generally,our method can perform well with the PSNR value about 0.1dB~1dB higher than the state-of-art methods.By using the image patch selection method,the algorithm can speed up about 40% without losing the image quality.