Algorithms Of Image Feature Extraction In Image Sparse Restoration

With high quality image widely used in varieties of fields such as industry and daily life, image restoration based on sparse representation method attracts more attention since it overcomes inherent resolution limitation of imaging sensors. In many applications, only one low-resolution image is available. Therefore, sparse representation based method is applied to solve this ill-posed reconstruction problem by introducing joint dictionaries. For dictionary, landmark atoms provide representations of images. In this thesis, a new feature extractor is proposed to extract representative feature maps. Representative feature maps lead to more representative dictionary atoms. This thesis mainly includes the following contents:First, characteristics of image features used in image restoration are described. In image restoration, much information is lost in high resolution image to low resolution image generation. High frequency information in low resolution image is important in sparse representation, since it is needed for predicting lost high frequency information in high resolution image. Some kind of low-level visual feature, such as edge and texture features, represents high frequency information. First and second order differential operator, wavelet decomposition and feature extractors got from convolutional neural network is introduced in detail. These kind of extractors can extract feature maps containing edge and texture features.Next, sparse representation of image patches is introduced in detail. Image restoration is patch-wise based method. Image patches can be represented as a sparse linear combination of atoms in an appropriate overcomplete dictionary. In compressed sensing, the sparse representation can be recovered from the downsampled signal under kind condition. Low resolution image patches are downsampled version of high resolution image patches. The sparse representation recovered from low-resolution image patches is used to reconstruct high resolution image patches with high resolution dictionary.In image restoration,feature of image are used as prior information. First,total variation deconvolution in image reconstruction is introduced. And then in image restoration based sparse representation, first and second order gradient maps in horizontal and vertical direction are used as feature maps. Secondly, interpolation method using wavelete decomposition is introduced and wavelete decomposition of image in horizontal, vertical and diagonal direction is also used in dictionary learning for sparse coding. To compare these two kinds of feature maps in image restoration, some experiments are conducted and results are compared in objective criterion and visual criterion.Finally, since existing feature extractors are universally applicable in feature extraction. General feature extractors are not able to extract representative features. To overcome these drawbacks, data-dependent feature extractor(DDFE) is proposed, which are derived from training image datasets. In this thesis, the DDFE is obtained from a convolutional neural network and able to extract representative features from lowresolution images. From experimental results, it is proved that representative features extracted by the DDFE play an important role in learning a good dictionary, aiming to improve image super resolution results. Finally, experiments are conducted on natural image set and the results are compared with other feature extractors. Experimental results show that data-dependent feature extractors perform better than other feature extractors.