| This dissertation is concerned with model designation andalgorithm implementation of the application of compressed sensing and itsrelated theories on image recognition and image flow recovery.First of all, image recognition problems under large-scale and small-scale training sample sets are discussed in this dissertation, together withthe one with its test sample in extreme cases, such like blocking, shift etc.,respectively. Different combination strategies and improvements of fea-ture extractor and classifier are put forward for these three different cases.Experimental results demonstrated the feasibility and effectiveness of threeproposed algorithms.Further, in view of so many disadvantages of traditional BP algorithm,such like its sensitiveness to the initial weights, slow convergence, highpossibility of over fitting and the selection problem of number of hiddenlayer nodes, we introduced the conceptions of sparsity of neural networkhidden layer nodes and group-sparsity of multi-kernel neural network hid-den layer nodes. Base on these, sparse coded random weighted networksalgorithm and sparse coded multiple kernel random weighted networks al-gorithm are proposed. Experimental results confirmed the existence of thesparsity of neural network hidden layer nodes and group-sparsity of multi-kernel neural network hidden layer nodes, and verified a fact that our pro-posedalgorithmscaneffectivelyavoidtheover-fittingphenomenon, whichis usually caused by the complexity of neural network structure, and havean outstanding performance when the optimal number of hidden layer n-odes is unclear.At the end of this paper, K-SVD algorithm is promoted to a tensor single value decomposition algorithm. In addition, based on T-SVD, animage flow recovery algorithm is proposed. The algorithm can realize, tosome extent, effective recovery of damaged image sequence. Experimen-tal results demonstrated the feasibility and effectiveness of three proposedalgorithms. |
PDF Full Text Request