Focusing on the low efficiency and poor effect of image style transfer, a feedforward residual image style transfer algorithm based on pre-trained network and combined with image texture feature analysis was proposed. In the algorithm, the pre-trained deep network was applied to extract the deep features of the style image, and the residual network was used to perform deep training and realize image transfer. Meanwhile, by analyzing the influence of input style image and content image texture on transfer effect, the corresponding measures were adopted for different input images to improve the transfer effect. Experimental results show that the algorithm can achieve better output visual effect, lower normalized style loss and less time consumption. Besides, according to the information entropy and moment invariant calculation of the input image to guide the setting and adjustment of the network parameters, the network was optimized pertinently, and good effect was obtained.

Deep neural network has powerful feature self-learning ability, which can obtain the granularity features of different levels by multi-layer stepwise feature extraction. However, when the target subject of an image has strong correlation with the background color, the feature extraction will be "lazy", the extracted features are difficult to be discriminated with low abstraction level. To solve this problem, the intrinsic law of feature extraction of deep neural network was studied by experiments. It was found that there was correlation between feature extraction preference and background color of the image. Eliminating this correlation was able to help deep neural network ignore background interference and extract the features of the target subject directly. Therefore, a data enhancement algorithm was proposed and experiments were carried out on the self-built dataset. The experimental results show that the proposed algorithm can reduce the interference of background color on the extraction of target features, reduce over-fitting and improve classification effect.

How to improve the accuracy of recommendation is an important issue for the current recommendation system. The matrix decomposition model was studied, and in order to exploit the group structure of the rating data, a Score Similarity based Matrix Factorization recommendation algorithm with Group Sparsity (SSMF-GS) was proposed. Firstly, the scoring matrix was divided into groups according to the users' rating behavior, and the similar user group scoring matrix was obtained. Then, similar users' rating matrix was decomposed in group sparsity by SSMF-GS algorithm. Finally, the alternating optimization algorithm was applied to optimize the proposed model. The latent item features of different user groups could be filtered out and the explanability of latent features was enhanced by the proposed model. Simulation experiments were tested on MovieLens datasets provided by GroupLens website. The experimental results show that the proposed algorithm can improve recommendation accuracy significantly, and the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) both have good performance.

Concerning salt-and-pepper noise removal and details protection, an image denoising algorithm based on switching kernel regression fitting was proposed. Firstly, the pixels corrupted by salt-and-pepper noises were identified exactly by efficient impulse detector. Secondly, the corrupted pixels were take as missing data, and then a kernel regression function was used to fit the non-noise pixels in a neighborhood of current noisy pixel, so as to obtain a kernel regression fitting surface that met local structure characteristics of the image. Finally, the noisy pixel was restored by resampling of the kernel regression fitting surface in terms of its spatial coordinates. In the comparison experiments at different noise densities with some state-of-the-art algorithms such as Standard Median Filter (SMF), Adaptive Median Filter (AMF), Modified Directional-Weighted-Median Filter (MDWMF), Fast Switching based Median-Mean Filter (FSMMF) and Image Inpainting (Ⅱ), the proposed scheme had better performance in subjective visual quality of restored image. At low, medium and high noise density levels, the average Peak Signal-to-Noise Ratio (PSNR) of different images by using the proposed scheme was increased by 6.02dB, 6.33dB and 5.58dB, respectively; and the average Mean Absolute Error (MAE) was decreased by 0.90, 5.84 and 25.29, respectively. Experimental results show that the proposed scheme outperforms all the compared techniques in removing salt-and-pepper noise and preserving details at various noise density levels.

Concerning highly accurate sub-pixel motion vector estimation, an accurate motion estimation algorithm based on upsampled phase correlation with kernel regression refining was proposed. Firstly, an upsampled phase correlation was computed efficiently by means of matrix-multiply discrete Fourier transform, and the initial estimation of motion vector with sub-pixel accuracy was achieved by simply locating its peak. Secondly, a kernel regression function was fit to the upsampled phase correlation values in a neighborhood of initial estimation. Finally, the initial estimation was refined with the location of peak found in the kernel regression fitting function, so as to obtain accurate estimation at arbitrary-precision. In the comparison experiments with some state-of-the-art algorithms such as Quadratic function Fitting (QuadFit), Linear Fitting (LinFit), Sinc Fitting (SincFit), Local Center of Mass (LCM) and Upsampling in the frequency domain (Upsamp), the proposed scheme achieved the average estimation error at 0.0070 in the case of noise-free, and increased the accuracy of motion estimation by more than 64%; while under the noise condition, the average estimation error of the proposed shceme was 0.0204, and the accuracy of motion estimation was improved by more than 47%. Experimental results show that the proposed scheme can not only improve the accuracy of motion estimation significantly, but also achieve good robustness to the influence of noise.

A Combined Prediction Scheme (CPS) and a concept of Prediction Accuracy Assurance (PAA) were put forward for the runtime of local and remote tasks, on the issue of inapplicability of the singleness policy to all the heterogeneous tasks. The toolkit of GridSim was used to implement the CPS, and PAA was a quantitative evaluation standard of the prediction runtime provided by a specific strategy. The simulation experiments showed that, compared with the local task prediction strategy such as Last and Sliding Median (SM), the average relative residual error of CPS respectively reduced by 1.58% and 1.62%; and compared with the remote task prediction strategy such as Running Mean (RM) and Exponential Smoothing (ES), the average relative residual error of CPS respectively reduced by 1.02% and 2.9%. The results indicate that PAA can select the near-optimal value from the results of comprehensive prediction strategy, and CPS enhances the PAA of the runtime of local and remote tasks in the computing environments.

To meet the requirements of multi-directional choice, a new approach to the invariant feature extraction of handwritten amount Chinese characters was raised, with Ridgelet transform as its foundation. As far as this approach is concerned, first of all, the original images would be rotated to the Radon circular shift by means of Radon transform. On the basis of the characteristic that Fourier transform is row shift invariant, then, the one-dimensional Fourier transform would be adopted in Radon field to gain the conclusion that magnitude matrixes bear the rotation-invariance as a typical feature, which was pretty beneficial to the invariant feature extraction of rotation. When this was done, one-dimensional wavelet transform would be carried out in the direction of rows, thus achieving perfect choice of frequency, which made it possible to extract the features of sub-line in the appropriate frequencies. Finally, the average values, standard deviations and the energy values would form the feature vector which was extracted from the Ridgelet sub-bands. The approaches mentioned in the paper could satisfy the requirements from the form automatic processing on the recognition of handwritten amount Chinese characters.

The paper analyzed the current measurement technologies and the existing localization methods of bottleneck bandwidth. Based on the analysis and comparison of their advantages and disadvantages, the paper put forward a new method of bottleneck bandwidth measurement and localization on the basis of meltable packet train method. The method can measure bottleneck bandwidth and localize it at the same time, which reduces the times and load of measurement effectively. In addition, the paper also proves the validity of the method from the theory and simulation experiment, and points out its advantages compared with other methods.

Set partitioning in hierarchical trees (SPIHT) was used to realize region of interest (ROI) image coding. Based on the characteristics of SPIHT, the properties of image wavelet coefficients after the coefficients associated with ROI were scaled up, and human visual specialties, a parameter of image quality (Q) and its formula were presented. Based on this, an improved algorithm was put forward, which could improve the subjective quality of image meanwhile ensure the quality of ROI. The experimental results show that the proposed method is effective and the formula of Q is correct, and we can choose Q to control the quality of ROI.

A multiple descriptions transform coding algorithm based on wavelet (WMDTC) is proposed. We first partitioned the wavelet coefficients of the highest and lowest pyramid levels into two sets. A pairwise correlating transform (PCT) was applied to each pair to generate two new sets of coefficients that were correlated between the sets. Coefficients in each set were encoded and decoded independently. Each description was then transmitted through separate channels. If only one channel was working, it was possible to estimate the lost description from another. We combined WMDTC with interpolation to improve reconstructed image. The algorithm performs well in performance simulations.