In order to better meet the accuracy and timeliness requirements of Chinese food dish recognition， a new type of dish recognition network was designed. The original YOLOv5 model was pruned by combining Supermask method and structured channel pruning method， and lightweighted finally by Int8 quantization technology. This ensured that the proposed model could balance accuracy and speed in dish recognition， achieving a good trade-off while improving the model portability. Experimental results show that the proposed model achieves a mean Average Precision （mAP） of 99.00% and an average recognition speed of 59.54 ms /frame at an Intersection over Union （IoU） of 0.5， which is 20 ms/frame faster than that of the original YOLOv5 model while maintaining the same level of accuracy. In addition， the new dish recognition network was ported to the Renesas RZ/G2L board by Qt. Based on this， an intelligent service system was constructed to realize the whole process of ordering， generating orders， and automatic meal distribution. A theoretical and practical foundation was provided for the future construction and application of truly intelligent service systems in restaurants.

Since current available sleep electroencephalogram data sets for sleep staging are all class imbalanced small data sets, it is hard to achieve ideal staging result by directly migration application of deep learning models. A deep automatic sleep staging model for class imbalanced small data sets was proposed, from the aspect of data oversampling and model training optimization. Firstly, a Modified Synthetic Minority Oversampling TEchnique (MSMOTE) was improved from the perspective of reducing the decision region, and the new technique was applied to generate the minority class samples in the original data sets. Then, the reconstructed class balanced data sets were used to pre-activate the sleep staging model. The 15-fold cross-validation experiment showed the overall classification accuracy was 86.73% and the macro-averaged F1-score was 81.70%. The value of F1 for the minimum class increased from 45.16% to 53.64% by using the data sets reconstructed by improved MSMOTE, to pre-activate the model. In conclusion, the model can realize the end-to-end learning for raw sleep electroencephalogram signals. It has a higher classification efficiency by comparison with recent advanced research and is suitable for the portable sleep monitors that work in conjunction with remote servers.

Due to critical security vulnerabilities of the protocols for Short Message Service (SMS), SMS spam come to the fore through numerous malicious pseudo base-stations, to spread fraud message or illegal advertisements. Nowadays, SMS spam negatively affects daily lives of the masses, even influences the stability of society. However, with respect to the properties as mobility and concealment of pseudo base-stations, exploring the trajectory and activity of pseudo base-stations is a difficult task. To solve this problem, a visual analytics scheme was proposed to trail pseudo base-stations via multi-users' SMS spam collected by mobile service provider. Multi-visualized views and a visual analytics system were designed based upon the proposed scheme. Moreover, a case study was presented to validate the proposed method and system, with the aid of dataset provided by the ChinaVis'2017 Challenge I. The result verifies the feasibility and effectiveness of the proposed method.

It is hard for existing image dehazing method to meet the demand of real-time because of high complexity, thus a fast image dehazing method combined with negative correction and contrast stretching transform was proposed to enhance the contrast and saturation of haze images. Contrast stretching transform was employed to negative image of input image to enhance the contrast, which saved computing time. Adaptive parameter was set for structure information got via Lipschitz exponent, it was composed of Lipschitz exponent and mean average function of local pixel block. Finally, the corresponding haze removed image with nature color and clear details was obtained by using Sigmoid function to stretch the image adaptively. The experimental results demonstrate that the proposed method can achieve a good subjective visual effect while ensuring the real-time performance, and meet the requirements of practical engineering applications.

Concerning the streak artifacts and impulse noise of the Low-Dose Computed Tomography (LDCT) reconstructed images, a statistical iterative reconstruction method based on adaptive weighted Total Variation (TV) for LDCT was presented. Considering the shortage that traditional TV may bring staircase effect while suppressing streak artifacts, an adaptive weighted TV model that combined the weighting factor based on weighted variation and TV model was proposed. Then, the new model was applied to the Penalized Weighted Least Square (PWLS). Different areas of the image were processed with different de-noising intensities, so as to achieve a good effect of noise suppression and edge preservation. The Shepp-Logan model and the digital pelvis phantom were used to test the effectiveness of the proposed algorithm. Experimental results show that the proposed method has smaller Normalized Mean Square Distance (NMSD) and Normal Average Absolute Distance (NAAD) in the two experiment images, compared with the Filtered Back Projection (FBP), PWLS, PWLS-Median Prior (PWLS-MP) and PWLS-TV algorithms. Meanwhile, the proposed method get Peak Signal-To-Noise Ratio (PSNR) of 40.91 dB and 42.25 dB respectively. Experimental results show that the proposed algorithm can well preserve image details and edges, while eliminating streak artifacts effectively.

A new denoising algorithm, Adaptive Total Generalized Variation (ATGV), was proposed for removing streak artifacts within the reconstructed image of low-dose Computed Tomography (CT). Considering the shortage that the traditional Total Generalized Variation (TGV) would blur the edge details, the intuitionistic fuzzy entropy which can distinguish the smooth and detail regions was introduced into the TGV algorithm. Different areas of the image were processed with different denoising intensities. As a result, the image details could be well preserved. Firstly, the Filtered Back Projection (FBP) algorithm was used to obtain a reconstructed image. Secondly, the edge indicator function based on intuitive fuzzy entropy was applied to improve the TGV algorithm. Finally, the new algorithm was employed to reduce the noise in the reconstructed image. The simulations of the low-dose CT image reconstruction for the Shepp-Logan model and the thorax phantom were used to test the effectiveness of the proposed algorithm. The experimental results show that the proposed algorithm has the smaller values of the Normalized Mean Square Distance (NMSD) and Normalized Average Absolute Distance (NAAD) in the two experiment images, compared with the Total Variation (TV) algorithm and TGV algorithm. Meanwhile, the two experiment images processed with the new method can obtain high Peak Signal-to-Noise Ratios (PSNR) of 26.90 dB and 44.58 dB, respectively. So the proposed algorithm can effectively preserve image details and edges, while reducing streak artifacts.

Many traditional machine learning methods tend to get biased classifier which leads to lower classification precision for minor class in sequential imbalanced data. To improve the classification accuracy of minor class, a new hybrid sampling online extreme learning machine on sequential imbalanced data was proposed. This algorithm could improve the classification accuracy of minor class as well as reduce the loss of classification accuracy of major class, which contained two stages. In offline stage, the principal curve was introduced to model the confidence regions of minor class and major class respectively based on the strategy of balanced samples. Over-sampling of minority and under-sampling of majority was achieved based on confidence region. Then the initial model was established. In online stage, only the most valuable samples of major class were chosen according to the sample importance, and then the network weight was updated dynamically. The proposed algorithm had upper bound of the information loss through the theoretical proof. The experiment was taken on two UCI datasets and the real-world air pollutant forecasting dataset of Macao. The experimental results show that, compared with the existing methods such as Online Sequential Extreme Learning Machine (OS-ELM), Extreme Learning Machine (ELM) and Meta-Cognitive Online Sequential Extreme Learning Machine (MCOS-ELM), the proposed method has higher prediction precision and better numerical stability.

Based on open source softwares of Computer Haptics, visualizAtion and Interactive in 3D (CHAI 3D) and Open Graphic Library (OpenGL), a virtual surgical system was designed for reduction of maxillary fracture. The virtual simulation scenario was constructed with real patients' CT data. A geomagic force feedback device was used to manipulate the virtual 3D models and output haptic feedback. On the basis of the original single finger-proxy algorithm, a multi-proxy collision algorithm was proposed to solve the problem that the tools might stab into the virtual organs during the simulation. In the virtual surgical system, the operator could use the force feedback device to choose, move and rotate the virtual skull model to simulate the movement and placement in real operation. The proposed system can be used to train medical students and for preoperative planning of complicated surgeries.

Many traditional machine learning methods tend to get biased classifier which leads to low classification precision for minor class in imbalanced online sequential data. To improve the classification accuracy of minor class, a new weighted online sequential extreme learning machine based on imbalanced sample-reconstruction was proposed. The algorithm started from exploiting distributed characteristics of online sequential data, and contained two stages. In offline stage, the principal curve was introduced to construct the confidence region, where over-sampling was achieved for minor class to construct the equilibrium sample set which was consistent with the sample distribution trend, and then the initial model was established. In online stage, a new weighted method was proposed to update sample weight dynamically, where the value of weight was related to training error. The proposed method was evaluated on UCI dataset and Macao meteorological data. Compared with the existing methods, such as Online Sequential-Extreme Learning Machine (OS-ELM), Extreme Learning Machine (ELM)and Meta-Cognitive Online Sequential- Extreme Learning Machine (MCOS-ELM), the experimental results show that the proposed method can identify the minor class with a higher ability. Moreover, the training time of the proposed method has not much difference compared with the others, which shows that the proposed method can greatly increase the minor prediction accuracy without affecting the complexity of algorithm.

Since the face images might be not over-complete and they might be also corrupted under different viewpoints or different lighting conditions with noise, an efficient and effective method for Face Recognition (FR) was proposed, namely Robust Principal Component Analysis with Collaborative Representation based Classification (RPCA_CRC). Firstly, the face training dictionary D₀ was decomposed into two matrices as the low-rank matrix D and the sparse error matrix E; Secondly, the test image could be collaboratively represented based on the low-rank matrix D; Finally, the test image was classified by the reconstruction error. Compared with SRC (Sparse Representation based Classification), the speed of RPCA_CRC on average is 25-times faster. Meanwhile, the recognition rate of RPCA_CRC increases by 30% with less training images. The experimental results show the proposed method is fast, effective and accurate.

Inspired by prior knowledge of face images' approximate symmetry, an algorithm based on symmetric Gabor features and sparse representation was proposed, which was successfully applied into face recognition in the paper. At first, mirror transform was performed on face images to get their mirror images, with which the face images could be decomposed into odd-even symmetric faces. Then, Gabor features were extracted from both odd faces and even faces to get the Gabor odd-even symmetric features，which could be fused via a weighting factor to generate the new features. At last, the newly obtained features were combined to form an over-complete dictionary which was used by sparse representation to classify the faces. The experimental results on AR and FERET face databases show that the new method can achieve high accuracy even when face images are under expression, pose and illumination variations.

In order to improve the quality of the solution in solving Quadratic Assignment Problem (QAP), an effective Max-Min Ant System (MMAS) was designed. Firstly, by using optimal iteration idea, the location and its corresponding task were selected randomly from the current optimal tour as the initial value of next iteration, so as to enhance the effectiveness of each searching in MMAS. Secondly, in order to increase the purpose of the search in every step, the incremental value of target function after adding new task was used as the heuristic factor to guide effectively the state transition. Then, the pheromone was updated by using the multi-elitist strategy so that it could increase the diversity of the solution. And an effective double-mutation technique was designed to improve the quality of solution and accelerate the algorithm convergence speed. Finally, a large number of data sets from QAPLIB were experimented. The experimental result shows that the proposed algorithm is significantly better than other algorithms in accuracy and stability on solving quadratic assignment problem.

The accuracy of the speech endpoint detection has a direct impact on the precision of speech recognition, synthesis, enhancement, etc. To improve the effectiveness of speech endpoint detection, an algorithm based on critical band and energy entropy was proposed. It took full advantage of the frequency distribution of human auditory characteristics, and divided the speech signals according to critical bands. Combined with the different distribution of energy entropy of each critical band of the signals respectively in the speech segments and noise segments, speech endpoint detection under different background noises was completed. The experimental results indicate that the average accuracy of the newly proposed algorithm is 1.6% higher than the traditional short-time energy algorithm. The proposed method can achieve the detection of speech endpoint under various noise environment of low Signal to Noise Ratio (SNR).

Considering clothes and accouterments, gait recognition method based on dynamic feature was proposed in this paper. Firstly, a value could be got by solving the Poisson equation in the gait shape area and a threshold function was constructed for dynamic feature of gait sequence. Secondly, the angle interval mean and variance of all values of the gait silhouette images in a sector region were computed. And the dynamic feature vector was constructed by them. Finally, Support Vector Machine (SVM) was used to classify the gait sequences with clothes and accouterments. The experimental results show the effectiveness of the proposed method in the CASIA gait database.

In order to solve the problem of geometric distortion and low efficiency in the process of biased skeleton trim, a new algorithm of biased skeleton trim based on intersecting cortical model was proposed. At first, according to inherent features of skeleton biased branch, definitions of endpoint and junction point were introduced and revised in the algorithm to accurately locate skeleton branch and biased branch. Then, with that information and the iteration number of intersecting cortical model, flameout condition of neurons spreading was set up. Finally, guided by that condition, the biased skeleton branch can be judged fast and trimmed accurately, with the aid of impulse dynamically generated by ignition neurons, which has biological nature of parallel transmission. Compared with conventional methods based on mathematical morphology, the experimental results show that the proposed algorithm has good performance in structural integrity of skeleton, as well as computation speed and anti-noise ability.

In order to solve the distortion of wool geometric dimension, resulting from image binarization with weak contour, an automatic binarization algorithm for Charge-Coupled Device （CCD） wool image was proposed with reference to a ramp-width-reduction approach based on intensity and gradient indices, using a classical global threshold method and a local one. In that algorithm, edge-pixel-seeking step was added and gray-adjusting factor was improved, with sobel operator and ramp edge model introduced, to increase processing efficiency and avoid human intervention. Besides, every sub image was processed by the mixed global and local threshold based on the analysis of Otsus and Bernsens methods to intensify edge details and decrease distortion. Compared with the traditional ways, the experimental results show that the new algorithm has good performance in automatic binarization with weak contour.

The existing methods of malware feature selection and reduction methods were studied. Current attribute reduction methods of malware do not take advantage of the information of feature selection evaluation function. So a method was proposed to order all features based on their value of information gain and their size, and used attributes order reduction method to get a reduction. An analysis of spatial and temporal complexity was given, and the overall design was given. Test results show that the application of attributes order reduction can obtain fewer reduction results in less time, and get higher classification accuracy using the reduction result.