Deep neural networks with 32-bit weights require a lot of computing resources, making it difficult for large-scale deep neural networks to be deployed in limited computing power scenarios (such as edge computing). In order to solve this problem, a plug-and-play neural network quantification method was proposed to reduce the computational cost of large-scale neural networks and keep the model performance away from significant reduction. Firstly, the high-frequency and low-frequency components of the input feature map were separated based on Octave convolution. Secondly, the convolution kernels with different bits were respectively applied to the high- and low-frequency components for convolution operation. Thirdly, the high- and low-frequency convolution results were quantized to the corresponding bits by using different activation functions. Finally, the feature maps with different precisions were mixed to obtain the output of the layer. Experimental results verify the effectiveness of the proposed method on model compression. When the model was compressed to 1+8 bit(s), the proposed method had the accuracy dropped less than 3 percentage points on CIFAR-10/100 dataset; moreover, the proposed method made the ResNet50 structure based model compressed to 1+4 bit(s) with the accuracy higher than 70% on ImageNet dataset.

Functional hyper-networks are widely used in brain disease diagnosis and classification studies. However, the existing research on hyper-network construction lacks the ability to interpret the grouping effect or only considers the information of group level information of brain regions, the hyper-network constructed in this way may lose some useful connections or contain some false information. Therefore, considering the group structure problem of brain regions, the sparse group Lasso (Least absolute shrinkage and selection operator) (sgLasso) method was introduced to further improve the construction of hyper-network. Firstly, the hyper-network was constructed by using the sgLasso method. Then, two groups of attribute indicators specific to the hyper-network were introduced for feature extraction and feature selection. The indictors are the clustering coefficient based on single node and the clustering coefficient based on a pair of nodes. Finally, the two groups of features with significant difference obtained after feature selection were subjected to multi-kernel learning for feature fusion and classification. The experimental results show that the proposed method achieves 87.88% classification accuracy by using the multi-feature fusion, which indicates that in order to improve the construction of hyper-network of brain function, the group information should be considered, but the whole group information cannot be forced to be used, and the group structure can be appropriately expanded.

The feature representation extracted from the functional connection network based on single brain map template is not sufficient to reveal complex topological differences between patient group and Normal Control (NC) group. However, the traditional multi-template-based functional brain network definitions mostly use independent templates, ignoring the potential topological association information in functional brain networks built with each template. Aiming at the above problems, a multi-level brain map template and a method of Relationship Induced Sparse (RIS) feature selection model were proposed. Firstly, an associated multi-level brain map template was defined, and the potential relationship between templates and network structure differences between groups were mined. Then, the RIS feature selection model was used to optimize the parameters and extract the differences between groups. Finally, the Support Vector Machine (SVM) method was used to construct classification model and was applied to the diagnosis of patients with depression. The experimental results on the clinical diagnosis database of depression in the First Hospital of Shanxi University show that the functional brain network based on multi-level template achieves 91.7% classification accuracy by using the RIS feature selection method, which is 3 percentage points higher than that of traditional multi-template method.

To deal with poor robustness in strong noise environment, weak adaptability to complex mixed noise that appear in the existing unsupervised image segmentation models, an improved noise-robust image segmentation model based on One-class SVM (Support Vector Machine) method was proposed. Firstly, a data outlier detection mechanism was constructed based on One-class SVM. Secondly, an outlier degree was introduced into the energy function, so that more accurate image information could be obtained by the proposed model under multiple noise intensities and the failure of weight-descend mechanism in strong noise environment was avoided. Finally, the segmentation contour was driven to the target edge by minimizing the energy function. In noise image segmentation experiments, the proposed model could obtain ideal segmentation results with different types and intensities of noise. Under F₁-score metric, the proposed model is 0.2 to 0.3 higher than LCK (Local Correntropy-based K-means) model, and has better stability in strong noise environments. The segmentation convergence time of the proposed model is only slightly longer than that of LCK model by about 0.1 s. Experimental results show that the proposed model is more robust to probabilistic, extreme values and mixed noise without significantly increase of segmentation time, and can segment natural images with noise.

It is difficult to get correct segmentation results of the images with unknown intensity and distribution of noise, and the existing models are poor in robustness to complex noise environment. Thus, a noise adaptive algorithm for image segmentation was proposed based on local intensity difference. Firstly, Local Correntropy-based K-means (LCK) model and Region-based model via Local Similarity Factor (RLSF) model were analyzed to reduce the sensitivity to noise pixels. Secondly, a correction function based on local intensity statistical information was introduced to reduce the interference of samples to be away from local mean to segmentation results. Finally, the active contour energy function and iterative equation integrated with the correction function were deduced. Experimental results performed on synthetic, and real-world noisy images show that the proposed model is more robust with higher precision, recall and F-score in comparison with Local Binary Fitting (LBF) model, LCK model and RLSF model, and it can achieve good performance on the images with intensity inhomogeneity and noise.

Aiming at the deficiencies of traditional Differential Evolution (DE) algorithm, low optimization accuracy and low convergence speed, a Cross-Population Differential Evolution algorithm based on Opposition-based Learning (OLCPDE) was proposed by using chaos dispersion strategy, opposition-based optimization strategy and multigroup parallel mechanism. The chaos dispersion strategy was used to generate the initial population, then the population was divided into sub-groups of the elite and the general, and a standard differential evolution strategy and a differential evolution strategy of Opposition-Based Learning (OBL) were applied to the two sub-groups respectively. Meanwhile, a cross-population differential evolution strategy was applied to further improve the accuracy and enhance population diversity for unimodal function. The sub-groups were handled through these three strategies to achieve co-evolution. After the experiments are totally run for 30 times independently, it is proven that the proposed algorithm can stably converge to the global optimal solution in 11 functions among 12 standard test functions, which is superior to other comparison algorithms. The results indicate that the proposed algorithm not only has high convergence precision but also effectively avoid trapping in local optimum.

The use of resting-state functional magnetic resonance imaging to study the functional connectivity network of the brain is one of the important methods of current brain disease research. This method can accurately detect a variety of brain diseases, including Alzheimer's disease. However, the traditional network only studies the correlation between the two brain regions, and lacks a deeper interaction between the brain regions and the association between functional connections. In order to solve these problems, a method was proposed to construct a functional connectivity network of high-order minimum spanning tree, which not only ensured the physiological significance of functional connectivity network, but also studied more complex interactive information in the network and improves the accuracy of classification. The classification results show that the resting-state functional magnetic resonance imaging classification method based on the functional connectivity network of high-order minimum spanning tree greatly improves the accuracy of Alzheimer's disease detection.

To solve the problems of complex planning method, too many man-made specified parameters and huge computation in the existing gait dynamic model, the gait generation approach of humanoid robot based on the data collected by Kinect to learn human gait was proposed. Firstly, the skeleton information was collected by Kinect device, human joint local coordinate system was built by the least square fitting method. Next, the dynamic model of human body mapping robot was built, and robot joint angle trajectory was generated according to mapping relation between main joints, the studies of walking posture from human was realized. Then, Robot's ankle joint was optimized and controlled by gradient descent on the basis of Zero-Moment Point (ZMP) stability principle. Finally, on the gait stability analysis, safety factor was proposed to evaluate the stability of robot walk. The experimental results show that the safety factor of walking keeps in 0 to 0.85, experctation is 0.4825 and ZMP closes to stable regional centres, the robot realizes walking imitating human posture and gait stability, which proves the validity of the method.

The community detecting algorithm was applied to human functional network to explore the mechanism of human brain. The brain functional data of 28 healthy subjects were collected by functional Magnetic Resonance Imaging (fMRI), and the brain functional network of human beings based on time series was constructed. A threshold range of vertices in the network was designated according to modularity and full connected network theory. The community structures were detected by using the hierarchical clustering algorithm and the greedy algorithm respectively, and the experimental results show that similar community structures have been obtained. Then different performances can be explored across the threshold by analyzing the modularity. An effective threshold range of vertices between 180 to 320 in brain network was proposed. Exploring the community structure is helpful to comprehend the mechanism of brain lesions, which provides a tool for diagnosis and treatment of brain diseases.