Aiming at the problem of the current common Convolutional Neural Network(CNN)in the expression and gender recognition tasks,that is training process is complicated,time-consuming,and poor in real-time performance,a realtime facial expression and gender recognition model based on depthwise separable convolutional neural network was proposed. Firstly,the Multi-Task Convolutional Neural Network(MTCNN)was used to detect faces in different scale input images,and the detected face positions were tracked by Kernelized Correlation Filter(KCF)to increase the detection speed. Then,the bottleneck layers of convolution kernels of different scales were set,the kernel convolution units were formed by the feature fusion method of channel combination,the diversified features were extracted by the depthwise separable convolutional neural network with residual blocks and separable convolution units,and the number of parameters was reduced to lightweight the model structure. Besides,real-time enabled backpropagation visualization was used to reveal the dynamic changes of the weights and characteristics of learning. Finally,the two networks of expression recognition and gender recognition were combined in parallel to realize real-time recognition of expression and gender. Experimental results show that the proposed network model has a recognition rate of 73. 8% on the FER-2013 dataset,a recognition rate of 96% on the CK+ dataset,the accuracy of gender classification on the IMDB dataset reaches 96%;and this model has the overall processing speed reached 70 frames per second,which is improved by 1. 5 times compared with the method of common convolutional neural network combined with support vector machine. Therefore,for datasets with large differences in quantity,resolution and size,the proposed network model has fast detection,short training time,simple feature extraction, and high recognition rate and real-time performance.

In view of poor performance of Coyote Optimization Algorithm (COA), a Best and Worst coyotes strengthened COA (BWCOA) was proposed. Firstly, for growth of the worst coyote in the group, a global optimal coyote guiding operation was introduced on the basis of the optimal coyote guidance to improve the social adaptability (local search ability) of the worst coyote. Then, a random perturbation operation was embedded in the growth process of the optimal coyote in the group, which means using the random perturbation between coyotes to promote the development of the coyotes and make full play of the initiative of each coyotes in the group to improve the diversity of the population and thus to enhance the global search ability, while the growing pattern of the other coyotes kept unchanged. BWCOA was applied to complex function optimization and Quadratic Assignment Problem (QAP) using hospital department layout as an example. Experimental results on CEC-2014 complex functions show that compared with COA and other state-of-the-art algorithms, BWCOA obtains 1.63 in the average ranking and 1.68 rank mean in the Friedman test, both of the results are the best. Experimental results on 6 QAP benchmark sets show that BWCOA obtains the best mean values for 5 times. These prove that BWCOA is more competitive.

To solve the problems of fine-grained image classification methods, such as highly complex methods and difficulty of using deeper models, a Deep Model Transfer (DMT) method was proposed. Firstly, the deep model was pre-trained on the coarse-grained image dataset. Secondly, the pre-trained deep model classification layer was trained based on inexact supervised learning by using fine-grained image dataset and transferred to the feature distribution direction of the novel dataset. Finally, the trained model was exported and tested on the corresponding test sets. The experimental results show that the classification accuracy rates on the STANFORD DOGS, CUB-200-2011 and OXFORD FLOWER-102 fine-grained image datasets are 72.23%, 73.33%, and 96.27%, respectively, which proves the effectiveness of DMT method on fine-grained image classification.

When the target apperance is influenced by the change of illumination, occlusion or attitude, the robustness and accuracy of target tracking system are usually frangible. In order to solving this problem, sparse representation was introduced into the particle filter framework for target tracking and a sparse cooperative model was proposed. Firstly, the target object was represented by intensity in the target motion positioning model. Secondly, the optimal classification features were extracted by training the positive template set and negative template set in the discriminant classification model, then the target was weighted by the histogram in the generative model. Subsequently, discriminant classification model and generative model were cooperated in a collaborative model, and the target was determined by the reconstruction error. Finally, every module was updated independently to mitigate the effects of changes in the appearance of the target. The experimental results show that the average center location error of the proposed model is only 7.5 pixels, meanwhile the model has good performance in anti-noise and real-time.

To solve the problem that traditional image classification methods deal with the whole image in a non-hierarchical way, an image classification method based on visual saliency detection was proposed. Firstly, the visual attention model was employed to generate the salient region. Secondly, the texture feature and time signature feature of the image were extracted by Gabor filter and pulse coupled neural network, respectively. Finally, the support vector machine was adopted to accomplish image classification according to the features of the salient region. The experimental results show that the image classification precision rates of the proposed method in SIMPLIcity and Caltech are 94.26% and 95.43%, respectively. Obviously, saliency detection and efficient image feature extraction are significant to image classification.

To solve the problem in most of conventional multi-task learning algorithms which evaluate risk independently for single task and lack uniform constraint across all tasks, a new hyper-spherical multi-task learning algorithm with adaptive grouping was proposed in this paper. Based on Extreme Learning Machine (ELM) as basic framework, this algorithm introduced hyper-spherical loss function to evaluate the risks of all tasks uniformly, and got decision model via iterative reweighted least squares solution. Furthermore, considering the existence of relatedness between tasks, this paper also constructed regularizer with grouping structure based on the assumption that related tasks had more similar weight vector, which would make the tasks in same group be trained independently. Finally, the optimization object was transformed into a mixed 0-1 programming problem, and a multi-objective method was utilized to identify optimal grouping structure and get model parameters. The simulation results on toy data and cylindrical vibration signal data show that the proposed algorithm outperforms state-of-the-art methods in terms of generalization performance and the ability of identifying inner structure in tasks.