Cascaded Deep Convolutional Neural Network (DCNN) algorithm is the first model that uses Convolutional Neural Network (CNN) in facial landmark detection and the use of CNN improves the accuracy significantly. This strategy needs to perform regression processing to the data between the adjacent stages repeatedly, resulting in complex algorithm procedure. Therefore, a facial landmark detection algorithm based on Asymmetric Convolution-Squeeze Excitation-Next Residual Network (AC-SE-ResNeXt) was proposed with only single-stage regression to simplify the procedure and solve the non-real-time problem of data preprocessing between adjacent stages. In order to keep the accuracy, the Asymmetric Convolution (AC) module and the Squeeze-and-Excitation (SE) module were added to Next Residual Network (ResNeXt) block to construct the AC-SE-ResNeXt network model. At the same time, in order to fit faces in complex environments such as different illuminations, postures and expressions better, the AC-SE-ResNeXt network model was deepened to 101 layers. The trained model was tested on datasets BioID and LFPW respectively. The overall mean error rate of the model for the five-point facial landmark detection on BioID dataset was 1.99%, and the overall mean error rate of the model for the five-point facial landmark detection on LFPW dataset was 2.3%. Experimental results show that with the simplified algorithm procedure and end to end processing, the improved algorithm can keep the accuracy as cascaded DCNN algorithm, while has the robustness significantly increased.

Focused on the problem that the huge number of parameters and calculations of Convolutional Neural Network (CNN) limit the application of CNN on resource-constrained devices such as embedded systems, a neural network compression method of statistics based network pruning and tensor decomposition was proposed. The core idea was to use the mean and variance as the basis for evaluating the weight contribution. Firstly, Lenet5 was used as a pruning model, the mean and variance distribution of each convolutional layer of the network were clustered to separate filters with weaker extracted features, and the retained filters were used to reconstruct the next convolutional layer. Secondly, the pruning method was combined with tensor decomposition to compress the Faster Region with Convolutional Neural Network (Faster RCNN). The pruning method was adopted for the low-dimensional convolution layers, and the high-dimensional convolutional layers were decomposed into three cascaded convolutional layers. Finally, the compressed model was fine-tuned, making the model be at the convergence state once again on the training set. Experimental results on the PASCAL VOC test set show that the proposed method reduces the storage space of the Faster RCNN model by 54% while the decrease of the accuracy is only 0.58%, at the same time, the method can reach 1.4 times acceleration of forward computing on the Raspberry Pi 4B system, which helpful for the deployment of deep CNN models on resource-constrained embedded devices.

In view of the problem of premature convergence and stagnation in the Differential Evolution (DE), the distributed memetic differential evolution was put forward. The idea of memetic algorithm was introduced into the DE algorithm. The distributed population structure and the combination strategy in memetic algorithm were applied. In the former strategy, the initial population was divided into multiple subpopulations according to the von Neumann topology and the periodical information exchange was realized among the subpopulations. And in the latter idea, the differential evolution was taken as an evolutionary frame that was assisted by pattern search to balance the exploration and exploitation abilities. The proposed algorithm made full use of advantages of the pattern search and differential evolution, set up an effective search mechanism and enhanced the algorithm to break away from local optima so as to satisfy the demand on population diversity and convergence speed of the search process. The proposed algorithm was run on a set of classic benchmark functions and compared with several state-of-the-art DE algorithms. Numerical results show that the proposed algorithm has excellent performance in terms of solution quality and convergence speed for all test problems given in this study.