Aiming at the problem of the jump existed in the first frame of human motion synthesis method based on Recurrent Neural Network （RNN）， which affects the quality of generated motion， a human motion synthesis method with hidden state initialization was proposed. The initial hidden state was used as independent variable， the objective function of the neural network was used as optimization goal， and the gradient descent method was used to optimize and solve the problem to obtain a suitable initial hidden state. Compared with Encoder-Recurrent-Decoder （ERD） model and Residual Gate Recurrent Unit （RGRU） model， the proposed method with initial hidden state estimation reduces the prediction error of the first frame by 63.51% and 6.90% respectively， and decreases the total error of 10 frames by 50.00% and 4.89% respectively. Experimental results show that the proposed method is better than the method without initial hidden state estimation in both motion synthesis quality and motion prediction accuracy. And the proposed method accurately estimates the hidden state of the first frame of RNN-based human motion model， which improves the quality of motion synthesis and provides reliable data support for action recognition model in real-time security monitoring.

Deep learning model with convolution structure has poor generalization performance in few-shot learning scenarios. Therefore， with AlexNet and ResNet as examples， a derivative-free few-shot learning based performance optimization method of convolution structured pre-trained models was proposed. Firstly， the sample data were modulated to generate the series data from the non-series data based on causal intervention， and the pre-trained model was pruned directly based on the co-integration test from the perspective of data distribution stability. Then， based on Capital Asset Pricing Model （CAPM） and optimal transmission theory， in the intermediate output process of the pre-trained model， the forward learning without gradient propagation was carried out， and a new structure was constructed， thereby generating the representation vectors with clear inter-class distinguishability in the distribution space. Finally， the generated effective features were adaptively weighted based on the self-attention mechanism， and the features were aggregated in the fully connected layer to generate the embedding vectors with weak correlation. Experimental results indicate that the proposed method can increase the Top-1 accuracies of the AlexNet and ResNet convolution structured pre-trained models on 100 classes of images in ImageNet 2012 dataset from 58.82%， 78.51% to 68.50%， 85.72%， respectively. Therefore， the proposed method can effectively improve the performance of convolution structured pre-trained models based on few-shot training data.