Optimizing CNNs for Static Pose Recognition Using Wavelet Features and Residual Constraints

With the continuous progress of deep learning technology, convolutional neural networks (CNNs) have made impressive achievements in the field of human action recognition. However, in order to further improve the performance of the model on complex datasets, optimizing the CNN structure and feature extraction methods is still a hot topic in current research. This experiment aims to improve the recognition performance of the model by selecting a suitable CNN model through the convolutional neural network comparison detection method, and elaborating the reason why Fourier transform is not used but wavelet transform is used to optimize the input data. NASNet-Large is chosen as the base model for the experiment, and a 2-level db4 Daubechies wavelet decomposition with MAD adaptive thresholding is used for preprocessing optimization. Under the unified dataset and experimental conditions, the effects of different processing methods on the model performance were comparatively analyzed through training and testing. The experimental results show that compared with the original NASNet-Large baseline model, the model recognition accuracy of progressive processing wavelet transform preprocessing with accurate adaptive thresholding and soft thresholding is improved by 0.76%, and the effectiveness of db4 wavelet in maintaining the integrity of the high-frequency features without taking excessive denoising is also verified, which provides a quantitative optimization strategy and theoretical support for research and application in related fields.