Improved Convolutional Neural Network based Feature Extraction Method

Deep learning algorithms based on convolutional neural networks have been widely researched and developed in the field of images. This helps in more accurate classification and recognition of images. In order to improve the recognition accuracy of convolutional neural network and optimize the learning performance of neural network, an improved dynamic adaptive pooling algorithm is proposed. First, an overview of the basic structure of convolutional neural networks, convolutional layers and pooling layer operations. Second, build a convolutional neural network model, study and compare different network pooling models. Finally, an improved dynamic adaptive pooling model is constructed for the case where the existing algorithm has a slow convergence speed. Experiment on handwritten database. The simulation results show that as the number of iterations continues to increase, the mean square error continues to decrease, and the recognition accuracy of the model continues to improve. The improved pooling method not only makes the feature extraction of the convolutional neural network more accurate, but also improves the convergence speed and model accuracy, and achieves the purpose of optimizing the network learning performance. This approach can be further extended to other models related to convolutional neural networks.