Construction of Image Recognition Model Based on Second-Order Local Feature Aggregation

This paper analyzes the mathematical model constructed by the Net VLAD method, and proposes a second-order function based on the Net VLAD method to solve the problem that the obtained image feature encoding is a first-order function of the local information of the image. Image feature encoding learning dictionary. Use the Net VLAD classifier as the fully connected layer of the CNN model, design the reconstruction residual weight, and redesign the image feature extraction module for the control of the reconstruction residual value range; when the image feature encoding in the Net VLAD learning dictionary is described by a first-order vector, there is a problem that it is difficult to derive and calculate the minimum value of image features. In the image dimensionality reduction transformation kernel, the first-order linear function is promoted to a second-order nonlinear function, and the image dimensionality reduction transformation kernel is used as a breakthrough, and the second-order nonlinear transformation kernel is used. The activation function has the characteristic of being derivable, and the local features of the image samples are subjected to dimensionality reduction processing to achieve more accurate dimensionality reduction calculations for image features. In the experiment, the accuracy of image feature classification, the recognition rate of image samples, and the training and testing efficiency of image samples are used as verification data. By using the CNN deep convolutional neural network model, the Net VLAD learning dictionary and the improved learning dictionary proposed in this paper, the Multiple image sample sets are compared for training and testing, which verifies that the improved learning dictionary based on Net VLAD proposed in this paper has obvious advantages.