CNN-NDOGSPM-CSVM: Advancing Image Feature Sparse Coding and Classification for Enhanced Recognition Performance

Abstract

Abstract This paper introduces a novel approach, CNN-NDOGSPM-CSVM, for image feature sparse coding and classification. The algorithm combines the analysis of Scale Invariant Feature Transform-Sparse Coding-Spatial Pyramid Matching and Convolutional Neural Networks-Sparse Coding-Spatial Pyramid Matching algorithms with the principles of Support Vector Machines (SVM). The objective is to enhance convergence control, improve the performance of the pyramid space model (SPM), and optimize the CSVM classifier. In the N-DOG-SPM model, emphasis is placed on controlling image feature sparse coding convergence efficiency and achieving real-time updatable sparse coding. The algorithm constructs an operation function using iterative operation direction and gradient parameters. However, initial experiments reveal challenges in achieving ideal convergence control. To address this, the algorithm focuses on constraints for local sparse coding and accurate calculation of control parameters. The SPM is enhanced by expanding the weight distribution range of the learning dictionary and feature sparse coding operation. This improvement addresses the negative effects of sparse coding iteration and value, resulting in improved convergence control and classification results. In the CSVM classifier, traditional minimum value calculation through differential derivatives is replaced. Bayesian recurrence theory and clustering classification theory are employed, while image feature sparse coding of Gaussian distribution is re-described. Posterior probability operation rules are used to solve key direction vectors, improving real-time updating accuracy. Through comprehensive evaluation, the experimental results demonstrate that the proposed CNN-NDOGSPM-CSVM model exhibits high reliability, accuracy, stability, and superior processing efficiency compared to classical image recognition models. The algorithm's performance is assessed in terms of image feature extraction, feature classification, image recognition accuracy, and processing efficiency.