Convolutional neural networks (CNNs) are widely used learning models in the field of computer vision. In the process of convolution initialization, the selection of the initial class center has a significant impact on the subsequent recognition accuracy and efficiency. To solve this problem, this paper proposes a CNN image classification enhancement method combined with K-means++ clustering algorithm. Firstly, the optimized K-means++ algorithm is used to cluster the image dataset to generate more representative class centers, to reduce the impact of initial class center selection on the classification task. To meet the special needs in the field of image recognition, this paper introduces the weight adjustment mechanism of image feature space and optimizes the K-means++ algorithm to make the clustering process closer to the high-dimensional feature distribution of image data. Subsequently, these optimized class centers were input into the CNN for deep feature extraction and classification. Experimental results show that the classification accuracy of the proposed method is significantly improved on multiple public image datasets. The classification accuracy on both CIFAR-10 and ImageNet datasets improved more than $\mathbf{4 \%}$. In summary, the proposed method provides a new enhancement strategy for image classification tasks, which has high practical value and application prospects.
Research on Image Classification Enhancement Method Based on K-means++ Clustering and CNN
2024-10-23
592506 byte
Conference paper
Electronic Resource
English
Simplified method of kernel fuzzy c-means clustering for image texture classification
British Library Online Contents | 2008
|British Library Online Contents | 2008
|Efficient fuzzy c-means clustering for image data
British Library Online Contents | 2005
|On Autonomous k-Means Clustering
British Library Conference Proceedings | 2005
|