Graph coloring based feature partitioning for multi-view convolutional neural networks

Convolutional neural networks, or CNNs, have shown impressive results for image classification applications. However, to further enhance their performance, this research explores the application of multi-view learning, which leverages multiple perspectives of image data. Initially, CNN is employed to fetch meaningful features from the images. Then, feature partitioning based on graph coloring is applied to generate different views of the dataset. Leveraging graph coloring to partition features based on correlation ensures minimal feature redundancy across views. Further, a model is trained on these views, and the predictions from these classifiers are then ensembled. This improves classification accuracy by enhancing the diversity of feature subsets used for multi-view learning. The experiments are conducted on eight standard image datasets, and the results are evaluated based on classification accuracy and rigorous statistical analysis. With a Friedman rank of 9.0, our proposed approach significantly outperforms other methods (ranging from 1.87 to 7.68) and the traditional CNN performance rank of 1.12. The outcomes justify that the graph coloring-based multi-view CNN model outperforms the traditional CNN model and other multi-view CNN approaches, demonstrating the efficiency of multi-view learning in boosting CNN performance for image classification tasks.